Handbook of Unmanned Aerial Vehicles [1 ed.] 9789048197064, 9789048197071

Citation preview

Kimon P. Valavanis George J. Vachtsevanos Editors

Handbook of Unmanned Aerial Vehicles

1 3Reference

Handbook of Unmanned Aerial Vehicles

Kimon P. Valavanis • George J. Vachtsevanos Editors

Handbook of Unmanned Aerial Vehicles With 1228 Figures and 233 Tables

Editors Kimon P. Valavanis John Evans Professor and Chair Department of Electrical and Computer Engineering Daniel Felix Ritchie School of Engineering and Computer Science University of Denver Denver, CO, USA

George J. Vachtsevanos Professor Emeritus School of Electrical and Computer Engineering The Georgia Institute of Technology Atlanta, GA, USA

ISBN 978-90-481-9706-4 ISBN 978-90-481-9707-1 (eBook) ISBN 978-90-481-9708-8 (print and electronic bundle) DOI 10.1007/978-90-481-9707-1 Springer Dordrecht Heidelberg New York London Library of Congress Control Number: 2014944662 c Springer Science+Business Media Dordrecht 2015  This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location, in its current version, and permission for use must always be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

To Stella and Panos, my children, and to Dina Kimon To my wife Athena and my granddaughters Ellie, Athena, Sophia and Martha George

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

By Peter van Blyenburgh President, UVS International1

Preamble The European remotely piloted aircraft systems (RPAS)2 industrial community consists of two stakeholder groups (Industry3 & SMEs/SMIs4 , which both encompass manufacturers and operators5 . The civil RPAS market comprises three user groups: commercial, non-commercial (incl. corporate operations & research), and governmental non-military (state & non-state flights) operators (See Fig. 1). Currently, the regulatory responsibilities in the European Union (EU) for civil remotely piloted aircraft (RPA) with a maximum take-off mass (MTOM) of more than 150 kg lay with the European Aviation Safety Agency (EASA), and for RPA with a MTOM of less than 150 kg with the relevant national aviation authorities (NAAs)6 .

Background Various initial national regulations relative to the operation of civil RPAS are now in place (Austria, Czech Rep., Denmark, France, Germany, Ireland, Italy, vii

viii

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

RPAS Ops Governmental

Non-Governmental

Blyenburgh ©

M U T U A L I Z A T I O N

Military Non-Military

Public

State Flights Security related

Police Customs Border Guard Coast Guard

N O N

Not State Flights Incl. Safety related

Civil Protection Fire Fighters National Mapping Agencies

European Union

Flights on behalf of public EU agency (without national oversight)

M I L I T A R Y

Commercial Air Transport (Transport of Persons & Freight)

Scheduled Air Service Non-scheduled Revenue Operations Non-revue Operations

General Aviation

Corporate Operations Flight Training / Instruction Pleasure

Aerial Work / Specialized Operations

Commercial Non-Commercial (incl. Corporate Operations) Training / Instruction Other Miscellaneous

R P A S O P S

Fig. 1 Aerial Operations - Current & Near-Future RPAS Usage

Sweden, UK), are about to enter into force (Belgium, Finland, Lithuania, Norway, Switzerland), or are in preparation (Malta, The Netherlands, Spain) – See Fig. 2. These regulations principally concern Light RPAS7 and they are not harmonized on a pan-European level. The coming into force of these regulations, which is fairly recent in most of these countries, has brought about a dynamic growth in approved8 and authorized9 RPAS operators supplying aerial operation services with an evergrowing diversity of applications10 (See Fig. 3). For reasons of simplicity, approved and authorized operators are hereinafter jointly referred to as certified operators11. In fact, there are currently more than 1500 certified civil RPAS operators in the European Union (See Fig. 4). In practically all cases, current flight operations are taking place within visual line-of-sight (VLOS), at a flight altitude of less than 500 ft above ground level (AGL) with RPA with a MTOM of less than 25 kg, which are principally produced and operated by SMIs. It should also be mentioned that a significant amount of European NAAs facilitate RPAS operations by granting Permits-to-Fly on a case-by-case basis. The European RPAS industrial community now has as objective to initiate activities that will make it possible to start to enlarge the currently permitted flight envelopes, and to operate larger RPAS. A growing number of governmental authorities in various EU countries involved with internal security matters (e.g. municipal & national police, anti-terrorist squads, municipal fire brigades, forest fire fighters, coast guard, civil defence, environmental protection agencies), as well as FRONTEX (European agency responsible for EU border security), have manifested great interest in the use of RPAS. In addition, large corporate entities (e.g. electric grid operators, pipeline network operators, railway operators, oil companies) have woken up to the fact that RPAS could fulfil functions that would be extremely beneficial to their corporate operations. However, in all cases the use of RPAS will depend on the acceptability of the proposed RPAS flight hour cost.

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

MTOM

In place

MTOM

In Preparation

ix

Comments

01

Austria F

02

Belgium F

03

Bulgaria

04

Croatia

05

Cyprus

06

Czech Rep. F

< 150 kg VLOS BLOS

In place (May 2013)

07

Denmark F

< 150 kg VLOS

In place (Jan. 2004)

08

Estonia

09

Finland F

10

France F

< 25 kg

VLOS BLOS < 150 kg VLOS BLOS In place (Apr. ‘12) To be update in 2014

11

Germany

< 25 kg

VLOS

12

Greece F

13

Hungary F

14

Ireland F

< 20 kg

VLOS

In place May 2012)

15

Italy F

< 25 kg

VLOS

In place (Dec. 2013)

16

Latvia

17

Lithuania F

< 25 kg

VLOS

18

Luxembourg

19

Malta F

20

Netherlands F

< 25 kg

VLOS

21

Poland F

< 150 kg VLOS BLOS

22

Portugal

23

Romania F

24

Slovakia

25

Slovenia F

26

Spain F

27

Sweden F

< 150 kg VLOS

In place (Mar. 2013)

28

UK F

< 20 kg

In place (2002) Several updates since

Sub-Total 29

Iceland

30

Norway F

31

Switzerland F

< 150 kg

VLOS < 150 kg VLOS

< 150 kg VLOS

F

Expected mid 2014

In place (2013)

< 150 kg VLOS

< 150 kg VLOS

Expected mid 2014

< 150 kg VLOS

In preparation

< 150 kg VLOS

In place (2012) Update in preparation In place 2013

< 25 kg

VLOS

In preparation

VLOS 12

3

8

1

< 150 kg VLOS BLOS Expected mid 2014 Model aircraft rules apply

Sub-Total Total

Finalised (2013) but not in force

12

3

VLOS over people/ crowds 1

2

9

2

Directive expected 2014

RPAS aerial operations facilitated = Permit to Fly is granted by NAA, based on specific national rules, possibly on a case-to-case basis & for a limited duration (incl. in countries where no national regulation exists). © Blyenburgh

Fig. 2 RPAS Regulation in Europe

x

Foreword: Civil RPAS In European Union Airspace – The Road To Integration Explanation of Terms Commercial & Non-Commercial (including Corporate Operations)

- Aerial Advertising - Aerial Inspection - Aerial Monitoring - Aerial Observation & Surveillance - Aerial Patrol & Spotting - Aerial Photography, Video, Cinema - Aerial Survey & Mapping (Photogrammetry) - Aerial Spraying & Dispensing - Research & Scientific - Aerial Search & Rescue Assistance Blyenburgh ©

Flight Training / Instruction - Duo

(student instruction by licensed pilot)

- Solo (unaided student flight) - Check (verification of qualification of pilot license holder) Other Miscellaneous - Test / Experimental - Demonstration - Ferry / Positioning - Air Show / Race

Inspection: Examination with the intent to find faults, errors, problems, malfunctions or specific phenomena. Monitoring: Observing on a regular basis over a period of time. Observation: Examination of an activity, person, group, area or phenomena. Patrol: Searching for a specific activity, person, group or phenomena. Spotting: Looking for & noting geographical coordinates of an object or activity or phenomena. Surveillance: Close observation of an activity, person, group, area or phenomena. Survey: Detailed inspection of a geo-referenced section of the earth’s surface (including structures) with the purpose to study or measure altitudes, angles, distances. phenomena, on the land and the structures flown over

Fig. 3 Civil RPAS Aerial Work Categories

A considerable amount of civil RPAS related European Commission (EC) initiatives have taken place over the years, namely, the Hearing on Light RPAS (Oct. ’09), the High Level UAS Conference (July ’10)12 , the UAS Panel initiative (Jun. ‘11-Feb. ‘12)13 , the European RPAS Steering Group initiative (Jun. ’12 - today)14 , and a significant amount of RPAS-related study contracts have been financed by various EC Directorate Generals (DG)15 . It should be noted that two study contracts have recently been allocated by DG Enterprise & Industry concerning the societal topics of respectively Liability & Insurance, and Privacy & Data Protection. FRONTEX16 has financed several RPAS capability demonstrations in Spain and Greece to evaluate RPAS usage for illegal immigrant control. In addition, in November 2013 SESAR Joint Undertaking (SJU) announced the 9 consortia that had been selected for funded RPAS demonstration activities. The objective of these demonstrations, which will take place over a period of 24 months (ending in Oct. ’15), is to prepare the integration of operational improvements in SESAR’s ATM Master Plan that should contribute to the safe integration of RPAS operations into European non-segregated airspace starting in 2016. The conclusions of the EC UAS Panel17 , co-chaired by the EC’s DG Enterprise & Industry and DG Mobility & Transport highlighted the potential of civil RPAS for the development of a wide range of commercial and non-commercial, as well as non-military governmental, applications. In addition, the EC concluded that civil RPAS offer significant potential for job creation in the industry and services sectors, and can generate positive economic growth. The underlying principal of the EC UAS Panel was to focus on civil RPAS applications with societal benefits for which a viable business could be made. The need to accelerate the safe integration of RPAS into European airspace - taking into account societal issues such as responsibility, liability, insurance, privacy, data protection, and public acceptance – was also recognized. Furthermore, it was brought to light that, viable business cases for civil RPAS operations can currently, and for the foreseeable future, only be made for Light RPAS. The Joint Authorities for Rulemaking on Unmanned Systems (JARUS)18 federates the NAAs of 22 countries19 , as well as EASA and EUROCONTROL. JARUS contributes to the harmonisation and coordination of the rulemaking, certification, and operational approval of civil RPAS, remote pilots and operators. Its intent

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

Quantity Certificated

Operator

01

Austria F

02

Belgium F

03

Bulgaria

04

Croatia

05

Cyprus

06

Czech Rep. F

24

07

Denmark F

13

08

Estonia

09

Finland F

33

10

France F

497

11

Germany

Pilot

Manufacturer RPAS

Regulation In place (2014)

10

Finalised (2013) but not in force

29

31

In place (May 2013) In place (Jan. 2004)

Expected mid 2014 30

In place (Apr. 2012) To be updated in 2014 In place (2013) [Länder & Federal level (no reciprocity)]

400 (estimate)

12

Greece F

13

Hungary F

14

Ireland F QE

12

15

Italy F

1

16

Latvia

17

Lithuania

18

Luxembourg

19

Malta

20

Netherlands

21

Poland F

22

Portugal

23

Romania

24

Slovakia

25

Slovenia

26

Spain F

27

Sweden F

28

UK

29

Iceland

30

Norway

31

Switzerland F

In place (May 2012) 1

In place (Dec. 2013)

F

In place (Jan. 2002) Update mid 2014

F QE

In preparation F QE

13

30

In place (2012) Update in preparation In place (2013)

F

F

1

F QE

Sub-Total F

Sub-Total Total Certification:

In preparation

216

In place (March 2003)

212

In place (2002) Several updates since

1432

89

32

33

76

Expected mid 2014 Directive expected 2014

76 1508

89

32

33

A form of official recognition of compliance with the applicable requirements by means of the issuance of a certificate attesting such compliance.

F

RPAS aerial operations facilitated = Permits to fly are granted by NAA, based on specific national rules, possibly on a case-to-case basis for a limited duration(incl. in countries where no national regulation exists) .

QE

National Qualified Entity (QE) in place, QE from another country recognised, or QE in the process of being appointed. © Blyenburgh

Fig. 4 RPAS (MTOM) Related Certification in Europe

xi

xii

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

is to eliminate the need for each country to write their individual requirements and to facilitate reciprocal acceptance of RPAS-related certificates, approvals and licenses. In November 2013, JARUS published its first certification specification (CS-LURS)20 and the online comment period of its second specification (AMC RPAS.1309)21 ended on 28 March 2014. The creation of the European RPAS Steering Group (ERSG), co-chaired by the EC’s DG Enterprise & Industry and DG Mobility & Transport and representing 11 stakeholders22, was announced by the EC at the Paris Air Show in 2012. Its objective is to foster the development of civil RPAS operations by planning and coordinating all activities necessary to achieve the safe and incremental integration of civil RPAS into European air traffic starting in 2016. The ERSG has developed a comprehensive European RPAS Roadmap23, consisting of three complementary sub-sets24 that was remitted by Peter van Blyenburgh, as a representatives of UVS International, to Matthias Ruete, Director General Mobility & Transport & Philippe Brunet, Director DG Enterprise & Industry, as representatives of the European Commission at the Paris Air Show in 2013. The European RPAS Roadmap identified JARUS as a key player for its implementation - This fact will most probably bring with it the necessity to give JARUS, in the near future, a recognised status with a formal linkup to EASA.

The Current Situation On Monday 3 February 2014, SJU published a tender25 , which has as purpose to select a consortium that will be entrusted with the preparation of the “Definition Phase” related to the insertion of civil RPAS into the European Aviation System; this tender should be seen in light of the guidelines defined by the European RPAS Roadmap, and within the context of the Single European Sky initiative. In order to be truly representative of the entire European RPAS industrial community and be able to stay in line with the conclusions of the EC Staff Working Document “Towards a European strategy for the development of civil applications of RPAS26 ”, as well as the European RPAS Roadmap, the selected consortium should federate large industrial enterprises and relevant SMEs/SMIs (manufacturers & operators) with measureable experience, so that it can effectively assure that the consortium’s work will indeed address all the topics required for the incremental integration of all classes of civil RPAS, starting in 2016. The tender closing date was 2 April 2014 and the winning consortium should be announced before the summer of 2014. At a news conference that took place in Brussels, Belgium on 8 April 2014, Siim Kallas, the vice president of the European Commission and Commissioner for Transport, officially announced the acceptance of the European RPAS Roadmap. On the same day, a communication to the European Parliament and the European Council entitled “A New Era for Aviation – Opening the aviation market to the civil use of remotely piloted aircraft systems in a safe & sustainable manner” was published by the EC. The EC also referred to the European Summit of 19 December 2013, which called for action to enable the progressive integration of RPAS into the

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

xiii

total aviation system and European airspace from 2016 onwards. In addition, the EC stated that the progressive integration of RPAS should be accompanied by adequate public debate on the development of measures that address societal concerns. The implementation of the European RPAS Roadmap will be instrumental to lay the foundation for the creation of jobs and economic growth in the research, development, production and services sectors in Europe and will favourably position the European RPAS community on the world stage. This situation clearly demonstrates the coordinated and pragmatic European policy within the framework of the Single European Sky (SES) initiative, which has been made possible by the cooperation of multiple stakeholder groups. In order to contribute in a positive and pro-active manner to this initiative UVS International has instigated the creation of a multi-disciplinary study group27 to deal with the topics of responsibility, liability and insurance. It is the intent of this study group, which was announced in December 2013, to produce position papers relative to the topics identified in its Terms of Reference 27 and submit them for consideration to the European Commission, as well as all other interested parties. The first position papers produced by this group will be presented at the upcoming RPAS 2014 conference in Brussels, Belgium on 23-26 June 201428 .

Considerations & Recommendations RPAS encompass many different types of aircraft with a huge variety of airframe types and lift technologies, with maximum take-off weights ranging from several grams to well over 10 000 kg, with speeds varying from hovering to over 1 000 km/h, and with a very wide envelope of flight endurances (from minutes to days). Therefore, RPAS requires a very specific and proportionate approach to rulemaking, which does not impose an excessive burden on the SMEs/SMIs producing and operating them, but that also does not entail unacceptable administrative procedures for them, as well as for the NAAs. However, this proportionate approach should in no way comprise current aviation safety levels. The current regulatory responsibility in the EU (MTOM 150 kg: EASA), which is based on traditional airworthiness considerations, is not considered to be conducive for the creation of a coherent harmonised EU safety policy relative to RPAS. The challenge will be to create rules that are proportionate to risk, while taking into account some, or all, of the following points: MTOM, speed, system complexity, airspace class, population density in the area over-flown, and the specificity of operation. Besides the technical and research aspects relative to the air traffic management and related issues, that will be addressed by SJU’s aforementioned “Definition Phase” tender, there are a number of other just as critical topics that have to be addressed within the context of European and international harmonisation, and with the objective to be able to comply with the EU principal of free movement of persons, products and services throughout the 28 EU Member States.

xiv

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

These topics include, amongst others: a) Operator (commercial & non-commercial) certification & assessment; b) Remote pilot training (theoretical & practical), licensing & proficiency assessment; c) Airworthiness & continued airworthiness assessment; d) Design organization approval; e) Manufacturer organization approval. In addition, particular attention should be paid to harmonizing the criteria relative to: 1) Remote pilot training syllabus; 2) RPAS operation manuals; 3) RPAS instruction manuals (supplied by the manufacturer); 4) RPAS maintenance manuals; 5) RPAS & operation categories; 6) Terminology. As mentioned earlier, the following societal issues should also be addressed: a) Manufacturer, operator & remote pilot responsibility; b) Manufacturer, operator & remote pilot liability & insurance; c) Privacy & data protection; d) Positive awareness creation with the general public relative to RPAS. Many of these subject matters will have to be dealt with by industry through standards organizations (EUROCAE) and dedicated industry working groups, in close coordination with the EC, JARUS (and NAAs), EASA, EUROCONTROL, SJU, the European Space Agency (ESA), and the European Defence Agency (EDA). It should be noted that the two study contracts awarded by EC DG Enterprise & Industry relative to RPAS Liability & Insurance and Privacy & Data Protection should be seen within this framework. It should also be noted that a substantial part of the initially required rules can be developed by JARUS and can be harmonised by the participating NAAs. In view of the many “remote pilot schools” that are currently mushrooming up in a steadily increasing number of countries, the enormous differences between their training courses, the explosive growth of the number of certified RPAS operators, and the necessity of European harmonisation, there appears to be a urgent requirement to address all facets of remote pilot training, as well as the assessment of remote pilot schools. The recognition of remote pilot certificates (and operator certificates) beyond national borders should be a prime objective for the commercial RPAS community. Within the context of the aforementioned, and taking into account the limited resources of the majority of the NAAs, as well as the fact that many of the SMEs/SMIs involved with RPAS as manufacturers or operators are relative “new comers” to aviation, the potential role that Qualified Entities (QEs)29 can play should be closely evaluated. In addition, the concept of QE should be better understood by both NAAs and industry. Only then will it be possible for the relevant stakeholders to recognize the very positive role that QEs can play to help open up the commercial RPAS market, without putting an unnecessary burden on NAAs, in terms of finance and personnel. At the time of going to press, only one EU country (UK) had appointed two QEs. However, these QEs are also recognised in Ireland

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

xv

and The Netherlands. In the near future, QEs are expected to be appointed by the NAAs in three additional EU countries. Mutual recognition, without requirements for additional certificates, licenses, approvals or other documents relative to RPAS operators, remote pilots and RPAS, would appear to be the necessary way forward, in order to create a level commercial playing field for the European RPAS community, and it would influence investment in the civil RPAS sector in a positive way. Lastly, it should be recognised that the principal of “regulator proximity”, will be of critical importance for SMIs/SMIs in the EU. In order to reduce the costs of travel and fees, but also for cultural and linguistic reasons, as well as reasons of comprehension, it seems imperative that all certificates, licenses and approvals inside the EU are based on common rules, but that they are issued by the NAAs in their country’s native language. In order to realise the aforementioned, various stakeholder communities, including RPAS operators (commercial & non-commercial), remote pilot schools, and QEs, have to be federated on a European level. This task is being addressed by the European RPAS Coordination Council, which consists of representatives of the 10 European national RPAS associations (Austria, Belgium, France, Germany, Italy, The Netherlands, Norway, Romania, Spain, UK) and UVS International, in coordination with the national RPAS operator associations in Australia, Botswana, Namibia and South Africa.

Conclusion The European civil RPAS community has grown substantially and has come a long way in a relatively short period of time, and has now passed through the first of several gates on its way to incremental integration into European airspace. All RPAS stakeholders and particularly the European industrial community, now have to demonstrate understanding and maturity, and jointly work together in order to successfully get through the remaining gates on the road to full RPAS integration. By doing so, they will not only serve themselves, but they will also be doing an immense favour to the global RPAS community, by showing that full RPAS integration is indeed reachable objective.

Explanations 1

UVS International is a non-profit association registered in The Netherlands and operating out of offices in Paris, France which is dedicated to the promotion of RPAS (manufacturers & operators). In addition to having corporate, noncorporate & natural person members, it also federates European national RPAS associations & working groups* in Austria (AAI), Belgium (BeUAS), Denmark (UAS Denmark*), France (FPDC), Germany (UAV-DACH), Italy (ASSORPAS), The Netherlands (DARPAS), Norway (UAS Norway), Romania (UVS Romania), Spain (AERPAS), UK (ARPAS), as well as Australia (ACUO), and South

xvi

2

3

4

5

6

7 8

9

10

11

12 13 14 15 16 17 18 19

Foreword: Civil RPAS In European Union Airspace – The Road To Integration

Africa + Botswana + Namibia (CUAASA). UVS International has a partnership agreement with the European Spatial Data Research Network, a not-for-profit organisation linking national mapping and cadastral agencies with research institutes & universities for the purpose of applied research in spatial data provision, management & delivery. See: www.eurosdr.net UVS International represents over 800 companies & organisations. See www.uvs-international.org Remotely Piloted Aircraft (RPA) & remotely piloted aircraft system (RPAS) These terms are recommended by ICAO for current use [instead of unmanned aircraft system (UAS), unmanned aerial vehicle (UAV)] - The acronyms are invariant. Industry = Enterprise with > 250 employees & annual turnover of > Euro 50 million. Small & medium-sized enterprise (SME) & Small & medium-sized Industry (SMI) = < 250 employees & annual turnover < Euro 50 million. Operators = persons, enterprises or organizations engaged in or offering to engage in an RPAS operation. The NAA of each of the 28 EU member states is responsible in their country for rulemaking, certification & operational approval of civil RPAS with a MTOM 500

3,000

24–48

>2,000

20,000

24–48

>2,000

>20,000

>48

TBD

>30,500

TBD

1,500

12,000

2

300 0–500

4,000 50–5,000

3–4 2;000

Range category Close range Short range Medium range Long range

Typical max altitude (ft) 1,000 ft 15,000 ft 30,000 ft Above 30,000 ft

Table 5.5 Proposed classification based on class of airspace used Class Airspace class S&A Transponder 2-way ATC communication VLA/LOS Class G Not required Not required Not requireda VLA/BLOS Class G Requiredb Requiredb Not requireda MA Class A–E Required Required Required MA/A Class A Requiredb Required Required VHA Above Fl600 Requiredb Required Requiredb a Communication with ATC before operation may still be required b May be waived for certain types of operations or under certain conditions

Table 5.6 Source: Industrieanlagen-Betriebsgesellschaft mbH (2001) Category I II

Grade of deviation from planned/required flight path No deviations Minor deviations

III

Remarkable/considerable deviations

IV

Extreme deviations

Explanation/definition Deviations in altitude of not more than 100 ft. Lateral deviations of not more than a nautical mile. UAV is able to correct deviation within 10 s. Deviations in altitude of not more than 500 ft. Lateral deviations of not more than one nautical mile. UAV is able to correct deviation within 30 s. Deviations in altitude of more than 500 ft. Lateral deviations of more than one nautical mile. UAV is not able to correct deviation within 30 s.

differentiation parameters. Of course, the different categories presented can also impose different requirements on pilot certification, onboard control systems, etc.

5.1.3

Classification Based on Autonomy

Another way to categorize UAVs that is also of interest for certification purposes is based on their level of autonomy. Already UAVs are exhibiting autonomy in certain functions, and this trend is expected to increase, especially if one pilot is to control

88 Table 5.7 Autonomous control levels (Source: Clough (2002a))

K. Dalamagkidis

ACL 0 1 2 3 4 5 6 7 8 9 10

Level descriptor Remotely piloted vehicle Execute preplanned mission Changeable mission Robust response to real-time faults/events Fault/event adaptive vehicle Real-time multi-vehicle coordination Real-time multi-vehicle cooperation Battlespace knowledge Battlespace cognizance Battlespace swarm cognizance Fully autonomous

more than one aircraft at the same time, a possibility discussed in Protti and Barzan (2007). In 2005 the autonomous control levels (ACL) were proposed to measure autonomy. More specifically, ten (10) such levels were proposed in Clough (2002a) that were based on requirements like situational awareness, analysis, coordination, decision making, and operational capability. A list of the ACL is presented in Table 5.7. Each ACL is based on three aspects characterizing autonomy, namely, the level of independence from human involvement, the complexity of the mission, and the complexity of the environment (Huang 2007). Of course, some of the distinctions between the ACL defined in Table 5.7 may not be of value for regulatory purposes, and some of them are not applicable for civil UAV. A simpler classification that takes into account only the level of human involvement and is compatible with the four operational modes proposed in Federal Agencies Ad Hoc Autonomy Levels for Unmanned Systems (ALFUS) Working Group (WG) (2004) is provided below: • Remotely piloted: A certified pilot remotely controls the system either within LOS or with feedback from the UA sensors. • Remotely operated (semiautonomous): The UA is given high-level commands (waypoints, objects to track, etc.), and its performance is monitored by a trained operator. In this case the flying is performed by the UA itself, but all the decision making is delegated to a human. • Fully autonomous: The UA is given general tasks and is capable of determining how to accomplish them, even at the face of unforeseen events. It can also monitor its health and take remedial action after the occurrence of faults. It should be noted that as autonomy increases, new regulatory issues will arise. For example, Clough (2002b) and Protti and Barzan (2007) mention that highly autonomous systems may exhibit non-deterministic behavior, something that will be likely prohibited based on current regulatory approaches. It is also evident that issues of liability for highly autonomous systems may also pose questions that will need to be addressed.

5 Classification of UAVs

89

Table 5.8 NATO UAV classification guide from the Sep. 2009 JCGUAV The Joint Air Force Competence Centre (2010)) Normal operating Normal mission Class Category altitude radius Class I (less than Small >20 kg Up to 5,000 ft 50 km (LOS) 150 kg) AGL Mini 2–20 kg Up to 3,000 ft 25 km (LOS) AGL

Micro < 2 kg Class II (150–600 kg)

Tactical

Class III (>600 kg)

Strike combat HALE MALE

5.1.4

Up to 200 ft AGL Up to 10,000 ft AGL

Up to 65,000 ft AGL Up to 65,000 ft AGL Up to 45,000 ft AGL

5 km (LOS) 200 km (LOS)

Unlimited (BLOS) Unlimited (BLOS) Unlimited (BLOS)

meeting (Source: Example platforms Luna, Hermes 90 Scan Eagle, Skylark, Raven, DH3, Aladin, Strix Black Widow Sperwer, Iview 250, Hermes 450, Aerostar, Ranger

Global Hawk Predator B, Predator A, Heron, Heron TP, Hermes 900

Military Classifications

There are several military UAV classifications in use. The NATO JCGUAV presented in September of 2009 a classification guide based on MTOW; see Table 5.8. All UAVs are divided into three classes: Class I for those weighing less than 150 kg, Class II for those in the range 150–600 kg, and Class III for those over 600 kg. More information can be found in The Joint Air Force Competence Centre (2010). Class I is subdivided into small (20–150 kg), mini (2–20 kg), and micro. Class III is also subdivided but based on the operational role of the UAV. The JUAV CoE has defined its own categories that depend on the operational characteristics and other UAV attributes. These categories include tactical, operational, and strategic UAV that have different scope and operate under different commands (U.S. Department of Defense, Office of the Secretary of Defense 2007). Also defined are six levels of domestic use as shown in Table 5.9. A different categorization based on airworthiness requirements is shown in Table 5.10. For all categories, airworthiness and operator qualifications will need to be demonstrated. In addition to that, Cat I aircraft is limited to LOS operations.

90

K. Dalamagkidis

Table 5.9 Domestic use UAV levels and corresponding system attributes as defined by the JUAV CoE (Source: U.S. Department of Defense, Office of the Secretary of Defense (2007))

Level

Airspeed (KIAS)

Weight (kg)

Operating altitude (ft)

0 1 2 3 4 5

250 250 250 250 250 Any

0:9 0.9–9 10–594 595–5,625 5,625 5,625

1,200 3,000 18,000 18,000 18,000 18,000

Table 5.10 Military UAV categories and relevant UAV regulations (Source: U.S. Department of Defense, Office of the Secretary of Defense (2007)) Airspeed Category FAA regulation Airspace usage limit (KIAS) Cat I – R/C model aircraft None (AC 91-57) Class G 100a Cat II – Nonstandard aircraft FAR Parts 91, Class E,G, and 250a 101, and 103 non-joint-use D Cat III – Certified aircraft FAR Part 91 All None a Proposed

5.1.5

Classification Based on Ownership

Finally UAVs – like other aircrafts – can be categorized based on their ownership as public or state when they are owned and operated by public entities like federal agencies or local law enforcement and civil when they are owned by industry or private parties; see Hempe (2006).

References Civil Aviation Authority of New Zealand, Unmanned aerial vehicles. Wellington, New Zealand, 2007 B. Clough, Metrics, schmetrics! How do you track a UAV’s autonomy? in Proceedings of the AIAA 1st Technical Conference and Workshop on Unmanned Aerospace Vehicles, Portsmouth, 2002a B.T. Clough, Unmanned aerial vehicles: autonomous control challenges, a researcher’s perspective, in Cooperative Control and Optimization, chap. 3, ed. by R. Murphey, P.M. Pardalos (Kluwer, Dordrecht/Boston, 2002b), pp. 35–53 K. Dalamagkidis, K. Valavanis, L. Piegl, On Integrating Unmanned Aircraft Systems into the National Airspace System: Issues, Challenges, Operational Restrictions, Certification, and Recommendations, Intelligent Systems, Control and Automation: Science and Engineering, vol. 36, 2nd edn. (Springer, Dordrecht/New York, 2012) European Aviation Safety Agency (EASA), A-NPA, No. 16/2005, policy for unmanned aerial vehicle (UAV) certification, K¨oln, Germany, 2005 Federal Agencies Ad Hoc Autonomy Levels for Unmanned Systems (ALFUS) Working Group (WG), Autonomy Levels for Unmanned Systems (ALFUS) Framework – Version 1.1. NIST, Gaithersburg, MD, USA, 2004 D.R. Haddon, C.J. Whittaker, Aircraft airworthiness certification standards for civil UAVs. UK Civil Aviation Authority, 2002

5 Classification of UAVs

91

D. Hempe, Unmanned aircraft systems in the United States. Presented to the US/Europe international safety conference, Vienna, 2006 H.M. Huang, Autonomy levels for unmanned systems (ALFUS) framework: safety and application issues, in: Proceedings of the Performance Metrics for Intelligent Systems (PerMIS) Workshop, Gaithersburg, 2007, pp. 48–53 Industrieanlagen-Betriebsgesellschaft mbH, CARE innovative action – preliminary study on integration of unmanned aerial vehicles into future air traffic management. Final report, 2001 Joint JAA/Eurocontrol Initiative on UAVs, A concept for European regulations for civil unmanned aerial vehicles (UAV). Final report, 2004 M. Protti, R. Barzan, UAV autonomy – which level is desirable? – which level is acceptable? Alenia aeronautica viewpoint, in Platform Innovations and System Integration for Unmanned Air, Land and Sea Vehicles (AVT-SCI Joint Symposium), Florence, 2007, pp. 12–1–12–12 Range Safety Group, Range Commanders Council, Range safety criteria for unmanned air vehicles – rationale and methodology supplement. Supplement to document 323-99, 1999 Range Safety Group, Range Commanders Council, Common risk criteria standards for national test ranges: supplement. Supplement to document 321–07, 2007 The Joint Air force Competence Centre, Strategic concept of employment for unmanned aircraft systems in NATO, Kalkar, Germany, 2010 U.S. Department of Defense, Office of the Secretary of Defense, Unmanned systems roadmap 2007–2032. Report, 2007 P. van Blyenburgh, UAV systems: global review. Presented at the Avionics’06 conference, Amsterdam, 2006 R.E. Weibel, R.J. Hansman, Safety considerations for operation of different classes of UAVs in the NAS, in Proceedings of the AIAA 4th Aviation Technology, Integration and Operations Forum and AIAA 3rd Unmanned Unlimited Technical Conference, Workshop and Exhibit, Chicago, 2004 A. Zeitlin, UAS S&A standards: challenges and progress, in UAS Yearbook 2009/2010 (UVS International, Paris, France, 2009), pp. 134–136

6

Military and Civilian Unmanned Aircraft George J. Vachtsevanos and Kimon P. Valavanis

Contents 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 6.2 Military Unmanned Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 6.3 Civilian Unmanned Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

Abstract

This chapter is based on information retrieved from the included references at the end of the chapter. It gives a brief overview of the global unmanned aircraft market, followed by a review of the military and civilian unmanned systems sectors. It aims at registering the current state of the art in the unmanned aviation arena projecting what lies ahead.

G.J. Vachtsevanos () Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 96, © Springer Science+Business Media Dordrecht 2015

93

94

6.1

G.J. Vachtsevanos and K.P. Valavanis

Introduction

The unmanned aircraft arena has seen unprecedented levels of growth over the past 20 years with military applications dominating the field. However, the 2013 year was the critical and the turning point year for using unmanned aircraft in civilian and public domain applications for the following three important reasons: • On 20 June 2013, the European Remotely Piloted Aircraft Systems (RPAS) Steering Group, ERSG, which was set up by the European Commission in July 2012 and was composed by a group of stakeholders of the main organizations and experts interested in the integration of RPAS into the European aviation system, handed over to the European Commission the “Roadmap for the Integration of Civil Remotely Piloted Aircraft Systems into the European Aviation System.” This was the result of the mandate received by the ERSG to establish a roadmap for the safe integration of civil RPAS into the European aviation system, aiming at an initial RPAS integration by 2016. This roadmap aims at facilitating decisions taken by the involved organizations, providing transparency and efficiency in planning different initiatives, and supporting the coordination of related activities in Europe. • On 7 November 2013, the U.S. Federal Aviation Administration (FAA) presented the “Roadmap for Integration of Civil Unmanned Aircraft Systems (UAS) in the National Airspace System (NAS),” which was the result of the joint effort between the FAA and the UAS Aviation Rulemaking Committee (ARC). The purpose of this roadmap is to outline the tasks and considerations that need to be completed in order to enable UAS integration into the NAS. This roadmap is the aftermath of, and in accordance with, the Congressional mandate in the FAA Modernization and Reform Act of 2012, Pub. L. 112-95. It is expected that this roadmap will be updated accordingly in subsequent publications, incorporating lessons learned and findings and also revising target dates toward efficient integration of UAS into the NAS. • On 30 December 2013, the FAA announced the selection of six UAS research and test site operators across the USA that are tasked to study a broad range of issues and challenges related to unmanned aircraft. The six test sites can operate until 13 February 2017, with Congress expected to consider extending the test sites beyond that date. This announcement was followed by the statement that “The FAA is committed to the safe and efficient integration of UAS into the NAS, thus enabling this emerging technology to safely achieve its full potential.” Further, the International Civil Aviation Organization (ICAO), which is a special agency of the United Nations (UN), has long been committed to “the safe and orderly development of international civil aviation throughout the world,” being the organization that also sets standards and regulations necessary for aviation safety, security, efficiency, and regularity, as well as aviation environmental protection. The above actions from the ERSG and the FAA along with the backing of ICAO have cautiously opened the door to utilizing unmanned aircraft in civilian and public domain applications. A “domino effect” is expected to accelerate research

6 Military and Civilian Unmanned Aircraft

95

and development; increase funding; create new jobs; generate provable novel and cutting-edge technologies for increased safety and security of unmanned aircraft; contribute to developing functional and operational “standards” for unmanned aviation; enforce closer collaboration and coordination among national and international interested parties; bring closer academia, government, industry, and the private sector; and engage interested parties to discussing and eventually resolving ethical, legal, and privacy issues, to say the least and name but a few opportunities. It has been estimated that the international UAV market will be worth up to $80 billion by 2020. The USA is the world’s largest operator of unmanned aircraft and the front runner in producing and using technology in military services since the 1950s, also expected to continue to represent the biggest market as a result of requirements, spending on the systems, and research and development funding into the technology. The USA was the highest spender on UAVs in 2009 with the North American region (including Canada) spending $2.7 billion. The results from the continued deployment of UAS in Afghanistan and the surrounding area as well as Iraq have also contributed to their increased use and procurement. The U.S. Department of Defense spent more than $17 billion up to 2013 for new UAV systems and technology with $5.0 billion on UAV research and development. The international market for UAVs/UAS has also grown in recent years with other countries worldwide, recognizing the capabilities and widespread utility of unmanned systems and investing heavily in these new technologies. It is anticipated that, as a result of the previous announcements, spending and investment will increase and the market numbers will shortly be revised. Before continuing with a review of the military and civilian use of unmanned aircraft, and since several terms are used to describe such systems, a brief explanation is provided to clarify the context in which such names are used. The terms UAVs and UAS have distinct definitions and it should be clear why, at first, UAVs became UAS: A UAV refers to a pilotless aircraft, a flying machine without an onboard human pilot or passengers. As such, “unmanned” implies total absence of a human who directs and actively pilots the aircraft. Control functions for unmanned aircraft may be either onboard or offboard (remote control). The term UAS was introduced by the U.S. Department of Defense, followed by the FAA and the European Aviation Safety Agency (EASA). This term was meant to signify that UAS are aircraft and as such airworthiness will need to be demonstrated and they are also systems consisting of ground control stations, communication links, and launch and retrieval systems in addition to the aircraft itself. Similar distinctions are true for RPA versus RPAS. Other names include remotely piloted vehicles (RPVs), a term that was used in the Vietnam War, while the U.S. military also calls them RPA a term used to include both the aircraft and the pilot, while the UK and ICAO have designated them as RPAS, to demonstrate the presence of the man in the loop to control them. The ICAO recommended definition for RPAS is “A set of configurable elements consisting of a remotely-piloted aircraft, its associated remote pilot station(s), the required command and control links and any other system elements as may be required, at any point during flight operation.”

96

6.2

G.J. Vachtsevanos and K.P. Valavanis

Military Unmanned Systems

There exist many different types of unmanned aircraft depending on size, payload, and endurance. There is increased interest for Mini UAVs (MUAVs) and Micro Air Vehicles (MAVs) mostly used for short-range close reconnaissance tasks (mostly by Special Forces) and in urban environments, usually operated at company or battalion level. Lightweight MUAVs/MAVs are basically launched by hand or by a bungee catapult. MUAVs/MAVs can usually be disassembled and be carried in a backpack. Power is mostly provided by lithium polymer, lithium sulfide, or zincair batteries, with range typically being no more than 20 km and endurance of up to 2 h. Application wise, the MUAV/MAV and the ground control station (GCS) usually have line-of-sight (LOS) air-to-ground radio relay or microwave links due to the short-range limitation. Although MUAVs/MAVs have extensively been used by the military deployed in Iraq and Afghanistan, they are also suitable for civilian use being relatively inexpensive to purchase and operate. As such, using swarms or multiple MAVs is projected to be a key area in future UAV development that will enable more effective intelligence, surveillance, and reconnaissance (ISR) due to the increased availability of multiple sensors during a mission. Larger UAVs/UAS operated in Afghanistan and Iraq have demonstrated their usefulness in Counter-Improvised Explosive Device (C-IED) tasks. Vertical TakeOff and Landing UAVs (VTOL UAVs) are particularly seen as suitable for C-IED with the air vehicle capable of hovering at a distance to find and locate Improvised Explosive Devices (IEDs). VTOL UAVs are additionally seen as useful in urban environments, for maritime ISR tasks from naval vessels, and in areas where normal fixed-wing UAV operations might be difficult due to terrain, threats, or insufficient support personnel deployed at forward bases to recover the UAV. Militaries have largely procured Tactical UAVs (TUAVs) to provide ISR support to ground forces with ranges typically up to 250 km and an endurance of up to 15 h. TUAVs either are catapult launched or have automatic takeoff and landing (ATOL) capabilities. The USA also uses and refers to a class of UAV systems smaller than the TUAV but larger than the MUAV as the small UAS (SUAS or sUAS). Medium-altitude long-endurance (MALE) and high-altitude long-endurance (HALE) unmanned aircraft offer low-cost alternatives to manned aircraft for strategic reconnaissance tasks, with some having an armed capability to undertake close air support (CAS). GCS operators use beyond-line-of-sight (BLOS) satellite data links, typically Ku-band, to control and receive data from these MALE/HALE UAVs and C-band LOS data links for shorter-range missions, takeoff, and landings. MALE/HALE UAVs provide persistent ISR at long ranges being equipped with electro-optical/infrared (EO/IR) sensors and ground surveillance radar systems without the limitations of manned aircraft, particularly the physiological effect on aircrew flying long endurance missions. Combined solar/electric or hydrogenpowered HALE UAVs could have an endurance of months or years. Some unmanned HALE platforms like solar-powered stratosphere UAVs and lighter-than-air (LTA) airships are likely to be more cost-effective alternatives to manned weather aircraft, satellites, and other space-based platforms. LTA platforms are particularly cheaper

6 Military and Civilian Unmanned Aircraft

97

alternatives to conventional MALE/HALE UAVs. The average LTA airship unit cost is estimated to be $3.0 million with flying operating costs about $50 per hour, compared to up to $10,700 for the MQ-9 Reaper or the Heron 1 UAV that includes the costs of fuel, maintenance, and deployment of personnel. Unmanned combat aerial vehicles (UCAVs) are expected to replace and complement manned strike aircraft in the future, providing an intelligence, surveillance, target acquisition, and reconnaissance (ISTAR), Suppression of Enemy Air Defenses (SEAD), and deep penetration strike in environments where there is a high threat from enemy air defenses. However, at this stage, technologies have not matured to the point to enable UCAVs to replace manned air superiority fighter aircraft. Very low-observable (VLO) or low-observable (LO) stealth characteristics must be incorporated into the UCAV’s airframe to reduce its radar cross section (RCS), infrared (IR) signature, and radio frequency (RF). Weapons and sensors are carried internally to further reduce the UCAV’s RCS. In comparison to their manned aircraft equivalents, UCAVs promise to be much cheaper to develop and procure, although support costs might be high to maintain airframe coatings and materials that give the aircraft its stealth capabilities. A potential future UAV application is for delivering cargo and supplies to frontline troops or in a humanitarian aid operation, as well as for casualty or medical evacuation in areas considered too risky to operate manned land vehicles or helicopters. The USA and Israel are currently investigating UAV technologies in these roles, which may as well transfer to the civilian domains. Another type of military UAV is the so-called attack UAV. This type of UAVs is fitted with a high-explosive warhead that is flown by a GCS operator and then loiters over a target area using their onboard sensors to identify and then attack a target. The main attack UAV is the Harop manufactured by Israel Aerospace Industries (IAI). The delta-wing Harop carries a 23 kg warhead. The European missile company MBDA has also developed an attack MAV designated the Tactical Grenade Extended Range (Tiger) that carries a warhead less than 0.5 kg. The U.S. Army has experienced a significant rise in the use of UAVs with nearly 4,000 air vehicles now in service compared to 54 a decade ago. The U.S. Army’s Aviation Center at Fort Rucker, Alabama, is now home to the U.S. Army Unmanned Aircraft Systems Center of Excellence (UAS CoE), which is tasked with integrating and coordinating Army UAV use as well as defining current and future UAV concepts for the service. The “U.S. Army Roadmap for UAS 2010–2035,” published by UAS CoE in April 2010, defines the service’s long-term UAV procurement plans and how these will be integrated and operated. At battalion level, the U.S. Army operates the RQ-11B Raven MUAV manufactured by AeroVironment located at Monrovia in California. The electrically powered 1.9 kg Raven entered U.S. Army service in 2006. It has a length of 0.9 m and a 1.37 m wingspan. The Raven can carry a 0.18 kg payload comprising a 5 zoom EO camera and IR camera equipped with a laser rangefinder that can designate targets at a range of 25 km. It has a maximum speed of 83 km/h, a range of up to 10 km, and an endurance of 90 min. U.S. Army Raven training is undertaken by the 2nd Battalion, 29th Infantry Regiment, and 197th Infantry Training

98

G.J. Vachtsevanos and K.P. Valavanis

Brigade at Fort Benning in Georgia. Preliminary UAV training is undertaken at the UAS Training Battalion (UASTB) at Fort Huachuca, Arizona. The Raven is also operated by the Netherlands, Italy, and Spain, and it has also been leased in the past by the UK. The U.S. Army has spent some $232 million on R&T and the procurement of the Raven, totaling more than 2,000 air vehicles with a total requirement for 3,000. About $37.6 million was spent on 312 air vehicles in 2011 with orders to continue through to 2014. The Raven is also in service with the U.S. Marine Corps (USMC) with more than 1,020 acquired since 2008 at a cost of about $71 million, with 16 more air vehicles procured in 2011 at a cost of $32.5 million. The U.S. Army’s SUAS Product Improvement Program (SUAS PIP) plans upgrades to the existing Raven UAS, such as new modular sensor suites for enhanced C-IED and CBRN tasks, an increased range of 15 km, and interoperability with unmanned ground vehicles (UGVs) and unmanned ground systems (UGS). After canceling the Future Combat Systems (FCS) program and the acquisition of further MUAVs/MAVs, the US Army now focuses its efforts on acquiring systems like the AeroVironment Wasp and the larger Puma AE (All Environment) as well as the gMAV (Gasoline Micro Air Vehicle) VTOL MAV. Nano-UAVs and other MAVs are also planned with the U.S. Army previously testing in 2004 the 0.8 lb Tactical Mini UAV (Tacmav) manufactured by Applied Research Associates (ARA). The Tacmav is now marketed by ARA as the Nighthawk. The electric-powered 0.43 kg Wasp has a 0.72 m wingspan, an endurance of 45 min, and a 5 km range. It was developed by AeroVironment with the Defense Advanced Research Projects Agency (DARPA) with more than 200 air vehicles acquired by the USAF Special Operations Command (AFSOC) as part of its Battlefield Air Targeting Micro Air Vehicle (BATMAV) program in 2006. Full rate production of an unspecified number of Wasp IIIs was approved by AFSOC in 2008. The USMC also ordered the MAV in 2007 at a cost of $19.3 million being assigned to units at platoon level. The Puma AE is a hand-launched electric 5.9 kg MUAV with a wingspan of 2.8 m. It has an endurance of 2 h and a range of 15 km. It shares a GCS with the Raven and the Wasp. It was selected by U.S. Special Operations Command (USSOCOM) in 2008 to meet its $35 million All Environment Capable Variant (AECV) requirement for a cost-effective shipborne UAV. It entered service in 2009 replacing an earlier version of the Puma. The gMAV is manufactured by Honeywell Defense & Space at Albuquerque, New Mexico, as the T-Hawk (Tarantula-Hawk) and is largely used for C-IED tasks with the US Army and joint EOD units. It is designed to operate in urban environments, and 90 systems (180 air vehicles) have been acquired by the U.S. Navy. The 11 kg ducted fan gMAV is 0.38 m in length and has a 0.33 m diameter. It can carry a 0.9 kg payload of interchangeable gimbaled EO or IR sensors as well as avionics and RF radio pods. The gMAV is powered by a 4.4 hp (3 kW) heavy fuel boxer twin engine, providing a maximum speed of about 74 km/h and the ability to fly in a 37 km/h wind. It has an endurance of 50 min, a range of 10 km, and a maximum service ceiling of 10,000 ft. About 83 gMAV Block I have been delivered since 2007 with plans for up to 166 to be acquired. Block II/III featuring a gimbaled sensor payload has since been fielded by units in Afghanistan with the type due to equip Infantry Brigade Combat Teams (IBCT) from 2011. A larger 18 kg

6 Military and Civilian Unmanned Aircraft

99

version powered by a heavy fuel engine has also been developed for the U.S. Army’s Brigade Combat Team Modernization (BCTM) Increment two program with first flight in 2012. The U.S. Army operates at brigade level the AAI RQ-7B Shadow 200 TUAV, which was delivered in 2004–2008. Some $350 million has been spent on the acquisition of the Shadow for the U.S. Army since 2007. The 208 kg Shadow 200 features a twin-boom and inverted V-tail. It is 3.6 m in length with a height of 1.0 m and a wingspan of 6.2 m if equipped with an enhancement kit. It can carry a payload of up to 36 kg that includes a 35 lb IAI Tamam Plug-in Optronic Payload (POP) POP-300 IR/EO sensor with laser pointer/rangefinder that is capable of detecting targets up to 10 km. The Shadow is powered by a single 38 hp (28 kW) UEL AR741 Wankel engine that provides a maximum speed of 203 km/h or a cruise speed of 166 km/h. Range is 110 km with an endurance of up to 9 h and maximum altitude of 15,000 ft. The Shadow has a Tactical Automatic Landing System (TALS) or can be launched from a hydraulic catapult launcher. The complete Shadow UAS comprises four air vehicles, two GCS and ground data terminals, one portable GCS, and four RVTs. Three air vehicles are transportable by an air vehicle transporter (AVT). The fourth is used as a spare. Some 48 Shadow UAVs were ordered in 2007 by the USMC to replace the RQ-2 Pioneer with deliveries being completed in 2009. In January 2010, AAI was awarded a $39 million contract for the supply of two air vehicles for the U.S. Army and one for the USMC with deliveries in March 2011. Growing military UAV use has mostly been driven by the USA’s operational experience of using these platforms in most theaters, Israel’s adoption of the systems for intelligence and reconnaissance tasks, as well as usage in the Kosovo War, in Afghanistan, and Iraq. UAV technology has now much improved and become more reliable, although there do still remain some problems. Nevertheless, it has been forecast that a large percentage of manned military aircraft missions will be undertaken by UAVs in the near future. It is also projected that the military UAS market will remain of higher value than the civilian market for the foreseeable future due to a higher allocation of funding drawn from national government defense budgets, although tightened spending could have an effect on some programs and procurement. Industry has provided additional funding toward research and development costs of national UAV programs or has developed their own systems with the hope of attracting of domestic and international orders.

6.3

Civilian Unmanned Systems

Possible civilian UAV applications include scientific research, search and rescue, emergency response, traffic control tasks, infrastructure support, aerial photography, forest protection and wildfire monitoring, environmental monitoring, energy and electrical facility monitoring, pipeline inspection, and coast guard support, to name but a few possible applications. Further, UAVs may also be used for crisis management operations during and after natural disasters like hurricane Katrina and Sandy, or after terrorist attacks, to survey disaster zones or to look for survivors. The use of UAVs to transport civilian air cargo could be a lucrative area in the future.

100

G.J. Vachtsevanos and K.P. Valavanis

In the short term, earth observation and surveillance are expected to account for the biggest share of the civilian UAV market. Increasing opportunities within this market has led to the establishment of numerous small and medium enterprises (SMEs) developing low-cost systems for civilian applications. However, there remains the danger of duplication as too many companies are formed offering similar systems. Europe in particular now offers the most diverse choice of UAVs compared to any other region in the world, but may struggle to find sufficient demand worldwide to guarantee that they will be able to focus on UAV technology in the long term. The development of low-cost systems will be a major factor in sales. To ensure low-cost UAV development, some major companies have opted to convert off-the-shelf manned platforms into UAVs. The BAE Systems Herti, for example, is based on the J6 Fregata motor glider developed by Poland’s J&AS Aero Design. The company has additionally used the MT-03 autogyro, which is produced by RotorSport, UK, in partnership with Germany’s AutoGyro, as a basis for its Ampersand UAV. France’s Sagem has converted the German Stemme S15-VT powered glider into the patroller. The USA’s Aurora Flight Services, Russia’s Irkut, and Israel’s Aeronautics Defense Systems have all opted to convert the Austrian Diamond DA42 Twin Star twin-engine aircraft into a UAV or optionally piloted vehicle (OPV) that can be flown manned or unmanned. Italy’s Selex Galileo, meanwhile, has been working with the Italian UAV company Unmanned Technologies Research Institute (UTRI) to develop the Short Take-Off and Landing OPV (STOL OPV) surveillance platform based on the two-seat Xenon gyroplane produced by Poland’s Celier Aviation. The level and the degree of use of some VTOL UAVs, TUAVs, and MALE UAVs for law enforcement and ground and border surveillance will basically be determined by funding from national governments with a leased fee-for-service arrangement more likely in an effort to keep costs down. An outright UAV purchase might not be seen as financially viable if taking into account that a complete UAS like the BAE Systems Herti will not cost much less to buy than a single Eurocopter EC145 – a platform being regularly chosen for police air support tasks. A manned platform like a helicopter is likely to remain the preferred choice as it provides more versatility, although a UAV has longer endurance time, lower operating costs, and a smaller logistics footprint. Outsourcing where a private contractor provides UAV services is already a growing trend among cash-strapped European militaries. Focusing on law enforcement tasks and in addition to the surveillance role, some UAVs could be fitted with a Long Range Acoustic Device (LRAD) loudspeaker and lightweight nonlethal weapons to reduce the risk to police officers on the ground. France’s Tecknisolar Seni test flew its Bourdon MUAV in 2004 fitted with a 1.1 kg Verney-Carron Flash-Ball nonlethal defense weapon that fires 44 mm rubber balls, plastic rounds, tear gas, or dye-ball markers. France’s SMP Technologies plans to incorporate a 0.5 kg taser stun gun, which it manufactures into an I-Drone quadrotor UAV for law enforcement tasks. Austria’s Schiebel has already flown its Camcopter S-100 VTOL UAV with a LRAD in a military test. There is currently a wide array of UAV applications in the civilian world, and many more are envisioned for the future. Among them, UAVs monitoring the Arctic,

6 Military and Civilian Unmanned Aircraft

101

deployed to survey areas affected by earthquakes, monitoring marine mammals, use for civilian scientific research or environmental monitoring, weather/atmospheric data collection, oceanographic data collection, agricultural monitoring, and highaltitude geological mapping of magnetic, radiological, and gravimetric data and more, among many others.

6.4

Discussion

There is no doubt that the international market for UAVs or UAS or RPA or RPAS has grown in recent years. The use of these systems instead of manned platforms has particular benefits for an operator, namely, lower procurement costs than for manned aircraft and a small logistics footprint. There is additionally a low risk to human life. Military use is currently the biggest market for UAVs and likely to do so for the foreseeable future, despite growing interest in the systems for civilian uses. Europe and the rest of the world were initially fairly slow to adopt UAVs with the USA and Israel using systems operationally from the 1970s. Earlier UAV technology was unreliable with flight control systems and sensor technology, endurance, and weather conditions limiting usage. UAV technology has now much improved and become more reliable, although there do still remain some problems. Military UAVs are generally operated in restricted airspace away from civilian air traffic, but serious problems are posed when both military and civilian operators start to use the systems in congested airspace, including possible midair collision. Current major efforts are trying to put in place general harmonization rules for UAV operations within civilian or controlled airspace with each state responsible for its own air regulations. Several regional efforts have been launched to draw up guidelines for UAV use in controlled airspace. Despite progress, there is still need for new technology to ensure that UAVs can operate safely in civilian airspace. One important area is the development of sense and avoid (S&A) technology that will enable the UAV to autonomously identify a manned aircraft and then make corrections to its flight path to avoid a midair collision. Integrating UAVs into congested airspace remains one of the main areas of current studies that will draw on existing and future UAV technology. There are still other problems to overcome, particularly for the military, to ensure that UAVs can remain operable, such as the need to develop secure data links to prevent possible hostile jamming and data sent from the UAV’s sensors from getting into the wrong hands. The USA and Israel have already had problems with enemy forces intercepting unsecure data broadcast from UAVs. There is additionally the problem of nations and hostile forces developing and gaining anti-UAV technology. The USA’s Raytheon has already successfully used its Laser Phalanx naval close-in weapon system to shoot down a UAV in tests. There is additionally the problem of the loss of data links between the GCS and the UAV, which has been the most common cause of crashes for the past decade. There is particularly a need to allocate sufficient satellite bandwidth to control UAVs and to provide a means of transmitting communications and real-time data from sensors to ground control stations and

102

G.J. Vachtsevanos and K.P. Valavanis

other receiving facilities. The USA has had problems with the allocation of enough bandwidth that has led it to use commercial satellite operators in addition to its own military satellite systems. The London Satellite Exchange (LSE), for example, is one commercial operator that has been providing Ku-band links for UAVs for several countries, including France. Research is being undertaken to address some of these issues. In Europe, for example, a study is being conducted by the European Space Agency (ESA) on European and Canadian satellite technologies. The Advanced Research on Telecommunication Satellite Systems (ARTES) program also includes planned studies into satellite communications (satcom) systems for civilian UAVs. A separate study dubbed SIGAT has also worked on identifying appropriate military radio spectrum for UAV integration into general air traffic in Europe. UAVs featuring autonomous flight capabilities could lower bandwidth use with the operator just responsible for guidance to the target area or making the decision to release weapons. Some of these potential roles for UAVs could, however, be controversial, and their use by law enforcement agencies might not be fully supported by the public. There are also issues about the export and proliferation of UAVs. The informal Missile Technology Control Regime (MTCR) prevents the proliferation of certain UAVs, particularly systems that could be used to deliver weapons of mass destruction. It has restricted the sale of MALE systems, as has the USA’s own export laws. Several UAV manufacturing nations, such as China, are not part of the MTCR that has led to concerns that the technology could end up in the hands of terrorist organizations. MTCR has, however, restricted sales of UAVs to nations outside Europe. The USA’s General Atomics Aeronautical Systems Inc (GA-ASI) has been keen to export of an unarmed version of the Predator MALE UAV, dubbed the Predator XP.

6.5

Conclusions

Although UAVs have been and are a familiar presence in war zones, the picture is changing and roadmaps have been produced to integrate UAVs into the global airspace. The civilian sector is attracting, of course, the interest of UAV OEMs and users, worldwide. In some cases, the civilian market is driven by specific needs such as border patrol, forest fire monitoring, rescue operations, etc. In other cases, the need is more generic and the UAVs may find multiple application domains, like traffic monitoring, police surveillance, etc. UAV applications are set to explode in the commercial market once airspace regulations are clearly defined and published. The Association for Unmanned Systems International (AUVSI) and UVS International hold annual shows and conferences lobbying governments for greater access to civilian skies. The airspace that belongs to manned systems still presents a controversial barrier to the civilian UAV market segment as regulatory agencies had limited significantly UAV access to the national airspace. For example, the FAA’s line-of-sight restriction and other certification requirements have been a major roadblock. However, the

6 Military and Civilian Unmanned Aircraft

103

2013 year is a turning point year as roadmaps for integrating UAS into the national airspace have been published both in the USA and Europe. Regardless, we are witnessing a growing role of relatively low-cost, nearly expendable UAVs in military and civilian operations. The use of these systems, instead of manned platforms, is providing particular benefits to the operator, namely, lower procurement costs and a small logistics footprint. UAVs are currently a growth industry in the aviation sector, with scores of new companies competing for a slice of the market. We are about to witness major changes as UAV technologies are becoming more reliable and safer, and their utility is expanding rapidly, accompanied by demand and the rapid growth of the industry. Reliability is likely to be a key issue for UAVs aimed at civilian use as the industry continues to lobby aviation regulators to gain access to skies that for the most part remained off-limits. A major challenge that needs to be addressed and solved before widespread use of UAVs relates to privacy, legal, and ethical issues.

References Army Unmanned Aircraft System Operations, FM-3-04-155, 14 May 2008 FY2009–2034 Unmanned Systems Integrated Roadmap, Office of Secretary of Defense, 6 Apr 2009 https://extranet.acq.osd.mil/uwir/ Joint Concept of Operations for Unmanned Aircraft Systems, Joint Requirements Oversight Council, Nov 2008 Joint Direct Support of Aerial Intelligence Surveillance Reconnaissance, (JDSAISR) Initial Capabilities Document (ICD), 6 Aug 2010 W. Lewis, A. Boydston, Qualification and reliability of complex electronic rotorcraft systems, in AHS Specialists’ Meeting on Systems Engineering, Hartford, 15–16 Oct 2009 Operational Environment 2009–2025, TRADOC, Aug 2009. (v5) Sense and Respond Logistics: Co-Evolution of an Adaptive Enterprise Capability: DoD Office of Force Transformation, Feb 2004 Sense & Respond Logistics Technology Roadmap, Office of the Deputy Under Secretary of Defense, Mar 2009 The U.S. Army Unmanned Aircraft Systems Strategic Vision 2007, Unmanned Futures Branch, CRD The U.S. UAS Future’s White Paper, 1 Oct 2009 Unmanned Aircraft System Airspace Integration Plan, Office of the Secretary of Defense, Draft Version 1.7, 30 Oct 2009 Unmanned Ground Systems Roadmap, Robotic Systems Joint Project Office, July 2009 Unmanned Ground Systems Roadmap, Robotic Systems Joint Project Office, Addendum, July 2010 Unmanned Systems Integrated Roadmap, FY2011–2036, DOD Reference Number: 11-5-3613 U.S. Army Roadmap for UAS, 2010–2035

Section II UAV Design Principles Daniel J. Pack and Graham Drozeski

7

UAV Design Principles: Introduction Kimon P. Valavanis and George J. Vachtsevanos

UAV Design Principles presents the basic design principles of UAVs and the components that comprise a complete unmanned aircraft system, avoiding in depth technical details. It aims at creating the “overall picture” with respect to UAV functionality and, consequently, to indirectly motivate the reader about the challenges that need to be overcome in order to make UAVs fully operational, reliable, and safe. Computational and Experimental Design of a Fixed-Wing UAV by RyaciotakiBoussalis and Guillaume presents a comprehensive methodology for the design and manufacture of all aspects of a prototype fixed wing UAV, which also allows for utilizing multi-segmented flight control surfaces on the wings of such UAVs. Adding multiple segments to UAV wings creates smaller control surfaces; thus, refined adjustments to UAV performance while airborne may be accomplished. UAV Handbook: Payload Design of Small UAVs by Gruber, Kwon, Hager, Sharma, Yoder, and Pack discusses payload design issues for small UAVs, which are essential to overcome various small UAV constraints imposed by stringent weight, power, and volume. The efficacy of design principles is demonstrated with the help of the payloads for a fixed wing UAV developed by the Center for Unmanned Aircraft Systems Research at the U.S. Air Force Academy, and it is shown that the system requirements for specific applications are closely related to those for other small UAV applications.

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 133, © Springer Science+Business Media Dordrecht 2015

107

108

K.P. Valavanis and G.J. Vachtsevanos

Small UAV Design Development and Sizing by Brandt describes a rationale for selecting a low-risk UAV configuration and a methodology for sizing and optimizing the shape of both fueled and electric-powered aircraft. Three examples of low-risk UAV configurations are shown to be essentially similar from the point of view of optimization and sizing. The sizing methodology for fueled aircraft is described in detail, and a similar one for electric-powered aircraft is described to the extent that it differs from that for fueled aircraft. Typical values of technology-related parameters used in sizing calculations are given, followed by an example of a small UAV designed using the proposed methodology. Systematic Design Methodology and Construction of Micro Aerial Quadrotor Vehicles by Phang, Li, Chen, and Lee presents guidelines to optimally and systematically design and construct an ultralight micro aerial quadrotor vehicle, including details of the steps to design a stable quadrotor with not more than 50 g take-off weight and flight duration of 8 min. The methodology covers, sequentially, structural analysis of the micro quadrotor air frame, design in a 3-D virtual simulator, design of the avionic system, controller implementation via simulation, and component fabrication and implementation. Finally, performance is evaluated, tested, and confirmed with real flight missions. Dexterous UAVs for Precision Low-Altitude Flight by Jiang and Voyles follows the concept of force closure (a term from the dexterous manipulation community) to introduce a new type of dexterous six degrees-of-freedom (6 DOF) UAV, which provides the unique capability of being able to resist any applied wrench, or generalized force-torque, providing more precise control during low-altitude flight. The major challenge that needs to be overcome in such designs is that lowaltitude flight usually introduces ground-effect disturbances and other backwash issues, while in the new field of aerial mobile manipulation it often includes close to structures operations for inspection or manipulation purposes. The presented design differs considerably from typical helicopters or quadrotors, which cannot instantaneously resist or apply an arbitrary force in the plane perpendicular to the rotor axis, thus, making such designs inadequate for complex mobile manipulation tasks. Collectively, the five chapters provide a plethora of ideas and design principles that prepare the reader to capture more advanced technical challenges.

8

Computational and Experimental Design of a Fixed-Wing UAV Helen Ryaciotaki-Boussalis and Darrell Guillaume

Contents 8.1 Odyssey UAV Mission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1.2 Objective of the Design and Development of the UAV. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1.3 Mission Requirements of the UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Odyssey UAV Design Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Initial Sizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Aerodynamic Design Process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.3 Geometry and Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.4 Airfoil Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.5 Structural Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.6 Engine Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Odyssey UAV Software Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Structural Analysis: Femap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.2 Composite Stress Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.3 Modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.4 Deflections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.5 Normal Mode Shapes and Mode Frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.6 Stress and Strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.7 Computational Fluid Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 FOSS System Application to the Odyssey UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.1 Introduction to Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.2 Computational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

110 110 111 111 112 112 114 115 115 120 120 122 122 122 123 125 125 126 126 127 127 130 132

H. Ryaciotaki-Boussalis () Department of Electrical and Computer Engineering, NASA University Research Center (SPACE), California State University, Los Angeles, CA, USA e-mail: [email protected]; [email protected] D. Guillaume Department of Mechanical Engineering, NASA University Research Center (SPACE), California State University, Los Angeles, CA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 121, © Springer Science+Business Media Dordrecht 2015

109

110

H. Ryaciotaki-Boussalis and D. Guillaume

8.4.4 Location of Fiber Line and Effects of Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.5 Number of FBG Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.6 Sensor Integration and Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.7 Analysis and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Flight Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5.1 X-Plane Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5.2 Autonomous Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

132 134 134 136 136 136 137 140

Abstract

This chapter presents a detailed example for the design and manufacture of all aspects of a prototype unmanned aerial vehicle (UAV). It further allows for the utilization of multi-segmented flight control surfaces on the wings of these UAVs. Adding multiple segments to UAV wings creates smaller control surfaces. By introducing smaller control surfaces, a wing can make refined adjustments to UAV performance while airborne. This unique technique will (i) apply localized correcting forces to the UAV, (ii) reduce structural deformation, (iii) minimize drag contribution due to control surface actuation, (iv) suppress and control structural resonance due to lift forces and vibrational modes, (v) reduce the weight of the structure, and (vi) improve the endurance of flights.

8.1

Odyssey UAV Mission

8.1.1

Introduction

In recent years, UAVs have played a major role in the security of the nation and have gained popularity over traditional full-size piloted aircrafts. Further, the use of UAVs has become a favored solution for important tasks requiring air operations. Examples of areas for air operations include academic research, earth environmental studies, fire surveillance, and reconnaissance. This chapter describes the thought process and techniques required to design and develop a fixed-wing UAV. It presents a case study for the design of a particular UAV underway at the Structures, Propulsion, And Control Engineering (SPACE) – a NASA-sponsored University Research Center (URC) of Excellence at the California State University, Los Angeles (CSULA). The main objective was to develop a multi-mission UAV, named “Odyssey,” which could be adapted to different mission requirements. The Odyssey UAV (Fig. 8.1) incorporates a flight control system (FCS) into a twin boom aircraft configuration, which is designed to be capable of carrying a payload of up to 14 lb, remaining in flight for up to 3 h and operating at a maximum altitude of 10,000 ft above sea level. The design encompasses the major areas of aircraft design, which include aerodynamics, structures, avionics, and propulsion. Three scale prototypes were constructed and flown by the SPACE Center team as part of the concept verification, allowing the team to experiment and learn from a

8 Computational and Experimental Design of a Fixed-Wing UAV

111

Fig. 8.1 Odyssey UAV

number of manufacturing techniques prior to the construction of the Odyssey UAV. This UAV was also used to develop and apply novel techniques for real-time structural health monitoring (SHM). The integration of a system for SHM of an aircraft allows for reduction of weight while maintaining a high level of confidence in UAV design. These SHM data are used for fault detection, while a control system is used to assure system reconfiguration.

8.1.2

Objective of the Design and Development of the UAV

During the previous years, the NASA University Research Center of Excellence (URC) has concentrated gaining the experience, skills, and manufacturing techniques for the development of small, lightweight UAVs that have excelled in the area of endurance. The Odyssey UAV project was focused on the design of a multimission/multipurpose air system that can operate autonomously with a maximum payload carrying capacity of 14 lb.

8.1.3

Mission Requirements of the UAV

The mission requirements were selected for the design of a multipurpose UAV that is capable of wide variety of missions with minimal structural modifications. The following are the specific requirements that have been used to design the Odyssey UAV: Endurance: 3 h Payload: 14 lb maximum

112

H. Ryaciotaki-Boussalis and D. Guillaume

end

Clim b

b

Desc

end

Clim

Cruise Back

Desc

3280 ft

Cruise Out

Payload Drop Takeoff

Landing

180 miles

Fig. 8.2 Mission profile

Cruise speed: 81.4 ft/s Cruising altitude: 3,000 ft (above sea level) Max. altitude: 10,000 ft (above sea level) The endurance capabilities of the air-frame were a trade-off between the payload weight and the mission. The mission profile shown in Fig. 8.2 was adopted for the development of the Odyssey UAV. However, the final mission will be determined by NASA once the UAV is developed. As shown in Fig. 8.2, the UAV was designed to take off within a distance of about 300 ft, climb, reach cruise altitude, and maintain a speed of 81.4 ft/s. The UAV will then cruise for 3 h, covering 180 mi, and then finally will decent to a predetermined way point for a payload drop.

8.2

Odyssey UAV Design Specifications

8.2.1

Initial Sizing

The first step in designing Odyssey was to perform a feasibility study by identifying existing UAVs with similar geometry, size, and mission requirements. Using available data from existing UAVs, a parametric study was performed to determine an initial wingspan, power requirement, and UAV maximum speed. The following table shows the UAVs and corresponding parameters that were considered for this study (Table 8.1). Figures 8.3–8.5 graphically show a comparison of the existing UAV’s weight versus cruise speed, weight versus wingspan, and weight versus horsepower. These figures were used to determine the first estimate of the UAV sizing including (1) gross weight, (2) wingspan, and (3) engine size. Specifically, the UAV gross weight was estimated to be between 100 and 150 lb, the wingspan estimated to be between 11 and 16 ft and the engine power estimated to be of at least 7 HP.

8 Computational and Experimental Design of a Fixed-Wing UAV Table 8.1 UAV parametric study UAV Weight (lb) Wingspan (ft) Aerosonde ScanEagle Insight Integrator Aerolight Orbiter Silver Fox Manta Coyote Killerbee Tern Mako Gull 24 Skylark II Neptune Viking 100 Gull 44 Luna x 2000 D1 Shadow 200

28:8 39:6 44 129:8 88 14:3 26:8 51:7 14:1 30:8 129:8 129:8 39:6 77 129:8 149:6 204:6 81:4 74:8 169:4

5:6 10:2 10:2 15:7 13:1 7:2 7:9 8:8 4:8 10 11:4 12:8 8:9 13:8 7 12 16:4 13:7 10:8 12:1

Cruise speed (mi/h)

113

HP 1:7 1:3

N/A 82 82 98:4 84:7 51 63:8 65:6 92:9

8

1:2

11 12 12 4 5:4 15 16 19 6:7 3:5 38:2

101:1

63:8 86:5

Ceiling (ft) 16; 400 19; 496:3 19; 680 9; 997:4 17; 994:1 11; 995 15; 973:6 19; 906:3 15; 088 9; 997:4 9; 997:4

7; 996:6

13; 120 4; 998:7 14; 996:2

Weight vs. Cruise Speed 80.000

Cruise Speed (mph)

70.000 60.000 50.000 40.000 30.000 20.000

Weight vs. Cruise…

10.000 0.000 0.000

50.000

100.000 Weight (lbs)

Fig. 8.3 Weight versus cruise speed

150.000

114

H. Ryaciotaki-Boussalis and D. Guillaume Weight vs. Wing Span 20.000

Wing Span (ft)

15.000 10.000 5.000 Weight vs. Wing Span 0.000 0.000

50.000

100.000 150.000 Weight (lbs)

200.000

250.000

200.000

250.000

Fig. 8.4 Weight versus wing span

Weight vs. HP

50.00 Weight vs. HP

40.00

HP

30.00 20.00 10.00 0.00 0.000

50.000

100.000 150.000 Weight (lbs)

Fig. 8.5 Weight versus horsepower

8.2.2

Aerodynamic Design Process

The traditional aerodynamic analytical relationships were utilized for developing the aerodynamic UAV model by the following steps: Step 1: Make selections such as the airfoils used for the main wing and tail, the configuration of the plane such as pusher or puller, the span, location of wings relative to each other and chord length. Step 2: Perform analysis on the geometry of the airfoils selected for the aircraft. Once the geometry of the aircraft is selected, it is possible to estimate lift-to-drag ratios, static margin, and performance capabilities. The degree of refinement increases as the number of finalized parameters is increased. The increase in refinement begins to reveal the cost of construction and the types of manufacturing techniques that will be required. This is the approach that was used to design and manufacture the Odyssey UAV. Step 3: Use the Athena Vortex Lattice (AVL) computational fluid dynamics software package to model the aerodynamic characteristics of the proposed UAV

8 Computational and Experimental Design of a Fixed-Wing UAV

115

structure. Step 4: Compare chosen aerodynamic characteristics to actual flight data. These aerodynamic characteristics should provide baseline aerodynamic parameters for state-space model to be used in the development of an autonomous controller. Following the steps mentioned above allows for the completion of the UAV design. For UAV design, a decision must be made between increased maneuverability and stability of the aircraft. For the aerodynamic design of the Odyssey UAV, priority was given to an increase in stability and payload rather than an increase in maneuverability. To achieve this, a twin boom pusher assembly with a rectangular wing was selected. Since the payload of the Odyssey UAV includes a forward-facing camera, the pusher configuration also facilitates undisturbed imaging.

8.2.3

Geometry and Configuration

The Odyssey UAV design has a rectangular main wing that has a 15 ft wingspan, a mean aerodynamic chord of 1.75 ft, and an aspect ratio of 8.57. In order to simplify the design and analysis, the main wing does not change geometry along its span nor is it swept. The aircraft is estimated to have a gross weight of 140 lb. The airfoil selected for the main wing is an Eppler214, and the airfoil for the horizontal tail and vertical stabilizers is NACA0012. These airfoil types were selected so that the UAV will exhibit a high lift at low speeds and provide adequate moments for stable pitch. Other aircraft specifications are shown in Tables 8.2–8.4 below. The control surface sizing was based on the results of the stability studies conducted in AVL. Based on the results, the sizes of the control surfaces for the aircraft are shown in Table 8.5. Figure 8.6 shows a representation of the UAV main geometry which includes the main wing and the horizontal and vertical tails. The yellow areas represent the control surfaces.

8.2.4

Airfoil Selection

This section describes the process in determining final geometry of the UAV based on CFD analysis and empirical calculations. As mentioned in the previous section, the airfoil types were selected so that the UAV will exhibit a high lift at low speeds and provide adequate moments for stable pitch. Several experimental studies were performed in order to identify the final airfoil selection for the wing and horizontal tail/vertical tails. The airfoil types for these wing and tail combinations included in these studies were the Eppler214 and NACA0012, Clark Y and NACA2412, and Eppler214 and NACA2412. The Eppler12 and Clark Y were chosen for the main wing, while the NACA0012 and NACA2412 were chosen for the horizontal and vertical tails. Since the interaction of the wing and tail is vital for the performance of the aircraft, the airfoil selection was based on the following design criteria: Lift/drag (L/D) must be greater than 15. Good trim conditions (which implies good pitch stability).

116 Table 8.2 Wing specifications

H. Ryaciotaki-Boussalis and D. Guillaume

Wing (Eppler214 airfoil) Weight Wing loading Aspect ratio Taper ratio Wing area Wingspan Root chord Tip chord Average chord length Quarter chord sweep angle Leading-edge sweep Span efficiency Thickness ratio Total surface area

W D 140 lb W/S D 5.33 lb/ft2 AR D 8.57 Ct/Cr D 1 S D 26.25 ft2 b D 15 Cr D 1.75 ft Ct D 1.75 ft C = 1.75 ft L D 0ı L D 0ı e D 0.505 t/c D 0.137 Ssurf D 67.5 ft2

Table 8.3 Vertical stabilizer specifications

Horizontal stabilizer (NACA0012 airfoil) Span b D 6 ft Root chord Cr D 1 ft Tip chord Ct D 1 ft Planform area S D 6 ft2 Taper ratio lD1 Aspect ratio AR D 6 Thickness ratio t/c D 0.12 Sweep angle L D 0ı Total surface area Ssurf D 12 ft2

Table 8.4 Horizontal stabilizer specifications

Vertical stabilizer (NACA0012 Airfoil) Span b D 2.25 ft Root chord Cr D 1 ft Tip chord Ct D 0.675 ft Ave. chord Ca = 0.8375 Planform area S D 0.675 ft2 Taper ratio l D 0.675 Aspect ratio AR D 7.500 Sweep angle (LE) L D 15.52ı Total surface area Ssurf D 4.500 ft2

Table 8.5 Control surface specification

Chord (in.) Span (in.)

Flaps

Aileron

Elevator

Rudder

6.825 30

6.825 20

3.6 64.8

3.9 27

8 Computational and Experimental Design of a Fixed-Wing UAV

117

Fig. 8.6 Representation of UAV geometry and control surfaces

L/D 20 15 10

L/D RATIO

5

−20

−10

0

0

10

20

30

−5

Fig. 8.7 Angle of attack versus lift-to-drag ratio

−10

E214 & NACA0012

−15

CLARK Y & NACA2412 E214 AND NACA2412

−20 AoA

Figure 8.7 shows the different lift-to-drag ratios for the selected airfoil combinations, and all performed with an L/D ratio greater than 15. Figure 8.8 shows the pitch stability of the aircraft with the different airfoil combinations. It can be seen that Clark Y and NACA2412 and E325 and NACA2012 have negative pitch stability. Since the objective of the Odyssey UAV is to be naturally stable during steady flight, this negative pitch stability is not desired. The selected airfoil combination that provided the most desired performance for the Odyssey UAV was the NACA00012 for the main wing and the Eppler214 for the tails. The final set of experiments for this study was to evaluate the drag performance of the different airfoil combinations. Figure 8.9 demonstrates that the Eppler214 and NACA0012 combination has the lowest drag.

118

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.8 Pitch stability

Pitch Stability 0.1 E214 & NACA0012 CLARK-Y & NACA2412

0.05

E214 & NACA2012 −20

−10

0

0

10

20

30

AoA

−0.05 −0.1 −0.15 −0.2 −0.25

cm

DRAG POLAR 0.25

0.2

CD

0.15

0.1

0.05

E214 & NACA0012 CLARK Y & NACA2412 E214 & NACA2412

Fig. 8.9 Drag polar

−1

0

0

1 CL

2

3

Based on the analysis that was performed, the optimal airfoil combination selected for the Odyssey UAV was the Eppler214 and NACA0012 since this combination provides (1) optimal L/D, (2) the lowest drag polar, and (3) the most stable aircraft performance. The following Table 8.6 shows the aerodynamic data obtained from both empirical studies and CFD analysis.

8 Computational and Experimental Design of a Fixed-Wing UAV Table 8.6 Aerodynamic coefficients from empirical studies Skin friction coefficient Cfe Zero lift parasite drag coefficient Span efficiency Induced drag factors

Air density (sea level) Total engine power Prop efficiency Total power available Speed Max curve slope Stall speed Maximum lift-to-drag ratio

CD,O D eD k3 D k2 D k1 D KD pD PD Hpr D PD VD Clmax D Vstall D L/D D

119

0.0055 0.0184 0.5 0.0735 0 0.0245 0.0980 0.0024 9.80 0.70 6.86 81.40 1.06 57.20 18.8

slug/ft3 HP HP ft/s ft/s

These coefficients are determined from both analytical studies and CFD analysis using AVL. The AVL software uses the vortex lattice method to estimate the lift and drag forces on the modeled aircraft. This software package was used during the design of the Odyssey UAV to ensure that the design provided adequate lift for both the UAV structure and the 14 lb payload. The software was also used to determine aircraft stability and to provide other necessary aerodynamic parameters. The CFD study also determined that the wing and the horizontal tail must have an incidence angle of 2ı . The necessary angles of attack for the main wing were found by optimizing the lift-to-drag ratios at cruising conditions. In addition to this, AVL was used to perform studies to investigate the overall performance of the aircraft at different speeds and altitudes. Optimizing an aircraft design minimizes fuel consumption, minimizes drag, and reduces the loads that the structure experiences. By applying CFD to an aircraft design, performance of the UAV design can be simulated. The simulated performance can be used to determine if off-the-shelf components, such as an engine, can be used while maintaining design parameters of the aircraft. If off-the-shelf parts can be utilized rather than custom parts, both time and cost of manufacturing are reduced. For the Odyssey, the simulation showed that an off-the-shelf engine met the aircraft requirements and a Desert Aircraft model DA-120 engine was purchased. Another important parameter in aircraft design is the static margin. This is the distance between the neutral point (i.e., the entire aircraft’s aerodynamic center) and the center of gravity divided by the length of the chord. The static margin is used in determining the longitudinal stability of an aircraft. A positive static margin on an aircraft means that the center of gravity is in front of the neutral point, while a negative static margin on an aircraft means that the center of gravity is behind the neutral point. As the static margin increases, the longitudinally

120

H. Ryaciotaki-Boussalis and D. Guillaume

stability increases. However, as the static margin increases, maneuverability decreases. As stated previously, the stability of the Odyssey UAV was determined to be more important than its maneuverability, so this UAV was designed with a static margin of 15 %.

8.2.5

Structural Requirements

The structural design of a UAV is based on the aerodynamic loads generated during each stage of the mission which include takeoff, climb, cruise, diving, and landing. These loads are determined from CFD and form the flight envelope of the Odyssey UAV as seen below in Fig. 8.10. Based on the maneuverability requirements of the Odyssey UAV design, the maximum acceleration loading was set to 3.0 g. This is the largest acceleration that Odyssey UAV will experience in the flight envelope. Thus, the aircraft structure was designed to meet the 3.0 g loading condition while maintaining a factor of safety of 2.0. Initial weight estimates obtained from a finite element model (FEM) indicated that an aircraft structure designed to meet these loads would weigh approximately 140 lb.

8.2.6

Engine Selection

Once the aerodynamic analysis is complete, the propulsion system can be selected using the following considerations: (1) the size of the propulsion system must be able to meet the minimum thrust requirements necessary to overcome the drag of the aircraft and achieve lift with an estimated maximum headwind; (2) the type of propulsion system used for an aircraft, such as piston engines, turboprop, and turbofan, turbojet, will influence the performance characteristics of the aircraft; (3) the amount of thrust output by the engine compared to the weight of the engine needs to be maximized; and (4) the thrust-specific fuel consumption (TSFC) which compares the ratio of the rate of fuel consumed to the thrust produced needs to be minimized. The temperature air surrounding the aircraft affects the performance of the aircraft since decreased air density leads to decreased lift and thrust. An aircraft designed with high maneuverability and high-speed requirements will require high thrust requirements (low engine efficiency), while an aircraft designed with simple cruise requirements will have low fuel consumption (high engine efficiency). For UAV propulsion systems designed for a specific mission, trade-offs must be made between thrust and efficiency. Piston engines are internal combustion engines that convert the reciprocating motion of pistons to rotational motion to spin a propeller. These engines are ideal for cruising speeds since they are built to operate at subsonic speeds and provide low thrust-specific fuel consumption. A propeller-driven engine has low maneuverability and cannot produce as much thrust as newer jet engines, but the low TSFC for

8 Computational and Experimental Design of a Fixed-Wing UAV Fig. 8.10 V-n diagram: operational

121

V-n Diagram -Operational

4 3 2

n

1 0

0

10 20 30 40 50 60 70 80 90 100 110 120 130

−1 −2

V-n

−3 Velocity (ft/s)

Fig. 8.11 Thrust required

propeller-driven engines is ideal for endurance missions. Alternatively, airbreathing jet engines, such as turbojet or turbofan, are another type of internal combustion engine which turbines instead of pistons to compress air. These airbreathing engines are capable of producing high thrust but are inefficient and heavy. Since both stability of the aircraft and flight endurance were the main mission requirements of the Odyssey UAV, a piston-type engine was chosen. Figure 8.11 shows the thrust requirements of the piston engine at various speeds, while Fig. 8.12 shows the amount of power required at various speeds. The fuel tank for Odyssey UAV is custom made, and it is capable to hold a fuel capacity that would allow for a maximum 3 h of flight time.

122

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.12 Power required

8.3

Odyssey UAV Software Analysis

8.3.1

Structural Analysis: Femap

A Nastran FEA model was developed to determine the allowable stresses within the aircraft skin and internal structures. For simplicity, only structural elements of the aircraft are modeled. All panels were modeled using 2-D orthotropic materials and laminate properties to determine membrane forces, bending moments, and transverse shear loads. The fuselage and wing skins are composed of sandwich panels made from composite materials. The material properties of the face materials were then defined within the FEM using core material properties specified by the manufacturer. Geometry for the finite element model was imported from Solid Works, and the imported surfaces were meshed with the laminate properties to simulate the sandwich panels.

8.3.2

Composite Stress Analysis

The structure’s margin of safety was determined utilizing the Tsai-Hill Failure Index (FI) to predict the failure in composite materials. The FI is defined as FI D

12  X2



1 1 1 C 2 2 2 X Y Z

 1 2 C

2 22 12 C Y2 S2

where 1 = calculated plane stress of the sandwich panel in direction 1 2 D calculated plane stress of the sandwich panel in direction 2

8 Computational and Experimental Design of a Fixed-Wing UAV

z

6. y x

Output Set: NX NASTRAN Case 1 Contour: Laminate Max Failure Index

123

0.25 0.24 0.23 0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0.

Fig. 8.13 Fuselage stress contours (Tsai-Wu failure index)

12 D calculated in-plane shear stress of the sandwich panel X D normal allowable stress of the sandwich panel in the X direction Y D normal allowable stress of the sandwich panel in the Y direction S D in-plane shear allowable of the panel

8.3.3

Modeling

All internal loads were applied using mass elements when the laminate properties of the fuselage skin, bulkheads, and internal components were modeled. All mass elements are attached to the structure using interpolation elements (rigid RBE3) in the FEM model. The function of the RBE3 is to attach the force or mass of a single dependent node to the average independent node and will not increase the stiffness of the model. RBE3 can also be considered as a “force link,” which distributes the forces of the independent nodes to the dependent nodes by interpolation. The fuselage structure load was applied by specifying the aircraft accelerations in the FEA model as seen in Fig. 8.13. An engine thrust load of 60 lb was also applied to the engine assembly mass element, and the fuselage was constrained at the interface between the bulkheads and the main spar of the wing. The structure of the wing-boom-tail assembly and all internal ribs were modeled with the use of laminate properties. For the wing-boom-tail assembly (Fig. 8.14), all lift loads were applied using nodal loads. A distributed lift load, which decreased toward the wing and tail tips, was calculated by incorporating a maximum gust load of 3.0 g and the aircraft accelerations obtained from the FEA model. The moments

124

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.14 EA analysis of wing structure

Fig. 8.15 Composite wing structure

about the quarter chord of the main wing were also applied to the model by using nodal loads. The wing is constrained at the interface between the bulkheads of the fuselage and the main spar of the wing. Figure 8.15 shows the main load-bearing structure within the wing which is a composite spar that has a box beam configuration. The spar is composed of alternating layers of both bidirectional and unidirectional carbon fiber. The layers of unidirectional carbon fiber are orientated in the spanwise direction to maximize the tensile and compressive strength of the carbon fiber in bending. The bidirectional layers of carbon fiber are orientated in a 45ı angle with respect to the spanwise direction to increase the strength of the spar under torsional loading. Regions where the spar experiences the least bending moments are manufactured with fewer layers of carbon fiber. The loading transferred from the tail through the booms add additional local stresses. To alleviate large stress concentrations, the spar and local sandwich panels were reinforced with additional layers of carbon fiber (Fig. 8.16). The base main wing was optimized and reinforced by increasing the number of layers of carbon fiber to counteract the large bending moments experience near the root chord. As the bending moment decreases with distance from the root, the number of layers of carbon fiber is decreased to prevent the risk of peak stress concentrations. Based on previous structural analysis, additional reinforcement was required in regions where the central wing, the wing tip extensions, and the booms connect.

8 Computational and Experimental Design of a Fixed-Wing UAV

125

Fig. 8.16 Carbon fiber layers for spar

Fig. 8.17 Simulated deflections caused by in-flight loads

8.3.4

Deflections

The results of the FEM show that the aircraft in a 3.0 g maneuver is capable of withstanding an ultimate aerodynamic load of 720 lb and can have expected wing tip deflections up to 1.69 in. (Fig. 8.17).

8.3.5

Normal Mode Shapes and Mode Frequencies

A structure modal analysis must be performed since uncontrolled vibrations can be catastrophic for the aircraft. Based on the finite element modeling shown in Fig. 8.18, the first structural bending mode frequency of the Odyssey UAV main wing was determined to be 8.25 Hz.

126

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.18 Analysis of the first structural bending mode frequency

8.3.6

Stress and Strain

The FEM of the Odyssey UAV wing-boom-tail was used to predict the stress and strain fields in the structure while under load. The greatest stress concentration occurs near the root of the wing, while the peak stress concentration occurs near where the fuselage attaches to the wing (Fig. 8.19). The direction of composite material needs to be specified during manufacturing. The skin material used in the Odyssey UAV is a bidirection carbon fiber weave which has equal stiffness and strength in two directions. However, the spar is composed of unidirectional carbon fiber, which is stiffer and stronger in the tow direction. FEM allows the user to specify the tow direction of the composite material. Since the largest contribution to stress on the wing structure is due to bending, it is critical to analyze the strain contours in the wingspan direction (Fig. 8.20).

8.3.7

Computational Fluid Dynamics

The Athena Vortex Lattice was used as the CFD software to calculate the aerodynamic parameters and the flight dynamics of the Odyssey UAV. It is a free resource software package that is an effective tool to approximate and refine the flightdynamic characteristics of a custom-built rigid-body UAV. The two major approximations made by the Athena Vortex Lattice include the following: (1) the software treated the fuselage as a slender body and did not consider certain drag features (booms and the main landing gear) of the Odyssey UAV model, and (2) the software used an average skin friction drag coefficient. The actual skin friction drag coefficient requires extensive wind tunnel testing, and it typically makes a small contribution to the overall drag coefficient of the aircraft. Regardless of these approximations, the software can provide a reasonable aerodynamic model that can be refined from flight testing.

8 Computational and Experimental Design of a Fixed-Wing UAV

127

Fig. 8.19 Wing assembly stress contours (Tsai-Wu failure index)

Fig. 8.20 Strain contour in the span direction

Despite of its limitations, AVL is still a powerful tool in regard to designing and building custom aircraft and was crucial to the development of the Odyssey. The software can provide lift loading at varying angles of attack, flight stability characteristics, dynamic system matrices, and other parameters necessary to design, manufacture, and control a UAV (Fig. 8.21).

8.4

FOSS System Application to the Odyssey UAV

8.4.1

Introduction to Structural Health Monitoring

The integration of a system for structural health monitoring of an aircraft allows for reduction of weight while maintaining a high level of confidence in UAV design. In addition, a real-time SHM system with a novel robotic/UAV-to-human interface will reduce the risk of in-flight breakups by providing crucial flight data to the ground control station (GCS), including structural deformations, stresses, and loading. These SHM data will also be used for fault detection, while the proposed control system algorithms will assure system reconfiguration. Traditional sensing

128

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.21 Odyssey AVL model

instrumentation – including strain gauges, accelerometers, and thermocouples – tend to be bulky and heavy, limiting their application to a few sensors, usually near the wing root. Recent improvements in fiber optic strain sensing technology (FOSS) have enabled the use of embedded fiber Bragg gratings (FBGs), which can provide numerous distributed strain and temperature measurements from a variety of structural elements. A FBG is a photo-induced refractive plane that is etched onto a fiber-optic cable that reflects a low power signal when light pulses travel through it. As a FBG is stretched, undergoing strain, the spacing of the refractive plane changes as well. This change in the refractive plane spacing causes a change in the reflected signal’s index of refractivity. The change in the index of refractivity is directly proportional to the strain felt by the fiber-optic line. By using a Bragg wavelength (b) as a point of reference to track strain over time, the relationship for strain becomes K" D

  b b

where K is a proportionality constant based on the properties of the fiber-optic cable and  is the wavelength being monitored by the FOSS system. Because of their accuracy, light weight, small size, and flexibility, these fiber-optic sensors are ideal for aircraft with strict weight and size limitations. Given this ability to provide a large number of sensor measurements, there is the potential to develop and implement methods of monitoring, which have a higher accuracy because of the increased sensor density. Thus, a higher accuracy of monitoring can be achieved by this large number of sensor measurements, which facilitates control and system reconfiguration. A strain-based method, known as the displacement transfer functions (DTF), has been developed by the NASA Dryden Flight Research

8 Computational and Experimental Design of a Fixed-Wing UAV

129

Center. The strain-based shape prediction algorithm, also known as the displacement transfer function, is derived from classical beam theory (Euler-Bernoulli) which relates the measured strain of the structure to the theoretical curvature, slope, deflection, and cross-sectional twist angle equations for uniform cantilever beams. The Euler-Bernoulli equation is the general differential equation that describes the relationship between load and deformation of a beam: EI

d4 y.x/ D q.x/ dx4

where the variable x is the distance of the beam starting at the wing root and ending at the end of the beam, y(x) is the displacement of a beam as a function of distance from the root, q(x) is the loading function analogous to the lift generated by the wing, E the elastic modulus, and I is the second moment of area that is calculated with respect to the centroidal axis perpendicular to the applied loading. The moment-strain relationship of the classical beam differential equation is used to implement the shape prediction algorithm and states that the curvature, second derivative of displacement, of a beam is proportional to the applied bending moment: d2 y M.x/ D dx2 EI By relating the bending stresses to the definition of stress and strain in Hooke’s law, it can be shown that the curvature of the beam is directly proportional to the strain felt by the beam during bending: .x/ D

yields d2 y.x/ M.x/c ".x/ &.x/ D E" ! D I dx2 c

where c is the distance to the neutral axis of which the beam is bending. By utilizing the high spatial resolution of the strain measurements, it is possible to numerically integrate the second-order differential by use of finite difference method, without accumulating a large amount of error. By integrating the curvature equation and applying boundary conditions, the slope equation of the nth strain location can be written in the general form: tan n D

i l l h "0 C 2."1 C "2 C "3 C : : : C "n1 / C "n ."n1 C "n / C tan n1 D 2c 2c

3 2 n1 X l 4 tan n D "j C "n 5 "0 C 2 2c jD1 To obtain the displacement of the wing using the shape prediction algorithm, it is necessary to integrate the slope equation with the respect to the x-axis. Utilizing the high spatial resolution of the strain measurements, it is possible to numerically integrate the first-order differential by use of finite difference method, without

130

H. Ryaciotaki-Boussalis and D. Guillaume

accumulating a large amount of error. The displacement of the beam at any given strain sensing station is approximated by yo D 0 y1 D

&

tan 0 D 0

.l/2 .2"0 C "1 / C l  tan 0 C y0 6c

.l/2 .2"1 C "2 / C l  tan 1 C y1 6c After applying boundary conditions, the displacement equation of the nth strain location can be written in the general form: y2 D

yn D

i .l/2 h 2"0 C 3."1 C "2 C "3 C : : : C "n1 / C "n 6c C l.tan 1 C tan 2 C tan 3 C : : : C tan n1 /

# " n1 n1 X X .l/2 2"0 C 3 yn D "1 C "n C l tan i 6c iD1 iD1 This method coupled with the Dryden developed FOSS technology allows for realtime deformation shape prediction. The FOSS technology has shown great potential for applications on various other structures, including smaller-scale and fixed-wing UAVs, as well as space structures and is being used on the Odyssey UAV. Excessive deformations of these structures can potentially induce damage to the wing, which may lead to flight instabilities and even catastrophic in-flight breakups. For fixed wings, the FOSS system has the potential to allow for the prevention of aerodynamic flutter by monitoring resonance frequencies and by providing crucial status of the vehicle’s overall structural integrity. Currently, a computational model has been created and used to investigate and assess the viability of the deflection algorithm as a displacement-shape prediction tool for potential monitoring and control applications. The geometric and physical properties of the FEA model have been addressed, and shape deflection algorithms were developed by subjecting the model to three different loading cases. A number of sensitivity analyses have been investigated and include the placement of the fiber lines on the wing, the density of sensors along the lines, the measurement uncertainty, and the effects of loading on the wing.

8.4.2

Computational Model

The finite element model was created to represent an actual experimental swept plate that was designed, fabricated, and instrumented at NASA’s Dryden Flight Research Center. Fiber-optic strain sensors were placed at the top and the bottom surfaces of

8 Computational and Experimental Design of a Fixed-Wing UAV

131

50. 45.º

12. 2.5

45.º

Fig. 8.22 Swept plate dimensions (inches)

Fig. 8.23 Swept plate sensor layout

the experimental swept plate, and strain data were collected from the plate when it was subjected to various loading conditions. The swept plate is made of 6061-T6 aluminum with a Young’s modulus of E D 10 Msi. The swept plate is 12 in. wide and 0.19 in. thick and has a span of approximately 51 in. The plate is swept horizontally at a 45ı angle from the fixed end (Fig. 8.22). Three fibers are located at the top and the bottom surfaces, for a total of six fibers on the plate. Two fiber lines are placed 0.5 in. from the leading and the trailing edges, while the third fiber line is located in the middle of the plate (Fig. 8.23). There are a total of 102 fiber gratings along each fiber line, and strain measurements are available from each. A combination of Femap and Nastran packages were used to create and analyze the computational models. The FEA models were created such that their physical properties and geometric constraints represent the actual experimental model. Three loading cases have been considered. The uniform-load (UL) case consists of a total of 12 point loads, 6 along each the leading and trailing edges of the wing. The trailing-edge load (TE) case consists of six equally spaced point loads that are applied to the trailing edge of the wing. Finally, the single-point (SP) load case consists of a single point load located at the corner of the plate located at the tip of the trailing edge, opposite the root. The trailing-edge and single-point loading cases are investigated to assess the torsional load effects on the performance of the deflection prediction algorithm. In the case of the leading-edge loads, the 6 lb point loads are applied closest to the fixed end of the plate. In the case of the uniform load, all point loads are 2 lb each. In the case of the single-point load, the point load is 11 lb.

132

H. Ryaciotaki-Boussalis and D. Guillaume

FEA Model

Displacements, {y}

Normalized Estimation Error, {e}

FOSS Approach

∋

Strains, { }

Strain Extrapolation & Injection of Noise to Strain Data

FOSS Algorithm

Estimated ^ Displacements, {y}

Fig. 8.24 FOSS approach sensitivity analysis

8.4.3

Sensitivity Analysis

A simulation was performed where three fiber lines were placed on both the top and bottom of the swept plate. The fibers were placed 0.5 in. parallel to the leading and trailing edges and one directly in the middle of the plate. Real-life situations were considered when assessing the effectiveness of the deflection prediction algorithms. Placement of the fibers on the wing and spacing of the sensing stations will drastically affect the accuracy of the predictions. Strain and deflection data were extracted from the finite element model in order to compare the results of the algorithm. Because the algorithm requires the root deflection and strain information, the data was extrapolated to the root using curve fitting techniques and the known locations of the sensing stations. The strain data may be manipulated to add various levels of uncertainty as needed for the analysis. The deflections were estimated using the FOSS DFT transformations. For each study, the displacements have been estimated, and the corresponding normalized estimation errors, e, were calculated. This process is shown in Fig. 8.24.

8.4.4

Location of Fiber Line and Effects of Loading

The top plate fibers were considered when assessing the effects of fiber location and torsion from loading. The results of the study on the fiber locations and loading conditions are shown in Fig. 8.25. Each case shows the error in deflection prediction for a particular fiber line under three loading conditions. The first loading condition

8 Computational and Experimental Design of a Fixed-Wing UAV

b

140

Deflection percentage error (%)

Deflection percentage error (%)

a

Trailing-edge load

120

Single-point load Uniform load

100 80 60 40 20 0

133

140 Trailing-edge load

120

Single-point load Uniform load

100 80 60 40 20 0

0

10

20

30

40

50

0

Distance from fixed-end (inches) Trailing-Edge

Deflection percentage error (%)

c

10

20

30

40

50

Distance from fixed-end (inches) Middle

140 Trailing-edge load

120

Single-point load Uniform load

100 80 60 40 20 0 0

10

20

30

40

50

Distance from fixed-end (inches) Leading-Edge

Fig. 8.25 Loading study-fiber line

is the trailing-edge load case, followed by the single-point load and uniform-load cases. Figure 8.25a shows the results for the trailing-edge fiber line, Fig.8.25b shows the results for the middle fiber line, and Fig. 8.25c shows the results of the leadingedge fiber line. The deflection prediction algorithm obtained promising results for both the trailing-edge and middle fiber lines. However, the results of the leading-edge fiber line showed high errors near the root of the wing for all three loading conditions. Such high prediction errors and erratic behavior can pose large problems for monitoring applications during flight. As expected, the higher percentage errors are located near the fixed end of the swept wing, where deflections are minimal, while the prediction percentage errors quickly improve toward the tip of the wing.

H. Ryaciotaki-Boussalis and D. Guillaume 120 110 100 90 80 70 60 50 40 30 20 10 0

100 sensing stations 50 sensing stations 30 sensing stations 25 sensing stations

0

10

20

30

40

Deflection percentage error (%)

Deflection percentage error (%)

134

120 110 100 90 80 70 60 50 40 30 20 10 0

50

Distance from fixed-end (inches)

100 sensing stations 50 sensing stations 30 sensing stations 25 sensing stations

0

10

20

30

40

50

Distance from fixed-end (inches)

Fig. 8.26 Sensor spatial resolution comparison: uniform loading, middle fiber

8.4.5

Number of FBG Sensors

Fiber-optic technology using FBG provides the potential for providing high spatial resolution. In a real-life application, increasing the spatial resolution of the fiberoptic sensors provides increased accuracy in capturing strain values. However, increasing the spatial resolution raises computational considerations. The degree of resolution needs to be increased in order to obtain the desired accuracy while maintaining a reasonable computational demand. In order to assess the effects of spatial resolution, four cases were considered using a sensor spacing of 0.5, 1, 1.5, and 2 in. This resulted in a total number of sensing stations per fiber of roughly 100, 50, 30, and 25 sensing stations, respectively. Figure 8.26 shows the results of the uniform loading case on the middle fiber. As expected, increasing the spatial resolution greatly increases the accuracy of the algorithm. Additionally, the percentage errors increase at the free end as the spatial resolution decreases.

8.4.6

Sensor Integration and Loading

A Femap/Nastran FEA model was constructed and included the structural elements of the aircraft skin and internal structures, while the nonstructural items were excluded. All panels and the spar have been modeled using both 2-D orthotropic materials with laminate properties. For the wing-boom-tail assembly, all lift loads were applied by using point loads. The lift load was calculated by incorporating a maximum aerodynamic load of 3.0 g. The applied lift load was a distributed load that decreases toward the wing tips, classified as an elliptical load. The moments about the quarter chord of the main wing were also applied to the model by use

8 Computational and Experimental Design of a Fixed-Wing UAV

135

Fig. 8.27 Femap model with loading

Fig. 8.28 UAV wing strain direction

of nodal loads, and a distributed load was applied to the tail. The wing-boom-tail structure load was applied by specifying the aircraft accelerations in the FEA model. The wing was constrained at the interface between the bulkheads of the fuselage and the main spar of the wing. For this design study, strain data have been extracted from the FEA model of the UAV wing from along the quarter chord and on either side of the quarter chord. These nodes represent the potential placement of the fiber-optic line on the top of the UAV wing (Fig. 8.27). The direction of the strain is along the span of the wing and can be seen in Fig. 8.28. Unlike a simple aluminum plate, the UAV wing is not a homogenous structure. It is constructed using various materials and has complex changes in geometry. These abrupt changes in geometry or material properties can be observed from strain measurements. While the overall behavior of the strain from the root of the wing to the tip of the wing agrees with what would be expected in a cantilever beam-like structure, there are abrupt discontinuities in the strain at locations where the materials or geometry in the wing changes. This could be seen near 40 in. from the wing root. These areas are crucial to ensure the structural integrity of the wing.

136

H. Ryaciotaki-Boussalis and D. Guillaume 8

8 FEA deflection Calculated deflection

6 5 4 3 2

FEA deflection Calculated deflection

6 5 4 3 2 1

1 0

7 Deflection (inches)

Deflection (inches)

7

0 0

20

40

60

80

100 120 140 160 180

Station location starting w/ fixed end

0

20

40

60

80

100 120 140 160 180

Station location starting w/ fixed end

Fig. 8.29 Comparison of deflections for top right and left wing fibers

8.4.7

Analysis and Results

The analysis was performed by extracting the strain data along the span of the wing. Each side of the wing, left and right, was treated as an individual cantilever beam for the purpose of the DTF. Deflection and slope at the root of each fiber were assumed to be zero. In order to obtain evenly spaced strain sensing stations, the locations, strains, and deflections from the FEA model of the UAV were extracted and interpolated at evenly spaced intervals using a polynomial curve fitting function in MATLAB before using in the deflection equations and performing error analysis. Lastly, the half-beam depth was assumed to be equal along the span of the wing. Figure 8.29 shows the comparison of the calculated deflection along the quarter chord of the UAV wing using the FEA strain data and the DTF deflection equations. Thus, the displacement equations accurately predict the shape of the wing for both the right and left side of the wing. Using the results from the FEM, the fiber-optic lines have been implemented onto the Odyssey UAV. Two lines of sensors have been routed along the span of the UAV, and a series of diagonals run in between the two lines. The 2 lines of optic fiber that run in the spanwise direction have approximately 360 strain readings per fiber line and will capture strain due to bending. As shown in Fig. 8.30, the diagonal lines are aligned in the 45ı orientation and will be utilized to obtain strain due to torsion.

8.5

Flight Simulation

8.5.1

X-Plane Simulation Model

By connecting the Odyssey UAV flight control system hardware to the X-Plane flight simulator, real-time hardware-in-the-loop (HIL) simulations were conducted to verify the flight behavior of the design as well as the optimized theoretical gains. The geometry of the UAV was simulated using the Plane Maker software and then uploaded into the X-Plane virtual environment (Fig. 8.31). The aerodynamic

8 Computational and Experimental Design of a Fixed-Wing UAV

137

Fig. 8.30 Fiber optics on odyssey UAV

Fig. 8.31 X-plane simulation model

forces and moments calculated from this blade element theory model is a good representation of the aircraft in real environmental conditions. Specifically, wind speed, gust speed, shear angles, and turbulence were obtained in X-Plane as shown in Fig. 8.32 to analyze the dynamic stability of the UAV design.

8.5.2

Autonomous Design

8.5.2.1 Introduction to Flight Control System If time and resources are not a major concern, designing a custom UAV flight control system (FCS) is a viable option. This allows personalized selection of sensors,

138

H. Ryaciotaki-Boussalis and D. Guillaume

Fig. 8.32 Blade element physics with environmental disturbance

actuators, and processors as well as complete control over the underlying software architecture. The disadvantage of designing a custom FCS is that its design and implementation will have to proceed in pace with the design and the manufacturing of the UAV. Last-minute changes to design, manufacturing, or mission objectives may compromise previous avionic layouts or sensor precision/accuracy requirements in the FCS. In addition, an enormous amount of time is required to code, debug, integrate, and test the finalized FCS first in a simulation environment and then in actual flight. To save time and recourses, an option to custom designing the FCS is the integration of an “off-the-shelf” FCS/autopilot system such as Cloud Cap Technology’s Piccolo Autopilot or the ArduPilot Mega (APM) FCS into the UAV. The advantage of this choice is time and effort saved on design/layout/programming, but the disadvantage is the possibility of experiencing difficulties from software problems or hardware failures of a system that may have no transparency of the underlying hardware and software architectures. The “off-the-shelf” APM FCS was selected over the Piccolo Autopilot for the Odyssey UAV since there is more transparency in the APM hardware and software architectures. Further, the APM is more suited for control since it allows the setting of PID (proportional-integral-derivative) controller gains for the servos based on empirically and analytically derived UAV models.

8 Computational and Experimental Design of a Fixed-Wing UAV

139

8.5.2.2 Aerodynamic State-Space Models Various depending system model implementations may need to be implemented based on the mission requirements of the UAV. For high-speed high-maneuverability aircraft, a nonlinear system model may be necessary. However, for aircraft that may need to have considerably large changes in flight conditions, multiple linear models may be needed with the addition of an active gain scheduler if one controller is incapable of meeting the demands of all expected flight regimes. The plant model of an aircraft can be derived analytically based on the well-known equations of motion and linearized about an operating point. Alternatively, there are existing software packages, such as the AVL that will provide a linear plant model based on design data such as the mass distribution and geometry of the aircraft. The disadvantage of this approach is the unknown underlying assumptions in the creation of such a model. To address the uncertainty in the automatically created plant model delivered by AVL, an alternative approach was used to extract the aircraft’s aerodynamic stability and body-axis coefficients and build new state-space (SS) models under the following assumptions: (1) the longitudinal and lateral degrees of freedom (DOF) are uncoupled, and (2) the aircraft’s motion will mostly consist of small deviations from its equilibrium flight condition. Equations below show the SS structure assumed and the terms calculated based on available sensor data and assuming the FCS board is placed right at the center of gravity (CG) of the aircraft. 3 2 Pu Xu 7 6 Zu 6 w P 7 6 6 4 Pq 5 D 4Mu P 0 2 3 2 u 1 0 6 w 7 6 0 1 6 7 6 4 q 5 D 4 0 0  0 0 2

3 2 Yv Pv 6 Pp 7 6Lv 7 6 6 4 Pr 5 D 4Nv 0 P 2 3 2 v 1 6 p 7 6 0 6 7 6 4 r 5 D 4 0  0 2

0 Lp Np 1 0 0 0 0 0 0 0 0

Xw Zw Mw 0 0 0 0 0

3 2 32 3 Xıe u g 7 6 6 7 0 7 7 6 w 7 C 6 Zıe 7 ıe 4 5 4 5 q 0 Mıe 5  0 0 3 0 u 6 7 07 7 6 w 7 4 5 0 q 5 0 u0 Mq 1 32

1



3 2 3 32 0 Yır v .Yr  u0 / g   7 6 6 7 Lr 0 7 7 6 p 7 C 6 Lıa Lır 7 ıa Nr 0 5 4 r 5 4 Nıa Nıa 5 ır  0 0 0 0 32 3 0 u 6 7 07 7 6 w 7 4 5 0 q 5 1 

The uppercase letters denote aerodynamic derivatives/interactions between subscripted states or control surfaces. The states (u,v,w) represent orthogonal velocity components in a body-centered right-hand system with positive (x,y,z)

140

H. Ryaciotaki-Boussalis and D. Guillaume Root Locus 25 Imaginary Axis (seconds−1)

20 15 10 5 0 −5 −10 −15 −20 −150

−100

−50

0

50

100

Real Axis (seconds−1)

Fig. 8.33 Root locus for u with respect to ıe

measured along the nose, right wing, and undercarriage of the aircraft, respectively. The angular states .p; q; r; ; ; / denote angular rates of change in pitch, roll, yaw, pitch angle, bank angle, and yaw angle, respectively. The elevator, ailerons, and rudder are denoted by ıe, ıa, and ır, respectively. Stability analysis should be performed to determine whether the aircraft has any uncontrollable, unobservable, or unstable modes once the model has been derived. For example, Fig. 8.33 shows the root locus for the longitudinal acceleration .u) with respect to elevator input .ıe) for the Odyssey’s half-scale prototype. Further analysis showed that the SS models are both controllable with only one unstable mode corresponding to the lateral spiral mode (not uncommon for most aircraft). This chapter has shown the design process of a fixed-wing UAV using a case study. At the time of publication, this aircraft was designed, simulated, and constructed. The maiden flight of the Odyssey will be in about 4 months. Acknowledgments The authors want to acknowledge and thank the technical support provided by the NASA Dryden Flight Research Center, all the faculty, and the students who have contributed in the past and present to the design and development of the Odyssey UAV at the Structures, Propulsion, And Control Engineering (SPACE) – a NASA sponsored University Research Center (URC) of Excellence at the California State University, Los Angeles (CSULA).

References J.D. Anderson Jr., Aircraft Performance and Design (Mcgraw Hill, Boston, 1999) A. Derkevorkian, J. Alvarenga, J. Bakalyar, L. Richards, S. Masri, H. Boussalis, Evaluation of strain-based deformation shape estimation algorithms for control and monitoring applications, in 2012 SPIE Symposium on SPIE Smart Structures and Materials C Nondestructive Evaluation and Health Monitoring, San Diego, Mar 2012

8 Computational and Experimental Design of a Fixed-Wing UAV

141

A. Derkevorkian, J. Alvarenga, J. Bakalyar, L. Richards, S. Masri, H. Boussalis, Computational experimental studies of deformation shape prediction algorithms for control monitoring applications, in 5th European Conference on Structural Control – EACS 2012, Genoa, 18–20 June 2012 M. Emmons, S. Karnani, S. Trono, K. Mohanchandra, W. Richards, G. Carman, Strain measurement validation of Embedded fiber bragg gratings. Int. J. Optomechatronics 4(1), 22–33 (2010) Y. Fan, M. Kahrizi, Characterization of a FBG strain gage array embedded in composite structure. Sens. Actuators A: Phys. 121, 305 (2005) B.P. Ferdinand, J.E. Russell, D.T. John, Mechanics of Materials, 4th edn. (McGraw Hill, New York, 2003) W. Ko, W. Richards, Method for real-time structure shape-sensing. U.S. Patent 7520176, pp. 1–13, Apr 2009 W. Ko, W. Richards, V. Tran, Displacement Theories for In-Flight Deformed Shape Predictions of Aerospace Structure. National Aeronautics and Space Administration (NASA), 214612, 2007 K.S.C. Kuang, R. Kenny, M.P. Whelan, W.J. Cantwell, P.R. Chalker, Embedded fiber Bragg grating sensors in advanced composite materials. Compos. Sci. Technol. 61, 1379–1387 (2011) D. Lee, M. Mitrovich, A. Friedman, G. Carman, L. Richards, Characterization of fiber optic sensors for structural health monitoring. J. Compos. Mater. 36(11), 1349 (2002) T.H.G. Megson, Aircraft Structures for Engineering Students, 3rd edn. (Butterworth Heinemann, Amsterdam/London, 1999) T.E. Noll, J.M Brown, M.E. Perez-Davis, S.D. Ishmael, G.C. Tiffany, M. Gaier, Investigation of the Helios prototype aircraft mishap, Vol. I, Mishap Report. National Aeronautics and Space Administration (NASA) I, Jan 2004 D.P. Raymer, Aircraft Design: A Conceptual Approach. AIAA Education Series, 3rd edn. (AIAA, Reston, 1999) W.L. Richards, Characterization of embedded fiber optic sensors in advanced composite materials for structural health monitoring, in Proceedings of SPIE 5390, pp. 505–512, 2004. (About Stability Analysis Using XFLR5 by A. Deperrois – Nov 2010) N.L. Robert, Machine Design An Integrated Approach, 3rd edn (Prentice Hall, Upper Saddle River, 2006) SPACE Center website, http://www.calstatela.edu/orgs/space/ C.C. Thomas, Design of Aircraft (Prentice Hall, Upper Saddle River, 2003) K. Wood, T. Brown, R. Rogowski, B. Jensen, Fiber optic sensors for health monitoring of morphing airframes: I. Bragg grating strain and temperature sensor. Smart Mater. Struct. 9(2), 163 (2000) G. Zhou, L. Sim, Damage detection and assessment in fiber-reinforced composite structures with embedded fiber optics Sensors – review. Smart Mater. Struct. 11, 925 (2002)

9

UAV Handbook: Payload Design of Small UAVs Scott Gruber, Hyukseong Kwon, Chad Hager, Rajnikant Sharma, Josiah Yoder, and Daniel Pack

Contents 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Payload Mission Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 Preliminary Payload Design Budgets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1 The Power Budget. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.2 The Weight Budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.3 Volume Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 Subsystems Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.1 Communications Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.2 Single Board Onboard Computer Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.3 EO Sensor Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.4 Power Subsystem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5 Other Payload Design Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

144 145 146 147 148 149 149 149 155 158 160 162 163

Abstract

This chapter discusses the payload design issues for small unmanned aerial vehicles (UAVs). Details of several payload design principles to overcome various small UAV constraints imposed by stringent weight, power, and volume are discussed. Throughout the chapter, the efficacy of these principles is demonstrated with the help of the payloads for a fixed wing UAV developed by the

S. Gruber () • H. Kwon • C. Hager • R. Sharma • J. Yoder Academy Center for Unmanned Aircraft Systems Research, Department of Electrical and Computer Engineering, United States Air Force Academy, Colorado Springs, CO, USA e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected] D. Pack Department of Electrical and Computer Engineering, University of Texas at San Antonio, San Antonio, TX, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 84, © Springer Science+Business Media Dordrecht 2015

143

144

S. Gruber et al.

Center for Unmanned Aircraft Systems Research at the U.S. Air Force Academy, using this example payload design, a UAV is to be used to autonomously search, detect, localize, and track ground targets. The system requirements for the example application are closely related to those for other small UAV applications.

9.1

Introduction

Clearly, a small UAV has limited payload volume and weight capacity. These limitations drive several additional constraints: electronic components must avoid electromagnetic interference (EMI) interactions, a risk increases by their close proximity; power is limited by the space and weight available for energy storage devices such as batteries; total payload operational time is limited by the available; power processor speed is limited by the weight and power available, and sensor capabilities such as video frame rate and image resolution are limited by available weight allocation. Researchers and developers in academia as well as in industry have worked on designing UAV payloads. Pastor et al. designed a low-cost embedded hardware/software architecture for micro UAVs (Pastor et al. 2007), and Semke et al. suggested an approach to use remote sensing payload design for digital imaging on UAVs for educational development (Semke et al. 2007). Stuckel et al. proposed a payload design for a platform stabilization system to remotely deliver more stable imagery during flights (Stuckel et al. 2011). And Everaerts et al. provided various remote sensing payload designs for visual, infrared, laser, or atmospheric sensors (Everaerts et al. 2004). The focus of this chapter is to design payloads for small UAVs, which can autonomously perform specific intelligence, surveillance, and reconnaissance missions only with an onboard computer and the supporting subsystem payloads. The constraints of volume and weight play critical roles in the following major trade-offs for a small UAV payload: • Autonomy versus human operator • Autonomy (onboard processing) versus available communication bandwidth • Autonomy (onboard capability) versus flight time • Minimalist design versus flexibility and adaptability • Optimal design versus simplicity Designing a small UAV payload involves trade-offs among the basic subsystems shown in Fig. 9.1. Decisions made for one subsystem will affect the design of another subsystem. While managing this balance, the overall small UAV system constraints of weight and volume must be maintained. For example, if an extensive image processing algorithm is executed on images at a rate of five images per second, the single board computer (SBC) needs to handle the computational load. The larger the computer needs are the more power and cooling are necessary to complete the task. Since the payload battery capacity cannot be increased due to the overall weight restriction, the mission time has to be reduced. If mission time is not negotiable, can another choice be made that still accomplishes the goal of

9 UAV Handbook: Payload Design of Small UAVs

145

Fig. 9.1 Major payload subsystems

the mission? Can the image processing rate be reduced from five to one image per second, or can a less computational image processing algorithm be used? These are typical questions that must be answered before a design is chosen. This chapter provides a method to design an integrated payload that maximizes performance in a small UAV package. The chapter is organized to: (a) Define the mission requirements of the payload (Sect. 9.2). (b) Create preliminary weight, power, and volume budgets for the payload (Sect. 9.3). (c) Start an iterative design of the payload subsystems, constantly interacting among all subsystem designs to ensure that performance is maximized within the system constraints (Sect. 9.4). (d) Address other areas that can affect the payload design (Sect. 9.5). Throughout this chapter, examples are provided to help explain the related concepts. There is also an underlying emphasis to ensure EMI does not contribute to the loss of the UAV’s global positioning system (GPS) signal. Most small UAVs rely on GPS signals for navigation, so reliable performance of the GPS system is necessary for the success of a small UAV mission.

9.2

Payload Mission Requirements

Before a detailed design can be developed, the requirements of the payload must be defined. Some questions that should be asked that affect a small UAV payload design are:

146

S. Gruber et al.

• What is the flight duration required for the mission? • What is the specific task of the mission? For example, searching for specific targets, setting up surveillance posts, providing a communications relay, etc. • What are the detailed mission subtasks? How large is the search area? • How long will the payload systems need to be operated on the ground before the flight? • What is the maximum distance from the ground station that the UAVs will be operating? • How many UAVs will be in flight? – If multiple UAVs, how is collision avoidance incorporated? For a small UAV, altitude separation is typically used so that avoidance hardware isn’t necessary. • What are the target characteristics? – Required accuracy of localization – Moving or stationary targets – Number of targets – Target separation – Size, shape, and color of targets • What are the allowed frequencies that can be used for communication between UAVs and the ground station? • What are the situational awareness requirements of the ground station operator? – Imagery refresh rate – Image quality and size – UAV position, mission status, and health • Will other devices be operated in the same area that can cause communication interference? • What are the environmental requirements (vibration, shock, humidity, and temperature)?

9.3

Preliminary Payload Design Budgets

The preliminary payload design is governed by two limiting factors: payload weight and volume. Power, in the form of batteries, is the main contributor to the payload’s weight. Any reduction in power requirements of the payload subsystems will result in either an overall weight reduction or an increase in the possible mission time. Thus, in the flow diagrams used in each of the subsystem design sections, a reference to WPV is used to represent weight, power, and volume. Note that cost is another important factor in developing a design. This is specific to each small UAV application and is not addressed in this chapter. When developing a preliminary design budget, an investigation of likely components needs to be completed. These components can then be used to determine the preliminary allocations of weight and volume. Figure 9.2 is an example payload

9 UAV Handbook: Payload Design of Small UAVs

147

900 MHz GPS

2.4GHz Communication Radio (miniPCI) RS232 Autopilot

E-SATA SSD SBC

Ethernet

EO Sensor

Fig. 9.2 An example of a payload hardware design: SBC (single board computer), EO (electrooptic) sensor, SSD (solid-state disk), and E-SATA (external serial advanced technology attachment)

hardware design including major components and I/O between the components. To start the payload sizing, a preliminary power budget is useful since battery weight is a major weight contributor.

9.3.1

The Power Budget

For a small UAV gasoline propulsion system, having an onboard generator is unlikely due to the weight constraint, so a battery is necessary to support the payload system. If the propulsion system is electric, the propulsion batteries are typically used just for propulsion to maximize flight time, again requiring the payload to have its own battery. Note that if the payload has small power requirements, the small UAV design may opt to use the propulsion batteries to support the payload with the understanding that the flight duration will be impacted. This section will focus on an example payload system with its own battery. An estimated payload is used to compute an overall payload power requirement. Table 9.1 presents an example payload using an atom-based processor, a 2.4-GHz radio, and a security camera as the key payload components. Note that the camera in this example has electronic panning capability, so it addresses all requirements of an electro-optical (EO) sensor system. Since this table is developed early in the design, a 30 % management reserve is added to the power for use if needed as the design progresses. Based on this analysis, the SBC system (processor and disk) has a 14-W budget, the communications system has a 3-W budget, and the EO sensor system has a 3.6-W budget. Note that the current requirement for the wireless card assumes it is transmitting a little more than 50 % of the time and that the solid-state disk (SSD) has 50 % utilization. Based on the mission’s requirements, these average currents can be adjusted. The available battery life is 0.9 h. If the combination of ground time and flight time usage is close to 0.9 h, a higher capacity battery should be employed.

148

S. Gruber et al.

Table 9.1 Example power budget Component Aurora/Corona SBC (http://www. diamondsystems.com/products/aurora/) 32 GB SSD Ubiquiti XR2 wireless card (http://www.ubnt.com/xr2) Axis 212 PTZ camera (http://www.axis.com/products/cam 212/) Misc. parts (servo switch, RC receiver, etc.) Total watts

Voltage

Current

Watts

5

2.0

10.0

5 3.3

0.8 0.9

4.0 3.0

5

0.7

3.6

5

0.2

Adjusted for efficiency of DC-DC converter Adjusted for 30 % management reserve Battery voltage and current for required power Lithium polymer battery amp hours (AH) and available operational hours

Efficiency 0.9 0.7 Batt volt 14.8

1.0 21.6 Watts 24.0 34.3 Needed current 2.3

Batt AH 2.1

Available hours 0.9

Table 9.2 Example weight budget Component

Weight (g)

Percent weight

Aurora/Corona SBC w/enclosure 32 GB SSD, enclosure, and cable Ubiquiti XR2 wireless card, cables, antenna Axis 212 PTZ camera Misc. parts Power system (battery, DC-DC, wires) Total weight Adjusted for 20 % management factor Total weight estimate UAV payload weight allowance Additional weight buffer

400 200 100 500 100 480 1; 780 2; 225 2; 225 2; 270 45

18:0 9:0 4:5 22:5 4:5 21:5

9.3.2

20:0 100

The Weight Budget

The maximum available payload weight, an important design constraint, is predetermined by the aircraft’s lift and gross weight limits. Using the same components used for the power budget, Table 9.2 shows a preliminary weight budget. This example assumes that the small UAV allocation for maximum payload weight is 5 lb. In this case, the management reserve is 20 %. The resulting weight budgets are 27 % for the SBC system (includes the SSD), 4.5 % for the communications system, 22.5 % for the communications system, and 26 % for the power system which includes miscellaneous components.

9 UAV Handbook: Payload Design of Small UAVs

149

Location of batteries

Location of Axis camera Fig. 9.3 Kadet senior UAV

9.3.3

Volume Allocation

When considering the volume available for payload components, it is not sufficient to simply allocate volume for each component. Instead, components must be configured within the payload compartment(s) so that they are placed for proper payload assembly and do not cause any electronic interference to each other. Furthermore, components must be arranged so that the aircraft center of gravity is maintained for stable flight. For example, Fig. 9.3 shows a small UAV (Kadet Senior aircraft, http://www.sigmfg.com/IndexText/SIGRC58.html). To maintain the proper location of the center of gravity, the heaviest objects, the propulsion and payload batteries, were placed in the front of the fuselage, and the electro-optic camera was placed behind the landing gear.

9.4

Subsystems Design

The following sections are introduced with a flow chart that presents the design flow for each subsystem. The number in each flow element corresponds to the items in the accompanying lists that further explain the flow chart elements.

9.4.1

Communications Subsystem

The design flow of the communications subsystem is described in Fig. 9.4. 1. Define mission constraints. (a) Allowed frequencies. Unless licensed radios are available, the UAV will be using the unlicensed ISM (industrial, scientific, and medical) bands. For UAVs, the 902–928-MHz, the 2,400–2,483.5-MHz, and the 5,725– 5,850-MHz bands are commonly used. These bands can provide one to multiple 20-MHz channel bands, and compliant radios are readily available.

150

S. Gruber et al. Define 1 mission constraints Payload 2 constraints on comm. 3 Develop link budgets

Define 4 possible radios 5 Does any radio meet constraints?

Yes

If multiple,6 select radio per WPV

No 8 Negotiate 7 new constraints

No

Is distance too great?

Yes

No

9 Can multi-hop be implemented ?

Yes

Fig. 9.4 Communications subsystem design flow

Transmission power is restricted per FCC regulations; radios and antenna gain need to be selected that comply with these regulations. Note that the 902–928-MHz band is frequently used for the autopilot wireless link. (b) Maximum distance. This distance is used to determine the link budget. (c) Number of UAVs in flight will determine if a multi-hop network is possible and the total bandwidth available at the ground station. (d) Other devices operating in the same area create noise and limit the communication system range if broadcasting over the same frequencies.

9 UAV Handbook: Payload Design of Small UAVs

151

(e) Situational awareness requirements of the ground station operator drive the bandwidth requirements of the system. A number of key questions to be asked are: (i) What command and control messages are necessary and what is the update rate of the messages? (ii) What aircraft status messages need to be reported to the ground station? i.e., aircraft health, stage in mission, etc. (iii) What size and compression factor of images are acceptable? Imagery is the largest consumer of the communication bandwidth. If a communication link is broken during a mission, imagery collected during the lost link may not be useful for the operator. If the imagery is sent using Transmission Control Protocol (TCP), the images have guaranteed delivery. During a lost link, these images are placed in a transmission queue and sent when the link is reestablished. Critical command and control messages need to wait until the queue is emptied of images. If only the current imagery is needed for situational awareness, imagery should be sent using User Datagram Protocol (UDP) which does not guarantee delivery. Thus, the imagery will not sit in a queue, and command and control messages are not delayed when the link is reestablished. (iv) If multiple UAVs are used, how does the imagery need to be presented? To minimize the overall bandwidth, a thumbnail image can be displayed to the operator. The operator then has an option to select a thumbnail to start receiving larger imagery, reducing the overall system bandwidth. Table 9.3 is an example calculation used to estimate the bandwidth requirements. In this UAS configuration, four UAVs converse with the ground station. Only one operator-selected UAV sends high-resolution imagery (320  240 pixels) to the ground station, and all four UAVs send thumbnail images. The “Number/s” column defines how many messages per UAV are allowed, and the “Total

Table 9.3 Estimated communication bandwidth requirements Information Image 320  240 @ 10 % comp Image 80  60 @ 30 % comp Control messages Health messages Ground station messages Other Total Bandwidth efficiency Bandwidth available (Kb/s) 6,000 12,000 24,000

Size (Kb) 20 1 0.1 0.1 0.1 0.1

Number/s 10 4 5 1 2 1

25 % Kb avail 1,500 3,000 6,000

Kb avail 187.5 375 750

Total streams 1.25 5 5 5 5 5

Total Kb/s 250 20 2.5 0.5 1.0 0.5 274.5 Usage 146 % 73 % 37 %

152

S. Gruber et al.

streams” column is based on the number of UAVs in the system plus an increase of 25 % to assume 25 % of the messages are being relayed through another UAV (multi-hop). An assumed 25 % efficiency for the system accounts for message overhead and resends of lost messages. The total message bandwidth required is then compared to available data bandwidths for the 802.11-g protocol. The 6-Mb/s system does not have the capacity and a 12-Mb/s system is required. Since imagery is the largest bandwidth consumer, a reduction in frames per second could get the bandwidth requirement to a more manageable size: 2. Define payload constraints on the communication system. These are constraints generated by the other components of the payload. Some of the questions that must be answered are: (a) If multiple UAVs are used, does one UAV need to communicate with another, or do all UAVs communicate directly with the ground station? If UAVs communicate with each other, is the link between UAVs direct or can the messages be routed through an access point at the ground station? (b) What is the weight allocation for the communication system? (c) What is the power allocation for the communication system? (d) What are the input/output (I/O) interfaces available from the onboard computer? 3. Develop a communications link budget. A communications link budget is a method to estimate how well a communication channel will work for a specific system. It is a good tool for comparing multiple radios, antennas, and frequency bands. This section will not provide a detailed description but gives an example calculation. The link budget is based on the Friis equation (1946): PR D PT GT GR   2 . The power received by a radio, PR , is equal to the power transmitted 4 d by a second radio, PT , that is focused by the gains of the transmit and receive antennas, GT and GR , but is attenuated by the distance the signal has to travel. 2  The last term, 4  d , is usually referred to as the free space loss: the  represents the wavelength of the carrier wave, and d is the distance between the antennas. The equation can be expanded to include additional losses and represented in decibels to simplify the calculations. PRdB D PTdB C GTdB C GRdB  LF SdB  LC bl dB  LP ol dB  LP nt dB , where transmit and receive gains are the maximum gains of the antennas, free space loss .LF S / is defined above, cabling losses .LC bl / are due to the interconnections between the radio and the antenna, polarization loss .LP ol / is due to misalignment of the electric fields, and pointing losses (LP nt ) are due to misalignment of the physical antennas. The actual gain of an antenna varies as measured around the antenna. The typical antennas used on a UAV are dipoles and monopoles which have a gain pattern similar to a doughnut, with the antenna being the axle of the doughnut. These antennas are referred to as omnidirectional due to equal gain in all directions away from the antenna in a plane perpendicular to the antenna.

9 UAV Handbook: Payload Design of Small UAVs

153

0.5 wave antenna with 15 degree roll and 0 degree pitch 350 300 250

meters

200 150 100 50 0 −50 −100 −150 −100

0

100

200 300 meters

400

500

Fig. 9.5 Antenna pointing loss

The design of an antenna generates an alignment of its electric field, referred to as polarization. The most common type used in UAVs is linear polarization, where the electric field aligns with one plane. In the case of a dipole antenna, the field is aligned with the length of the antenna. Thus, for a UAV with a half-wave dipole antenna mounted vertically to the fuselage, the ground station antenna should also be mounted vertically. During flight, the antennas will not always be aligned due to the pitch or roll of the aircraft. This misalignment is referred to as polarization loss and is defined as 20 log10 (cos./), where  is the maximum misalignment angle. For missions where altitude is held constant,  can be approximated by the maximum roll angle allowed by the autopilot. Even if the antennas have no polarization loss, there will still be losses because the aircraft is at a different altitude from the ground station and when the aircraft is rolling away from the ground station. Figure 9.5 is an example of these two conditions. The blue outline is the gain pattern for a 1=2 wavelength dipole antenna, the red line is the line of sight, and the black lines mark the point in the antenna gain pattern where the gain is 1=2 of the maximum pattern gain. Antennas should be selected to ensure that the geometry of the UAV system stays within these 1=2 power or 3 dB points. When maintained, the maximum pointing losses can be assumed to be 3 dB. Let’s demonstrate the use of the link budget through an example. Two radio systems are being compared, a 2.4-GHz radio and a 900-MHz radio. A data bandwidth of 11 Mb/s needs to be maintained, resulting in transmit and receive powers presented in Table 9.4. Other system parameters are also defined in Table 9.4.

154

S. Gruber et al.

Table 9.4 Communications link budget problem definition Parameter 2.4 GHz radio Required receive power for 11 Mb/s (dBm) Transmit power for 11 Mb/s (dBm) Maximum distance (mile) Ground station antenna gain (dBi) Aircraft antenna gain (dBi) Maximum pointing error Maximum bank angle (ı ) Radio to antenna cable losses at each end (dB)

900 MHz radio

92

90

28 1 4 2 Half power beamwidth 15 0.5

28 1 4 0 Half power beamwidth 15 0.5

Table 9.5 Communications link budget example solution Parameter 2.4 GHz radio Power transmit for 11 Mb/s 28.0 Ground cabling loss 0.5 Ground station antenna gain 4.0 Ground pointing loss 3.0 (half power pointing error) Free space loss 104.4 Polarization loss 0.3 Aircraft pointing loss 3.0 (half power pointing error) Aircraft antenna gain 2.0 Aircraft cabling loss 0.5 Power received (total of above) 77.7 Required receive power for 11 Mb/s 92.0 Link margin 14.3

dBm dB dBi dB

900 MHz radio 28.0 0.5 4.0 3.0

dBm dB dBi dB

dB dB dB

95.8 0.3 3.0

dB dB dB

dBi dB dBm dBm dB

0.0 0.5 71.1 90.0 18.9

dBi dB dBm dBm dB

Table 9.5 presents the solution of the link budget. The free space losses are calculated as 2:4 GHz

radio W 2:45 GHz W 20  log10 ..4  1; 609/=0:122/ D 104:4 dB

900 MHz radio W 915 MHz W 20  log10 ..4  1; 609/=0:328/ D 95:8 dB And the polarization loss is calculated as PolarizationLoss W 20  log10 .cos.15ı // D 0:3 dB A good link margin should be above 10 dB. In this case, both radios exceed the link margin. Since both systems have sufficient link margin, other factors should

9 UAV Handbook: Payload Design of Small UAVs

4.

5. 6.

7.

8. 9.

155

be considered such as in-band interferers. If the autopilot wireless link uses the 900-MHz band, there is potential interference for that radio. It would then be better to choose one of the 2.4-GHz channels. Define possible radios. The link budget in Step 3 guides the choices of possible radio solutions. After possible radios are chosen, additional link budget comparisons can be made for final selections. Other considerations when choosing a radio are: (a) What I/O interfaces are available from devices connected to the radio? Typical radio interfaces available are mini-PCI and PCI-Express. There are also radios available for connecting to the USB or Ethernet busses. (b) What antenna connectors are available on the radios? A smaller connector, such as an SMA, is much lighter than the bulky N connector. (c) Select antennas based on required coverage of the mission area. Directional antennas are not feasible on a small UAV because there isn’t a sufficient weight budget to carry a pointing device. Also, the more directional an antenna is, the larger it is. Thus, small UAVs use omnidirectional antennas. On the other hand, the ground station can use a directional antenna if the UAV mission area is away from the ground station such that the UAVs will remain within the 3 dB beamwidth of the antenna. Does any radio meet the communication system constraints? Introduce WPV factors to make the final decision. Reduction in weight and power can either leave more allowance for other payload subsystems or increase flight time. If a solution cannot be found, identify what can be supported and work with the customer to develop new constraints. Multiple design loops are typically required to develop a robust system. Is the only constraint that wasn’t met due to having an insufficient link margin at the required distance? A multi-hop solution may be possible. In a multi-hop solution, messages from a distant UAV are relayed by another UAV. Note that this requires a doubling of bandwidth for any multi-hopped messages since they are sent first by the originating UAV and then again by the relaying UAV. Also, standard 802.11 a/b/g access points do not support multi-hop so an ad hoc network would need to be created.

9.4.2

Single Board Onboard Computer Subsystem

Note that references to the SBC refer to the entire computing system, including storage devices. Figure 9.6 shows the design flow. 1. Define mission constraints. (a) Mission time. Define how long the SBC will be running, both prior to takeoff and during the mission. (b) Environmental requirements affect the selection of the SBC system.

156

2.

3. 4. 5. 6. 7.

8. 9. 10.

11.

S. Gruber et al.

(i) Will the interconnection of the SBC with other systems handle shock and vibration? (ii) Will a hard disk be a viable solution or must a solid-state disk (SSD) be employed? (iii) Do expected humidity and temperature conditions affect the SBC choice? Define payload constraints on the SBC system. (a) What are the computing requirements for processing images? (b) What are the computing requirements for control algorithms? (c) What are the computing requirements for processing sensor outputs? (d) What I/O is needed to interface to autopilot, sensors, and communications systems? (e) What operating system is necessary to support software requirements? (f) What is the weight allocation for the SBC system? (g) What is the power allocation for the SBC system? Develop performance budgets. Table 9.6 is an example that assigns the SBC resources to different software tasks. Look for possible SBC solutions that meet the performance budgets. The solution must also consider storage requirements. Were any SBCs found that meet the requirements? If there were, does the SBC create any GPS noise? If multiple choices exist, select the best choice using WPV factors. Reduction in weight and power can either leave more allowance for other payload subsystems or increase flight time. If no viable computing solution can be found, new constraints must be negotiated with the customer. If an onboard SBC solution can’t be found, is a ground solution viable? Can a solution to the GPS EMI issue be found? (a) In attempts to minimize weight, an SBC without an enclosure may be selected. Without the enclosure, the Faraday cage that contains the EMI is not present and noise may interfere with GPS. (b) Directly connecting an SSD to the SBC via the serial advanced technology attachment (SATA) bus will generate severe GPS noise. SATA generates broadband noise to over 2 GHz. If requiring an external drive, either choose a parallel bus or incorporate SBC and SSD enclosures that support E-SATA and use the shielded E-SATA cable between them. Areas to consider for a ground-based computing solution: (a) A faster computer that doesn’t need to meet weight requirements could be available for ground use. If performing image processing, high-resolution images from the aircraft may be required. Larger images require greater communication bandwidth, which increases with the number of UAVs in use. Can the communications system handle this bandwidth? (b) With all processing on the ground, what happens to the success of the mission if messages are lost?

9 UAV Handbook: Payload Design of Small UAVs

157

Define 1 mission constraints Payload 2 constraints on SBC Develop 3 performance budgets Define 4 possible SBC

5

6

Does any SBC meet constraints?

Does SBC create GPS interference?

Yes

9 new constraints

No

If multiple, 7 select SBC per WPV

Yes

No Negotiate 8

No

Processing at ground station possible?

No

10 Can GPS solution be implemented?

Yes

Yes Ground 11 based solution

Fig. 9.6 SBC subsystem design flow

Table 9.6 Example of aircraft software budget table

Priority 1 2 3 4 5

Component Image capture Planning and control Image processing Data fusion Housekeeping Total

CPU allocation (%) 10 10 45 15 10 100

158

9.4.3

S. Gruber et al.

EO Sensor Subsystem

Though other sensors may be used, the electro-optical sensor is predominantly employed in small UAVs. This section presents design considerations for the EO sensor, but the methods can be readily applied to other types of sensors. The EO sensor subsystem includes the sensor hardware, image capture method, and any required gimbal mechanism for panning the sensor. Note that an inertial measurement unit (IMU) is not listed. For a small UAV, weight considerations limit the use of an IMU dedicated to the sensor. The aircraft attitude can be provided by the autopilot to the image processing system to geo-register the image. Figure 9.7 shows the design flow. 1. Define mission constraints. (a) If multiple UAVs are used, is collision avoidance achieved through altitude separation? Would a camera with variable zoom level be needed so that a common ground footprint is maintained for all altitudes that will be used? (b) Based on the target characteristics and ground coverage, (i) What image processing techniques need to be employed? (ii) Will sensor panning be required? (iii) What image quality and size are needed? (iv) How many images per second are required to process? (c) Environmental requirements. (i) Will aircraft vibration impact image quality and registration? (ii) Will the sensor meet humidity requirements? Will it meet shock requirements? 2. Determine constraints from rest of payload. The EO sensor needs to comply with its allocation of weight, power, and volume, but it also drives the SBC computational requirements. The image processing algorithm selection needs to consider a reasonable SBC capability for a small UAV. Sensor fusion techniques may be employed to account for inaccuracies in the image processing results. 3. Develop a decision matrix for the EO sensor. Table 9.7 presents an example decision matrix with three possible candidates. 4. Define possible EO sensors. Two general types of EO sensors are typically employed: analog and digital. The analog sensor either complies to National Television System Committee (NTSC) or Phase Alternation Line (PAL) formats, with NTSC common in the U.S. and PAL common in Europe. An analog sensor requires an image capture board to convert the image stream to individual images in a common digital format such as Joint Photographic Experts Group (jpeg). Digital sensors provide an image that can be directly processed, usually in a raw format. Each manufacturer of an EO sensor has specifications for their product, but the specifications don’t provide sufficient information to determine if the color quality will meet the project’s image quality requirements. After defining the best candidates, samples should be purchased so that image quality can be evaluated for the project’s application.

9 UAV Handbook: Payload Design of Small UAVs

159 Define 1 mission constraints Payload 2 constraints on sensors Define 3 decision matrix Define 4 possible EO sensors

5 Yes

Does sensor create GPS interference?

Analog

6 What is sensor technology?

No 8 Can GPS solution be implemented?

No

Is analog noise present in image?

7 Method available to acquire images?

No

Yes

9 Yes

Digital

No

Select 10 choice per WPV

Yes

11 Negotiate new constraints

Fig. 9.7 EO sensor subsystem design flow

5. Will the analog sensor generate GPS noise? Though a digital EO sensor should also be checked, NTSC analog sensors use a clock speed that has a harmonic signal at the GPS frequencies. This noise is further increased during the image capture process. PAL analog EO sensors use a different frequency and do not generate the same noise. 6. Determine path between analog and digital. 7. Many digital EO sensors do not directly connect to a common SBC bus such as USB or Ethernet; thus, a field programmable gate array interface must be used. Test interface units are typically available from the manufacturer of the EO sensor though they tend to be bulky and don’t have robust connections. Some complete sensors are available with either Ethernet or USB interfaces.

160

S. Gruber et al.

Table 9.7 Comparison of pan-tilt-zoom camera systems EVI-D100 (http://pro. sony.com/bbsc/ssr/ cat-industrialcameras/ cat-robotic/productEVID100/) Panasonic Sony 1/300 CMOS 1/400 Super HAD CCD 1;280  960 768  494 30 30 HTTP/FTP (digital) NTSC (analog) H.264 JPEG Y/C 0–350ı /30 to 90ı ˙100ı /˙25ı (mechanical) (mechanical) 18 optic + 12 digital 10 optic + 40 digital zoom (3.2–55.2ı zoom (6–65ı AOV) ı (H)/2.4–42.1 (V) AOV) 900 860 115  115  155 113  130  122 12.0 13.2

WV-SC385 (http:// AXIS-212 PTZ (http:// panasonic.net/pss/ www.axis.com/ security/products/hd/ products/cam 212/) wv-sc385/index.html) Manufacturer Image sensor type Pixel resolution Frame rate (fps) Video signal output Pan/tilt Zoom

Weight (g) Dimension (mm3 ) Power consumption (W)

Axis communications 1/200 CMOS 640  480 30 HTTP/FTP (digital) MPEG-4 JPEG ˙70ı /˙52ı (electronic) 3 electronic zoom

504 89  94  77 3.6

8. GPS EMI may be mitigated using a PAL versus NTSC sensor. If the UAV is large enough, separating the analog sensor and capture board from the GPS antenna may sufficiently mitigate the GPS interference. 9. Most analog sensors interlace its image which means that it creates all odd lines in an image and then the even lines of the image. This can create jagged edges since the UAV is moving while the odd lines are captured followed by the capture of the even lines. The performance of edge detection algorithms are impacted by the jagged edges. Other considerations: • If an analog sensor is used and images are transmitted to the ground station for processing, additional noise will be introduced from the analog wireless transmission. • If a greater pan angle is required just for orbiting an object, mounting the sensor at an angle and then always orbiting in the direction that uses the desired tilt of the camera may eliminate the need for a more cumbersome gimbal system.

9.4.4

Power Subsystem

The power system includes batteries and DC-DC converters. Figure 9.8 shows the design flow. 1. Define mission constraints. (a) Flight time will define the amount of battery capacity required for the mission. If the payload operates for a while on the ground before a UAV

9 UAV Handbook: Payload Design of Small UAVs

161

1 Define mission constraints

Payload constraints on power

Develop power budget

Define power solutions

Develop decision matrix

2

3

4

5

7 6

Negotiate new constraints

No

Does a power solution exist?

Yes

Implement power system

8

Fig. 9.8 Power subsystem design flow

takes off, additional capacity is needed. A battery backup switch may be employed which allows an external battery to be connected while on the ground and then removed before take-off. With a properly designed battery backup switch, the payload does not see a drop in voltage when the external battery is removed. 2. Payload constraints on the power system. (a) What is the weight allocation for the power system? (b) As subsystem designs evolve, are additional power requirements identified? 3. While completing the subsystem design, verify the power budget using the technique displayed in Table 9.1. 4. Nickel metal hydride (NiMH) and lithium polymer (li-po) batteries are the most common choices of battery technology in use for UAVs. Table 9.8 presents a comparison of the two types. Though a li-po is more expensive, it is a better choice for UAVs because of the higher power density for the same weight. Special care must be taken with li-po technology to ensure that no overcharging occurs and that the batteries are not discharged to less than 3 V per cell.

162

S. Gruber et al.

Table 9.8 Battery trade analysis Criteria

Nickel metal hydride

Lithium polymer

Cost Rechargeable Volume (cu. in.)/AH Weight (ounces)/AH Charge rate (Amps)

Half of lithium polymer Yes 13.6 18.1 1

Expensive Yes 6.6 6.9 8

5. DC-DC conversion is done with a switching power supply. DC-DC ATX supplies that have been developed for use in automobiles are a good choice. They can use a wide input voltage range and output 3.3-, 5-, and 12-DC voltages that are commonly used by SBCs, sensors, and communication systems. 6. If a power solution is not feasible, a new design needs to be considered with new constraints. 7. If no viable power solution can be found, new constraints must be negotiated with the customer. 8. Decision box regarding existence of viable power solution. 9. Implement the selected power system, always recognizing opportunities to reduce weight and size wherever possible.

9.5

Other Payload Design Considerations

• Flight Time Restrictions: Most small UAV missions will be flown during good visibility and daylight hours so that the aircraft can always be observed in case of erratic behavior. Select sensors to optimize this time of day. Thermal sensors have difficulty identifying targets against warm backgrounds. • Gas engines create greater vibration than electric motors. If a gas engine is required to increase the flight duration, place a vibration reduction mechanism at the engine instead of each of the payload subsystems. • Autopilots and thermal cameras often fall under U.S. International Traffic in Arms Regulations (ITAR). Even though you may not be shipping these items to foreign countries, they also cannot be used by foreign nationals located at the facility. Also, if any software that directly communicates with the drivers of these products is also ITAR controlled, it must comply with the same restrictions. • An autopilot with hardware in the loop (HIL) capability will greatly help in development of control algorithms. The SBC can be directly connected to the autopilot, and the control software can be executed in fight simulators. Conclusion

In this chapter, payload design principles for small UAVs were discussed and proposed. An example based on a system design used by the USAF Academy Center for Unmanned Aircraft Systems Research described various issues and the corresponding solutions for a system with an autonomous ground target detection

9 UAV Handbook: Payload Design of Small UAVs

163

and tracking intelligence, surveillance, and reconnaissance mission. In more detail, mission requirements (budget allocations for power, weight, and volume; and individual subsystem design) were explained with commercial hardware examples.

References J. Everaerts, N. Lewyckyj, D. Fransaer, PEGASUS: design of a stratospheric long endurance UAV system for remote sensing. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. XXXV (Part B), 29–33 (2004) H.T. Friis, Proc. Inst. Radio Eng. 34, 254 (1946) http://www.axis.com/products/cam 212/ http://www.diamondsystems.com/products/aurora/ http://panasonic.net/pss/security/products/hd/wv-sc385/index.html http://pro.sony.com/bbsc/ssr/cat-industrialcameras/cat-robotic/product-EVID100/ http://www.sigmfg.com/IndexText/SIGRC58.html http://www.ubnt.com/xr2 E. Pastor, J. Lopez, P. Royo, UAV payload and mission control hardware/software architecture. IEEE Aerosp. Electron. Syst. Mag. 22(6), 3–8 (2007) W.H. Semke, R.R. Shultz, D. Dvorak, Utilizing UAV payload design by undergraduate researchers for educational and research development, in ASME 2007 International Mechanical Engineering Congress and Exposition (ASME, New York, 2007) K.J. Stuckel, W.H. Semke, N. Baer, R.R. Shultz, A high frequency stabilization system for UAS imaging payloads. Struct. Dyn. 3, 1411–1419 (2011)

Small UAV Design Development and Sizing

10

Steven A. Brandt

Contents 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Configuration Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Fueled Aircraft Sizing and Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.1 Sizing Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.2 Aerodynamic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.3 Constraint Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Mission Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.5 Weight Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6 Electric Aircraft Sizing and Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.1 Sizing Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.2 Airframe Wing Loading Portion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.3 Battery Wing Loading Portion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.4 Motor and Prop Wing Loading Portion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.5 Solving the Sizing Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.6 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.7 Example: The Red Falcon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

166 166 167 168 168 169 171 173 173 174 174 175 176 177 177 178 180 180

Abstract

This chapter describes a rationale for selecting a low-risk UAV (Unmanned Aerial Vehicle) configuration and a methodology for sizing and optimizing the shape of both fueled and electric-powered aircraft. Three examples are given of low-risk UAV configurations. These are shown to be essentially similar from the point of view of optimization and sizing. The sizing methodology for fueled aircraft is described in detail, and a similar one for electric-powered aircraft is

S.A. Brandt Department of Aeronautics, United States Air Force Academy, Colorado Springs, CO, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 83, © Springer Science+Business Media Dordrecht 2015

165

166

S.A. Brandt

described to the extent that it differs from that for fueled aircraft. The chapter also gives typical values of technology-related parameters used in sizing calculations. It then shows an example of a small UAV designed using this methodology. The UAV design project showed that the methodology is useful in obtaining aircraft of minimum size for the performance and payload required.

10.1

Introduction

The two important questions which a small UAV designer must answer about the air vehicle itself are “What will this aircraft look like?” and “How big will it be?” This chapter describes a rationale for selecting a low-risk UAV configuration and a methodology for sizing and optimizing the shape of both fueled and electricpowered aircraft. It then shows an example of a small UAV designed using this methodology.

10.2

Configuration Selection

To answer the question “What will it look like?,” the savvy UAV designer might easily just ask “What do most other small UAVs look like?” Indeed, there is a very Darwinian process in the small UAV industry that tends to eliminate designs that are less useful, have less performance, or are more expensive and more difficult to use. Figure 10.1 illustrates three aircraft configurations commonly used in small UAVs. Note their similarities. Although they do not look exactly alike, the three example UAV designs in Fig. 10.1 all employ a single main lifting wing placed high on the fuselage and a stabilizing tail aft. Since the days of the Wright brothers, this general configuration, as opposed to canard or tailless flying wing configurations, has proven most successful. For reasons why, see Brandt (2004). The three configurations above differ in the details of their aft stabilizing surfaces or empennages. The left-hand one has an inverted T-tail arrangement with the vertical stabilizer and rudder mounted on top of a horizontal stabilizer and elevator. The middle one employs a V-tail which combines the horizontal stabilizing surfaces into a V-shape and combines the rudder and elevators into “ruddervators.” The righthand design adds to the V a downward-projecting vertical surface. Historically, none of these tail arrangements has proven significantly better than the others, so the choice is left as a matter of preference for the designer. The other main differences in the three configurations in Fig. 10.1 are the placement of the motor/propeller and the consequent shape of the rear fuselage. The two left-hand designs place their motor/propellers in the front, while the third places it at the aft end of the fuselage. The aft-facing motor/propeller then requires the extra downward-projecting vertical tail to protect it from hitting the ground on landing.

10 Small UAV Design Development and Sizing

167

Fig. 10.1 Typical UAV designs

In general, aft-facing motor/propellers have more cooling difficulties and result in heavier designs. The decision to use such a configuration would probably result from a design requirement to keep the nose area of the UAV clear for sensors or other mission payloads. If such a requirement does not exist for a particular design, then a forward placement of the motor/propeller is preferred. Neither of the main differences in the typical UAV designs just described has a large effect on the sizing and optimization which answers the second question “How big will it be?” Because aft placement of the motor/propeller usually results in a slightly heavier aircraft structure, it usually has a slight effect on the aircraft size. For initial UAV sizing and optimization, this slight effect can be ignored. Sizing and optimization methodologies differ somewhat between fueled aircraft and those that use electric power. This chapter will describe methods for both types of aircraft.

10.3

Fueled Aircraft Sizing and Optimization

The main difference between fueled and battery-powered aircraft is in the way the aircraft weight changes during the mission. For fueled aircraft, fuel is burned during the mission, so the aircraft’s weight decreases as the mission unfolds. In general,

168

S.A. Brandt

the weight of battery-powered aircraft does not change unless payload is dropped or otherwise released during the mission. Also, the amount of energy which can be extracted from a given mass of fuel is at present much higher than can be extracted from a similar mass of batteries. For these two reasons, fueled aircraft are often the least expensive way to achieve extremely long flights.

10.3.1 Sizing Equation The central equation in this analysis is a simple statement that the sum of the weights of the various components of the aircraft must equal the aircraft total weight: WTO D Wairframe C Wpayload C Wengine C Wfmission When divided by the wing area, the equation expresses the equality between the wing loading of the aircraft and the sum of the portions of that wing loading due to each of the components: Wairframe Wpayload Wengine Wfmission WTO D C C C S S S S S The total wing loading, the left side of the equality, is usually chosen in order to allow the aircraft to meet performance requirements specified by the customer. That analysis is called constraint analysis. The payload is specified by the customer, the airframe and engine wing loading portions are obtained from weight analysis, and the wing loading portion for mission fuel burn is obtained from the mission fuel fraction calculated in mission analysis. The strategy for using this equation to size and optimize the aircraft involves determining values for all but the payload wing loading portion term and then using the weight of the required payload to solve for the required wing reference planform area, S . The resulting equation is called the sizing equation: Wpayload SD (10.1) W Wairframe W WTO  S  engine  fmission S S S

10.3.2 Aerodynamic Analysis All of the analysis that follows depends on having good models for the aerodynamics of the UAV being sized. Methods for predicting aerodynamics of small subsonic aircraft are abundant and have acceptable accuracy. See Brandt (2004) for a simple, accurate aerodynamic prediction method. In many cases, a model of the UAV configuration in question may be tested in a wind tunnel for even better accuracy.

10 Small UAV Design Development and Sizing

169

10.3.3 Constraint Analysis Constraint analysis is a method for defining a solution space such that all design points within that space allow the aircraft to satisfy all of the customer’s performance requirements. The design points are defined by combinations of wing loading, WTO =S and thrust-to-weight ratio, or thrust loading, TSL =WTO , where WTO is maximum takeoff gross weight and TSL is sea-level static maximum thrust, that is, the maximum thrust the aircraft propulsion system can produce at sea level when the aircraft is sitting still on the ground. The equation for constraint analysis is based on a modification of the equation for specific excess power. Power is the rate of doing work, so for an aircraft thrust power is the thrust produced by the aircraft’s propulsion system, T , multiplied by the aircraft’s velocity, V . Excess thrust is the difference between the aircraft’s thrust and its drag, D. Drag is the resistance created by the air to the aircraft’s motion through it. Excess power is the excess thrust multiplied by the aircraft’s velocity and specific excess power is the excess power divided by the aircraft’s weight. Specific excess power, Ps , is a measure of the aircraft’s ability to increase its energy either by climbing or accelerating. All of these ideas are captured in the following equation for specific excess power: Ps D

dh V dV .T  D/V D C W dt g dt

where T is thrust, D is drag, W is weight, V is true airspeed, h is altitude, and g is the acceleration of gravity. Dividing by velocity yields D 1 dh 1 dV T  D C W W V dt g dt

(10.2)

The master equation for constraint analysis is obtained by substituting the following relations into (10.2):T D ˛ TSL , where ˛ is the ratio of available thrust to maximum sea level static thrust. Available thrust is the maximum thrust an aircraft is able to produce at full throttle for a given flight condition. W D ˇWTO , where ˇ is the specified weight fraction for that design requirement. This is usually specified as a value of 1 for takeoff and a value based on full payload and 50 % internal fuel for most mission performance constraints. q D 1=2 V 2 is a quantity called dynamic pressure. In the equation for q the symbol  represents the air density. The drag on the aircraft for a given flight condition is represented by a drag coefficient CD multiplied by dynamic pressure and the reference planform area of the wing, S . Planform area is the area of the shadow of the wing when the sun is directly overhead, including the shadow of the part of the wing that is inside the aircraft’s fuselage. The expression for CD in turn is a quadratic equation called the aircraft’s drag polar. The drag polar has three constants and a variable called lift coefficient, CL . Lift coefficient is related to the actual lift produced by the aircraft in the same way that drag coefficient is related to the aircraft’s drag. The constants

170

S.A. Brandt

in the drag polar are zero-lift drag coefficient CDo , the induced drag factor k1 , and a camber effect term, k2 . When an aircraft is producing zero lift, CL D 0 and CD D CDo . When an aircraft produces lift, normally when it is flying, additional drag results from swirling tornado-like vortices which form at the wingtips. This additional drag-due-to-lift is modeled by the k1 term in the drag polar. Wings with special shaping called camber have less drag for a given amount of lift, an effect modeled by the k2 term in the drag polar. The value of k2 is usually negative while k1 is positive in a typical aircraft drag polar: D D CD qS D .CDo C k1 CL2 C k2 CL /qS CL D

nW L D qS qS

These substitutions produce the “master equation” for constraint analysis: 2



TSL nˇ q CDo  C k1 D 4 WTO ˛ q WTO= S   ˇ 1 dh 1 dV C C ˛ V dt g dt

2 

WTO S



 C k2

3  nˇ 5 q (10.3)

The master equation is written in a form that expresses the minimum TSL /WTO which can achieve a given performance requirement at a particular value of WTO =S . A separate form of the master equation is written using values of velocity, load factor, acceleration, rate of climb, etc., for each design requirement specified by the customer. Lines representing these equations are plotted as boundaries on a thrustto-weight ratio vs. wing loading diagram. Combinations of design thrust-to-weight ratio and wing loading which fall outside the boundary specified by each constraint result in an aircraft design which will not meet that design requirement. When all the constraint boundaries are plotted, they define a solution space which, for all combinations of thrust-to-weight ratio and wing loading within the space, allows the aircraft to meet all the design requirements. Figure 10.2 illustrates a typical constraint diagram. Ideal and actual design points are shown by the circle and plus sign on the plot. The design point chosen on the constraint diagram specifies the value of WTO =S to be used in the sizing equation. It also specifies the value of TSL =WTO which, with a little manipulation, produces the value of Weng i ne =S required to solve the sizing equation. This manipulation requires only a knowledge of the technology level of the engine, as expressed by the thrust-to-weight ratio of the engine itself, TSL =Weng i ne . With this value in hand, Wengine =S is calculated as follows: Wengine D S

TSL WTO WTO S TSL Wengine

10 Small UAV Design Development and Sizing

171

4.5 MxMach

Thrust Loading, TsI/Wto

4

LoTurn

3.5

ABPs

3

MilPs

2.5

Cruise

2

HiTurn SloTurn

1.5

MilTurn

1

Takeoff

0.5

Landing

0 220

200

180

160

140

120

100

80

60

40

20

0

Actual Design Point Desired

Wing Loading, Wto/S, psf

Fig. 10.2 Typical constraint diagram

10.4

Mission Analysis

Mission analysis uses the aircraft’s drag polar and the design point determined from constraint analysis, plus an appropriate model for engine thrust-specific fuel consumption, to predict the amount of fuel the aircraft will burn on its design mission(s). The master equation for mission analysis is essentially the time integral form of the constraint analysis master equation substituting actual thrust for thrust available and multiplying by thrust-specific fuel consumption, c, to yield fuel flow rate: cT D WP f The minus sign in front of fuel flow rate indicates that aircraft weight decreases as fuel is burned. The integral of fuel flow rate yields fuel burn which, divided by weight, is the fuel fraction: Zt1 t0

cT dt D WTO

Zt1 t0

Zt1 D t0

WP f dt WTO 2 CDo  C k1 cq 4  WTO= S 

Zt1 cˇ

C t0



1 dV 1 dh C V dt g dt

nˇ q

2 

 dt

WTO S



 C k2

3  nˇ 5 dt q

172

S.A. Brandt

which is easily integrated for the most common cases of constant velocity climbs and cruise, constant altitude accelerations, and for relatively short mission legs where it is acceptable to assume the aircraft’s drag polar and weight fraction are approximately constant: Wf W0  W1 D D WTO WTO 3 2   2    nˇ W nˇ C Do TO 5 .t1 t0 /  C k1 C k2 D cq 4  q S q WTO= S   1 1 .h1  h0 / C .V  V0 1/ C cˇ (10.4) V g Equation (10.4) is known as the master equation for mission analysis. Note that WTO =S in (10.4) is, for this analysis, a constant, since it was chosen as a result of constraint analysis. The assumption of constant ˇ is not acceptable for long cruise or loiter legs, and the assumption of constant drag polar is not acceptable for transonic accelerations, but these are easily accommodated in computer implementations of the method by breaking long legs into many similar short legs. When the mission analysis master equation is applied to each leg of a typical design mission, the fuel fractions from all preceding legs of the mission are used to determine the weight fraction for each subsequent leg. When the fuel fractions for each leg are summed, the total fuel fraction for the mission is obtained: ˇtakeoff D 1

.since W D WTO /

ˇaccel D ˇtakeoff  ˇc lim b D ˇaccel 

Wftakeoff WTO

Wfaccel WTO

ˇcruise D ˇc lim b 

Wfclimb WTO

ˇloiter D ˇcruise 

Wfcruise WTO

Wftakeoff Wfaccel Wfclimb Wfcruise Wfloiter Wfmission D C C C C WTO WTO WTO WTO WTO WTO Wfmission =WTO is known as the design mission fuel fraction. It is multiplied by the design point wing loading to yield the value of Wfmission =S needed to solve the sizing equation: Wfmission WTO Wfmission D S WTO S

10 Small UAV Design Development and Sizing

10.5

173

Weight Analysis

Once mission analysis has predicted the mission fuel burn and fuel fraction, weightestimating methods are used to predict the structural weights per unit area of the airframe components. For a small UAV, it is often adequate to weigh a similar UAV airframe to obtain approximate weights per area for the various airframe components. Whether predicted or measured, the component weights per unit area are then scaled by the ratios of their component reference areas to the UAV reference planform area and summed to obtain the airframe wing loading portion needed to solve the sizing equation. Once a value of S is determined, it may be multiplied by the design point wing loading to obtain the sized takeoff gross weight:   WTO (10.5) WTO D S S The results obtained in (10.1) and (10.5) are final, provided some other size or volume constraints, such as the dimensions of the payload, do not force redesign of the aircraft configuration. The entire aircraft is just scaled up or down as necessary to achieve the specified wing area. It is more common, however, that the fuselage of a small UAV is sized by payload dimensions, so that when the aircraft is sized up or down, the fuselage size remains unchanged. In this common situation, after resizing, the entire analysis must be accomplished again. This is because resizing the wing and tail without resizing the fuselage changes the aircraft’s drag polar. Since drag polar is one of the required inputs to constraint analysis, constraint analysis must be reaccomplished, yielding new values for wing loading and thrust loading. A change in the configuration also changes the value of the airframe wing loading portion, and a new solution to the sizing equation results. The process must be repeated iteratively until a converged solution is obtained. Convergence may require human intervention, since the numerical scheme is not necessarily stable. In addition, the savvy UAV designer will explore variations of the configuration to find the optimum design, that is, that design which can meet all the customer’s requirements for the lowest cost.

10.6

Electric Aircraft Sizing and Optimization

The use of electric power in small UAVs has increased significantly as cost and weight of batteries have decreased. The advantages of batteries over chemical fuels for UAV storage, transportation, and deployment are obvious. As battery performance continues to improve, a larger proportion of UAVs will likely use electric power.

174

S.A. Brandt

10.6.1 Sizing Equation The sizing methodology for electric-powered aircraft is based on a similar analysis for solar-powered aircraft described in Brandt and Gilliam (1994). As in the method for fueled aircraft, the central equation in this analysis is a simple statement that the sum of the weights of the various components of the aircraft must equal the aircraft total weight. When divided by the wing area, the equation expresses the equality between the wing loading of the aircraft and the sum of the portions of that wing loading due to each of the components: WAirframe WBattery WMotor& Pr op WPayload WTakeoff D C C C S S S S S The strategy in using this equation to size and optimize an electric aircraft is identical to that used to obtain Eq. (10.1) for fueled aircraft and the resulting equation is similar: SD

Wpayload WTO S



WAircraft S



Wbattery S



Wmotor S

(10.6)

As for fueled aircraft, total wing loading is usually chosen in order to allow the aircraft to meet performance requirements specified by the customer, commonly represented on a constraint diagram such as Fig. 10.3. The details of constraint analysis for electric aircraft are identical to those for fueled aircraft, except that the thrust lapse, ˛, for electric aircraft behaves somewhat differently. From a constraint diagram the designer selects values for wing loading and thrust or power loading which will allow the aircraft to meet the customer’s needs. In the case of Fig. 10.2, the selected values are WTO =S D 1:1 and T =WTO D 0:21. The thrust values graphed in the constraint diagram are specified as those at static sea level conditions. Wind tunnel tests of the motor and propeller may be used to correct the thrust required for each of the constraints to sea level static values for plotting.

10.6.2 Airframe Wing Loading Portion As with fueled aircraft, determining a value for the airframe wing loading portion involves determining representative values for similar aircraft and estimating what is achievable with current and future structures and materials technology. Since weights of aircraft components typically vary with their areas, this airframe wing loading portion, once calculated for a given aircraft configuration, will not vary significantly as the aircraft is sized up or down. A nominal value for airframe wing loading portion for a conventional configuration small UAV constructed of carbon fiber and epoxy composite skin over a Styrofoam core is 0:2 lb=ft2 .

10 Small UAV Design Development and Sizing

175

Fig. 10.3 Example small UAV constraint diagram

10.6.3 Battery Wing Loading Portion The weight of the battery depends on the level of battery technology, represented by its energy density, e, the efficiency of the motor and propeller, motor and prop , and the total thrust energy which the motor and propeller must provide to overcome the aircraft’s drag and to increase the aircraft’s mechanical energy by climbing or accelerating. The total thrust energy required will therefore depend on the details of the design mission profile. Figure 10.4 illustrates the mission profile for a small backpack UAV. For this mission, the only change in energy occurs immediately after launch, when the aircraft climbs 500 ft. above its launch altitude. Throughout the mission, the aircraft will fly at an approximately constant speed, and since its weight will not change, neither will its drag. As a result, the total thrust energy required for the mission is given by E D WTO h C DVt (10.7) where h is the change in altitude during the initial climb, D is the drag throughout the mission, V is the velocity, and t is the total mission duration. The required weight of the battery is the total energy which must be provided by the battery divided by the battery’s energy density. This energy provided by the battery must be enough so that after losses in the motor and propeller, modeled by efficiency factors for these devices, there is useable thrust energy equal to that required by (10.7).

176

S.A. Brandt

Cruise at Best Endurance Speed Launch & climb to 500 ft AGL

Descend for Recovery

Cruise back at Best Endurance Speed

Reconnaissance Loiter

Fig. 10.4 Backpack UAV general mission profile

If, as is the case in the mission profile of Fig. 10.3, the energy to climb is small compared to the total energy required for the mission, then the weight of the battery required is approximated by WBattery D

DVt eMotor prop

(10.8)

Dividing (10.8) by the aircraft weight, WTO , which is equal to the required lift, L, yields WBattery D L e Vt  (10.9) Motor prop WTO D Note that the ratio L=D in (10.9) is a common means of describing the aerodynamic efficiency of an aircraft and is a function of the aircraft configuration and flight conditions. For the simple mission profile being considered, L=D will be essentially constant throughout the mission. Multiplying (10.9) by the total wing loading chosen from constraint analysis yields WBattery WTakeoff WTakeoff WBattery D D  S WTakeoff S S

Vt

L D eMotor prop

(10.10)

10.6.4 Motor and Prop Wing Loading Portion Finally, the motor and propeller wing loading portion is estimated using the thrust loading and wing loading obtained from constraint analysis and an estimate of available motor and propeller technology expressed as a motor and prop thrust-toweight ratio: WMotor T WMotor WTakeoff D  (10.11)  S WTakeoff T S The motor and prop thrust-to-weight ratio, along with values for motor and prop , may be obtained by testing several typical motor/prop combinations in a low-speed wind tunnel. Figure 10.5 illustrates the results of a typical test. It plots the variation of motor and propeller efficiency for the best motor/propeller combination tested, as a function of free-stream velocity.

10 Small UAV Design Development and Sizing

177

Motor/Propeller Efficiency as a function of Velocity

Efficiency (-)

0.8 0.6 0.4 0.2 0 0

20

40 Velocity (ft/s)

60

80

Fig. 10.5 Variation of efficiency with forward velocity for typical small UAV electric motor and propeller

10.6.5 Solving the Sizing Equation With all terms evaluated, Eq. (10.6) is solved for planform area. As with fueled aircraft, this result is then used in Eq. (10.5) to solve for sized takeoff gross weight. The results are final provided payload volume or some other size constraint does not force a change to the configuration. In the common case where the fuselage must remain a constant size while the wing is sized up or down, the resulting new configuration will require an updated drag polar and weight model. This in turn will require a complete new iteration through the sizing method. Frequently, several iterations are required to reach a final solution. This numerical scheme may be unstable, so human intervention may be required to enforce convergence. In general, adjusting the wing area to an average between that modeled in the previous iteration and that recommended by the new solution to the sizing equation will lead to faster convergence. In the extreme case, the analysis may not converge. This typically happens when specified mission duration is too long. When this occurs, radical changes to the configuration or radical new technologies like increased battery energy density may be required to meet the requirements.

10.6.6 Optimization The configuration modeling, aerodynamic analysis, constraint analysis, and sizing cycle just described can be repeated for a variety of aircraft configurations. If configuration parameters such as wing sweep, aspect ratio, airfoil thickness, and camber are varied systematically, an optimization carpet plot such as Fig. 10.6 can be obtained. From such a plot, an optimum wing planform or other configuration characteristic can be selected, which minimizes the aircraft size while still meeting all customer requirements.

178

S.A. Brandt 14

Wing Area (ft^2)

12 AR=2 AR=3 AR=4 AR=5 AR=6

10

8

6

4

2 3.5

4

4.5

5

5.5

6

6.5

7

Takeoff Weight (lb)

Fig. 10.6 Typical wing optimization carpet plot

10.6.7 Example: The Red Falcon Results of a recent UAV design effort at the United States Air Force Academy (USAFA) will illustrate how the method may be applied to a typical UAV design problem. The example design problem required a UAV small enough to fit in a backpack and when deployed able to carry a small spy camera, as depicted in Fig. 10.7. This camera, with its built-in downlink transmitter and its power supply weighed approximately 100 g or slightly more than 3 ounces. The design team selected rechargeable lithium-ion batteries, with an energy density of 68 W-h/lb, as their power source. The team measured motor and propeller efficiencies during wind tunnel testing. They also performed wind tunnel tests of several candidate aircraft configurations including pusher, canard, and flying wing variations but in the end settled on a conventional tractor propeller, tail aft configuration with an unswept wing with an aspect ratio (AR) of 4. Figures 10.8 and 10.9 illustrate the wing sweep and aspect ratio optimization plots for this configuration. In the end, an aspect ratio of 4 was selected for ease of packaging in the backpack and because the decrease in size for higher aspect ratios was minimal. The above design decisions led to an aircraft configuration which sized to a wing area of slightly over 1 ft2 and a maximum design flying weight of slightly over 1 lb. The aircraft, named the “Red Falcon,” is shown in Fig. 10.10. This aircraft flew a number of flight test missions to verify performance predictions and flying qualities. Although it never went into production, it provided a sizing example for a number of subsequent UAV developments. It served to verify the validity of the methodology described in this chapter and the practicality of using a small UAV for this design mission.

10 Small UAV Design Development and Sizing

179

Power Required, Preq (W)

Fig. 10.7 Daylight spy camera and downlink

40 39 38 37 36 35 34 33 0

5

10

15

20

25

30

Leading Edge Sweep (deg)

Fig. 10.8 Wing sweep optimization plot 2.2 W/ 4ft/s Climb @ V=50ft/s W/o 4ft/s Climb @ V=50ft/s

Wb (lb)

2 1.8 1.6 1.4 1.2 1 0

1

2

3 AR

Fig. 10.9 Wing aspect ratio optimization plot

4

5

6

180

S.A. Brandt

Fig. 10.10 The Red Falcon

10.7

Conclusion

This chapter has presented a rationale for the UAV designer to choose an aircraft configuration and a methodology for sizing it. By iteratively changing such aircraft configuration parameters as wing aspect ratio and camber, the designer can determine the smallest/cheapest UAV that can meet the design mission. This process is one way to optimize UAV designs for maximum performance and minimum size/cost.

References S.A. Brandt, F.P. Gilliam, A design analysis methodology for solar-powered aircraft. J. Aircr. 766–780 (1994) S.A. Brandt et al., Introduction to Aeronautics: A Design Perspective. AIAA Education Series, 2nd edn. (AIAA, Reston, 2004) J.R. Smith, P.G. Batish, S.A. Brandt, S.R. Morton, A student-developed sizing methodology for electric-powered aircraft applied to small UAVs, in WAC Paper 2000-01-5536, Proceedings of the 5th AIAA/SAE World Aviation Congress, San Diego, 10–12 Oct 2000

Systematic Design Methodology and Construction of Micro Aerial Quadrotor Vehicles

11

Swee King Phang, Kun Li, Ben M. Chen, and Tong H. Lee

Contents 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Virtual Design Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2.1 MSC Patran and Nastran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2.2 SolidWorks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3 Hardware Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3.1 Motor and Propeller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3.2 Electronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3.3 Power Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3.4 Avionics Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4 Dynamic Modeling and Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.1 Nonlinear Mathematics Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.2 Parameters Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.3 Controller Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.5 Flight Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

182 183 184 188 190 190 191 193 193 194 194 198 200 200 202 205

Abstract

This chapter presents a guideline to systematically design and construct an ultralight micro aerial quadrotor vehicle, capable of autonomous flight. The guideline details the steps to design a stable quadrotor with not more than 50 g takeoff weight while having a flight duration of 8 min. Optimization was done to all mechanical parts of the vehicle with the constraints of its weight and

S.K. Phang • K. Li Control and Simulation Lab, National University of Singapore, Singapore e-mail: [email protected]; [email protected] B.M. Chen • T.H. Lee Control and Simulation Lab, Department of Electrical and Computer Engineering, National University of Singapore, Singapore e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 116, © Springer Science+Business Media Dordrecht 2015

181

182

S.K. Phang et al.

its performance. The guideline first covers the structural analysis of the micro quadrotor air frame, followed by its design in a 3D virtual simulator. Then, design of avionic system such as processor and sensors will be discussed, followed by controller implementation with the aid of software simulation. Components of the micro quadrotor are then fabricated and implemented based on the optimum results obtained from various simulations. Finally, the feasibility of the constructed aircraft is tested and confirmed with real flight missions.

11.1

Introduction

With more requirements imposed on unmanned aerial vehicle (UAV) for military tasks of surveillance, reconnaissance, and detecting in obstacle-rich areas, clandestine military bases, radiant areas, and other dangerous regions, the desire and interest in research of micro aerial vehicle (MAV), or even nano aerial vehicle (NAV), are aroused (Wang et al. 2011). These requirements involve higher intelligence for simultaneous localization and mapping (SLAM), path planning, and obstacle avoidance, along with battery technology to maintain long endurance flight and communication with the ground control station (Achtelik et al. 2009; Eresen et al. 2012; Meier et al. 2011; Phang et al. 2010). In the recent few decades, the development of integrated circuits and micro-electromechanical systems (MEMS) triggered the emergence of smaller and lighter electronics and mechanical systems. It provides the possibility to shrink the size and weight of the UAV to a new milestone. In 1997, the Defense Advanced Research Projects Agency (DARPA) sets an official definition for MAV, which requires maximum dimension of the aircraft in any direction to be no greater than 15 cm and that the gross takeoff weight should not exceed 100 g, with up to 20 g devoted to payload. Based on these requirements, many UAV research groups initiated their MAV study with different design approaches, including fixed-wing, rotary-wing, flapping-wing, and other unconventional platforms (Al-Qadi et al. 2006; Michelson 2010; Petricca et al. 2011). One good example would be the Black Widow, an 80 g fixed-wing MAV developed by AeroVironment, which has a 15 cm wingspan and is claimed to be fully autonomous (Grasmeyer and Keennon 2011). It is, however, difficult for the fixedwing platform to show its capability in indoor environments or cluttered outdoor environments such as forest and urban environment full of building and other obstacles, due to its limitation of hovering and vertically taking off and landing (VTOL) ability. Flapping-wing is a promising platform with its VTOL and hovering capabilities like that of an insect or a bird. In 2007, a 3 cm size flapping-wing MAV was developed by Harvard University, biologically mimicking the dipteran insects (Wood 2007). Unfortunately, there is no mature development of autonomous flapping-wing platforms due to the sophisticated flapping mechanism and wing structure aerodynamic analysis. Ducted-fan platform is also a possible approach for miniature aerial vehicle. A 150 g ducted-fan platform was developed by University of Bologna. This platform requires extra fins to control its attitudes, resulting in a large takeoff weight (Naldi et al. 2010).

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

183

Recently, due to its extreme stability and mechanical simplicity, quadrotor platforms have aroused the research interests of many researchers. There are, however, not many quadrotor MAV platforms lower than 50 g capable of autonomous flight. A research group from Stanford University has developed a micro quadrotor “Mesicopter”, with an impressive weight of 3 g. It is, however, unable to take off due to its rotor efficiency problem (Kroo and Kuns 2001). Mellinger and his collaborators from GRASP lab have been working on the formation flight of quadrotor MAVs weighing 60 g each and have achieved impressive results (Mellinger et al. 2010, 2012). From the perspective of the platform approach, rotary-wing platforms, especially in the form of quadrotor, have the advantages in maneuverability and mechanical feasibility. In this chapter, a guideline to design a micro aerial quadrotor vehicle, or in short quadrotor MAV, with the following specifications will be presented: 1. Target weight not higher than 50 g 2. Largest dimension smaller than 15 cm in any direction 3. Flight duration of 8 min Systematic design procedures with scopes of the design blueprint in virtual 3D environment, functionalities analysis in simulations, hardware selection and avionic system design, electrical circuit and software debugging, mathematical modeling and parameters identification, model simulation and verification, as well as aircraft orientation control with real flight test will be discussed in detail. In regard to this work, great effort was made to investigate the feasibility of overall structure and to evaluate the platform from the aspects of weight, size, and power consumption with both experiments and simulations before manufacturing, which guarantees a flyable and stable platform. Main contents of this chapter are divided into five sections. Section 11.2 is the virtual design of the overall platform, with respect to mechanical structure and natural frequency analysis. Section 11.3 describes the hardware selection and avionic design procedures. Section 11.4 is the nonlinear mathematics model derivation concerning the aspects of kinematics, rigid body dynamics, and motor dynamics, followed by parameters identification methods and controller design. In Sect. 11.5, real flight test data is used to verify the mathematical model and control laws.

11.2

Virtual Design Environment

Before constructing the quadrotor platform, the design was done in virtual design environments. Differing from the usual UAV in larger scale, the natural modes of the structure or platform play an important role in determining the stability of the quadrotor MAV. As the aircraft gets smaller and lighter, the effect of its vibration due to natural modes is disastrous. Therefore in this section, two design processes in developing small-scale quadrotor MAV are proposed to systematically determine the optimum shape for the platform, in terms of lightweight and natural mode avoidance. The first process is to perform finite element analysis (FEA) on various platform shapes. FEA is a numerical approach in solving structural problems by discretizing

184

S.K. Phang et al.

the problem where the domain to be analyzed is divided into a number of discrete elements. Using this approach, it offers great flexibility to model complex geometries which would be near impossible if analytical approach is taken. By undergoing FEA, the natural modes and frequencies of the platform can be estimated numerically. With this, a best possible candidate can be chosen from among the designs. Upon obtaining an optimum structural shape of the platform according to the FEA results, the platform of the quadrotor MAV can then be designed with a 3D simulation software, where each individual part of the aircraft could be assembled in the software for an overall design. An advantage of utilizing simulation software is that it could estimate various important parameters of the designed aircraft, such as the weight, density, and the moment of inertia, in a way that the design could be customized freely in order to meet their requirements before they are fabricated to real physical objects.

11.2.1 MSC Patran and Nastran MSC Nastran, one of the most widely used FEA commercial software, is utilized in simulating the vibration properties of the quadrotor MAV. It is a useful tool which simulates static and dynamic cases for a wide variety of complicated structural problems. In addition to Nastran, a modeler software called Patran is employed for finite element modeling. Patran provides geometry modeling, meshing, boundary condition, and material properties set up for Nastran. It is also used for postprocessing purpose. Provided in Nastran, Lanczos (1950) method is utilized for eigenvalue and eigenvector extraction purpose. To investigate and design the quadrotor platform, single arm and full quadrotor models are constructed in Patran with the properties of carbon fiber composite material assigned to the software. The model is then fetched to Nastran to investigate its structural resonance. In corresponding specifically to the design of micro aerial quadrotor, the study is concentrated on natural mode analysis to determine the natural frequencies and mode shapes of the model. The natural mode analysis predicts the resonance for the structure and the type of resonance. On the other hand, the displacement response of the model can be determined using frequency response analysis when external steady-state oscillatory excitation (simulating rotor rotation) is applied.

11.2.1.1 Single Quadrotor Arm A quadrotor consists of four extended arms attached to a body which holds the onboard electronics. The aim of carrying out the single quadrotor arm analysis is to provide a useful summary on the performance of different quadrotor arm shapes, and then among the few possible candidates under the constrained of weight and size, a best solution in terms of shape, length, and dimension can be obtained. In this analysis, the quadrotor arm is approximated as a cantilever beam which has a fixed end and a free end.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

185

Fig. 11.1 Cross sections of five common beams

Five different types of cantilever beams are designed and analyzed. The cross sections for each of the beams are shown in Fig. 11.1. A few different parameters on the dimension of the beams will be varied, and the corresponding natural mode (1–4) will be compared. Note that the composite material for all beams to be simulated has the properties of carbon fiber code name carbon/epoxy T300/976. All simulations are done in Nastran by employing the discrete element with six degrees of freedom (DOF) per node, and all nodes at one of the tips are assumed to be fixed by setting all six DOF to zero. The first variable to be investigated is the length of the beam. It is well known that slender bodies are more easily exposed to vibration, or in other words, shorter beams are stiffer. To verify the relationship between the length of the beam and its natural modes, the rectangular beam (first cross section in Fig. 11.1) with different length is analyzed in the simulation. In Table 11.1, a summary of the Nastran natural mode analysis results obtained for this study is given. Based on the results, it is evident that natural frequencies for the first four modes increase as the structure becomes shorter. In general, beams of other shape show similar behavior, and thus the results are trivial and not to be included here. Despite the results favoring shorter beams, there are other restrictions on the minimum length of the quadrotor arms. One important factor would be the aerodynamic interferences between the rotors, which generally limits the minimum length of the quadrotor arms to be at least twice the rotor radius.

186 Table 11.1 Natural frequencies of thin plate with varying length

S.K. Phang et al.

Length (mm)

Natural frequency (Hz) Mode 1 Mode 2

50 60 70 80 90 100

614.6 426.85 313.62 240.3 189.73 153.68

3; 656:7 2; 546:1 1; 873:6 1; 437:1 1; 135:4 920:14

Mode 3

Mode 4

3; 803:4 2; 642:6 1; 942:1 1; 491:3 1; 175:3 952:05

10; 518 7; 313:4 5; 377:1 4; 137:1 3; 255:4 2; 637:5

Table 11.2 Natural frequencies of beam with different cross sections and thickness 1 mm thickness 0.5 mm thickness Type of cross section Weight (g) Mode 1 (Hz) Weight (g) Mode 1 (Hz) Rectangular 0.5328 426.85 0.2664 213.45 Rectangular hollow 1.776 3,032.4 0.9768 3,270.7 Circular hollow 1.3949 2,647.9 0.767 2,851.6 T shape 0.9768 1,902.8 0.5106 1,845.3 N shape 1.4208 2,717.5 0.7548 2,810.4

The second variable to be investigated is the thickness of the material. As two common thickness of carbon fiber sheet or beam available commercially are of 0.5 and 1 mm thick, the natural mode of all five general shapes of beams are analyzed and compared in 0.5 and 1 mm thickness. For fair comparison, the model for each cross section was constructed with length l D 60 mm, width w D 6 mm, and height h D 6 mm (see Fig. 11.1 for illustration). Table 11.2 shows the comparison between beam with thickness t D 1 mm and t D 0:5 mm, together with their calculated weight. One can notice that in general, the beam having a closed cross-sectional shape, i.e., the rectangular hollow and circular hollow beams, has a much higher natural frequency compared to other shapes. Also, the thickness of the material has little effect on them, while the weight could be half as light. The last variable to be investigated is the width and height of these closed crosssectional beams. It can be seen from Table 11.3 that the height variation of the rectangular hollow beam affects the first mode frequency more severely than the second mode frequency. On the other hand, the width variation of the beam affects the second mode frequency much more than the first mode frequency. In conclusion, utilizing a square beam with equal height and width will give similar frequencies for the first and second modes. This is further proven in Table 11.4 where the circular hollow beams were analyzed. In general, although the rectangular solid and T-shape configurations show better weight budget, their natural frequencies are comparatively much lower than that of rectangular hollow and circular hollow shapes, which are in the closed cross section forms. In the following section, a full quadrotor configuration with rectangular hollow beam of 3 mm height and 3 mm width will be investigated.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor Table 11.3 Natural frequencies of the rectangular hollow beam with different width and height (0.5 mm thickness)

Table 11.4 Natural frequencies of the circular hollow beam with different radius (0.5 mm thickness)

187

Height (mm)

Width (mm)

Natural frequency (Hz) Mode 1 Mode 2 Mode 3

Mode 4

3 3 3 3 3 3 4 5 6

5 6 7 8 9 10 6 6 6

1,625.9 1,651.8 1,670.3 1,683.5 1,693.0 1,699.4 2,218.0 2,756.0 3,270.7

14,904 17,415 19,762 21,954 23,571 22,983 17,908 18,185 18,312

2,515.8 2,979.7 3,430.8 3,871.3 4,302.7 4,725.6 3,105.6 3,199.4 3,270.7

9,673.4 9,718.5 9,708.0 9,656.9 9,576.0 9,472.4 12,845 15,699 18,312

Radius (mm)

Natural frequency (Hz) Mode 1 Mode 2

Mode 3

Mode 4

2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5

1,839.3 2,347.5 2,851.6 3,349.6 3,840.1 4,322 4,794.3 5,256.1

11,070 13,875 16,495 18,911 21,117 23,116 24,918 26,538

11,070 13,875 16,495 18,911 21,117 23,116 24,918 26,538

1,839.3 2,347.5 2,851.6 3,349.6 3,840.1 4,322 4,794.3 5,256.1

11.2.1.2 Full Quadrotor Configuration Due to structure stability shown by closed shape beams, the full configuration is then tested using rectangular hollow beam and circular hollow beam to form the arms attached to each corner of the main frame. The main frame is designed to be 28:28  28:28 mm with thickness 2 mm. Meanwhile, 60 mm long beam is employed as the quadrotor arm. Again, unidirectional carbon/epoxy T300/976 is used. In Patran, the main frame is modeled as 2-D shell element while the quadrotor arm using 1-D beam. For this analysis, the main frame is assumed to be rigid and thus fixed in all three translation DOF. For a quadrotor model with 3 mm width, 3 mm height, and 1 mm thickness rectangular hollow beam, the resulted natural frequency for the first and second modes are 971.92 and 5,843.3 Hz. On the other hand, the natural frequency for quadrotor model using circular hollow beam with outer diameter of 3 mm and thickness of 1 mm is at 940.35 Hz for the first mode and 5,070.7 Hz for second mode. Subsequently, dynamic analysis is performed to investigate the response of the quadrotor to oscillatory excitation produced by the propeller blades. In this simulation, a 0.25 N oscillatory force is applied to the tip of each quadrotor arms. The analysis is performed over frequency range of 0–4,000 Hz. Results obtained from the Nastran are displayed in Fig. 11.2a, b. From the response spectrum, the frequency at peak for both quadrotor models matches the first natural mode

188

Displacements, Translational

a

S.K. Phang et al.

1.50–003 1.25–003 1.00–003 7.50–004 5.00–004 2.50–004 0. 0.

7.00+002

1.40+003

2.10+003

2.80+003

3.50+003

4.20+003

2.80+003

3.50+003

4.20+003

Frequency

Displacements, Translational

b

1.50–003 1.25–003 1.00–003 7.50–004 5.00–004 2.50–004 0. 0.

7.00+002

1.40+003

2.10+003 Frequency

Fig. 11.2 Displacement at tip of the arm with rectangular (a) and circular (b) hollow beam

simulated earlier. Based on the simulation result, both models are suitable for miniature quadrotor design as they have the highest stiffness to weight ratio among the tested shapes. In conclusion, using a beam with a closed cross-sectional shape as the quadrotor arm can guarantee extreme stability for the whole quadrotor model.

11.2.2 SolidWorks An overall design blueprint of this quadrotor platform is essential to provide a broad overview of the quadrotor functionalities, size, weight, and appearance design. A 3D mechanical design software named SolidWorks, developed for efficient and quicker design of mechanical products and components, facilitates the design tasks for the platform. This 3D software is chosen as the design and analyzing software over other mechanical design tools due to a few advantages, such as increase design efficiency, ease of fabricating, and easy access to simulated geometric data.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor Table 11.5 Weight budget for quadrotor MAV

Components Battery Motor and propeller Quadrotor arm Quadrotor frame Onboard system Miscellaneous Total

Amount 1 4 4 1 1

Estimate weight (g) 10 3:5 1 4 10 2 44

189

Current weight (g) 9:8 3:59 0:93 2:13 7:32 1:7 39:03

Fig. 11.3 Quadrotor arm and frame design in SolidWorks

Based on the natural mode analysis in Nastran (see previous section), a crossshaped frame structure is designed with four rectangular hollow carbon fiber beams fixed into a holder locating the avionic system along with the battery. The design takes the following considerations: 1. The CG should be located on the z-axis at the geometrical center of the cross shape. 2. Mechanical parts should be carefully designed with the weight and moment of inertia symmetrically distributed along the x- and y-axis. 3. Materials with low density and high intensity are to be used as the frame and protection to satisfy weight limit. In order to accurately approximate the weight and dimension of the overall design, components are to be drawn on scale, with the correct weight and density assigned to it in SolidWorks. As weight is the main constraint to the design of the MAV, the design parts are minimized. Table 11.5 details the total parts in the design of the quadrotor MAV, with specific weight budget assigned to each of them. On the right of the table listed the real weight of the components or designed parts, which will be detailed here and also in the next sections. The structure design can be separated into two different parts: 1. Quadrotor arms: A protection scheme is developed as a shell to encapsulate the motor into a tight chamber to fix its position. Carbon fiber beam can be mounted and screwed to the side, as shown in Fig. 11.3a.

190

S.K. Phang et al.

Fig. 11.4 Mechanical structure layout and size of the quadrotor

2. Main body: The main frame structure has four slots for the quadrotor arms and contains two layers, where the avionic system is placed on the top and the battery is located in the lower level, as shown in Fig. 11.3b. It is fabricated with acrylonitrile butadiene styrene (ABS). The mechanical structure layout of the proposed quadrotor MAV is shown in Fig. 11.4a with all the parts assembled together as shown in Fig. 11.4b. Distance of two diagonal rotors is 142.54 mm, with a total height of 22.50 mm. Details of weight breakdown are reflected in Table 11.5. In this table, the estimated weight is approximated based on the design guideline shown in this chapter, while the current weight on the right of the table is the measured weight of the quadrotor MAV prototype, code name KayLion, made by National University of Singapore, following the guideline.

11.3

Hardware Design

Hardware system and the avionics are the core of any unmanned aircraft design. The hardware system includes the motors and propellers, avionics such as onboard processor (CPU), sensors like inertial measurement unit (IMU), and power supply. Suitable choices of the hardware and components will be proposed based on the requirement on their performance under the constraint on weight and size.

11.3.1 Motor and Propeller Motor and propeller sets are the main actuators of the quadrotor MAV. As each quadrotor consists of four sets of motor and propeller, they need to be chosen carefully as their characteristics must satisfy the design requirements. A few

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

191

important design requirements of the quadrotor MAV are directly related to the characteristic of the motor and propeller. They are the operation voltage and current consumption of the motors, the weight of the motors, and the maximum thrust it can produce with different propellers. Based on the requirements stated above, a 8,000 kV single cell brushed direct current (DC) motor is utilized. Combined with two sets of propeller, each has a clockwise and an anticlockwise propellers; the total weight of a single motor and propeller is approximately 4 g. Test bench experiment has proved that the combination could produce a maximum thrust of 16 g each, which combined to be approximately 1.5 times larger than the proposed MAV at 40 g.

11.3.2 Electronics The avionic system consists of electronic components for the electrical system of the quadrotor MAV. Different from the conventional design of the avionic system for larger UAV, where each component can be chosen from the commercially offthe-shelf (COTS) product, with firm packaging and plug-and-play connection ready, the choices of electronic components for the quadrotor MAV need to be made down to integrated chip (IC) level. A printed circuit board (PCB) incorporated all onboard electronics will then be designed.

11.3.2.1 Control CPU The processor is the core of the avionic system. In the compact MAV design, the main tasks of the processor are (1) collecting data from IMU sensor from serial port, (2) receiving command from the receiver in the form of pulse position modulation (PPM) signal, (3) decoding and analyzing data, (4) realizing flight control laws, (5) sending control signals in the form of pulse width modulation (PWM) signal to the electronic speed controllers (ESC), and (6) sending logging data to the logger via serial port. For the selection of the microprocessor for the abovementioned tasks, ATmega328P, a high-performance Atmel 8-bit AVR RISC-based microcontroller, is adopted because of the following features: (1) availability of various ports, such as UART, SPI, and PWM ports; (2) low power consumption; and (3) ease of coding. It is proved that the ATmega328P microprocessor is good enough to run the software for the tasks mentioned above within 5 ms. As a result, the control loop running in the software is user configurable up to 200 Hz, depending on the update frequency of the IMU sensor. 11.3.2.2 Inertial Measurement Unit An IMU is an essential sensor to any autonomous aircraft. It provides important motion measurements of the body it attached to, such as accelerations, angular rates, and magnetic values. It is noted that for most of the aircraft orientation control, 3-axis Euler angle measurements must be provided. However, it is not necessary that all IMU provides the angular measurements as they can be estimated by using an extended Kalman filter (EKF) (Jang and Liccardo 2007) or complimentary

192

S.K. Phang et al.

filtering (Yun et al. 2007). These filters are, however, computationally intensive; thus, they might be a burden for the onboard AVR microprocessor. Here, a solution is to find a powerful IMU with in-built EKF algorithm in a small and light package. The VN-100 SMD from VectorNAV is chosen in the design. It is a lightweight (3 g) and miniature (24  22  3 mm) high-performance IMU and Attitude Heading Reference System (AHRS). In this tiny package, it includes 3-axis accelerometer, 3-axis gyroscope, and a 3-axis magnetic sensor. An extra advantage of VN-100 is the in-built 32-bit microprocessor to compute and output a real-time and drift-free 3D orientation solution with EKF algorithm. The IMU output rate is user-configurable from 40 to 200 Hz, depending on the output methods.

11.3.2.3 Motor Speed Controller ESC is essential for each brushed motor used in the MAV design. The main contribution of ESC is to convert the commonly used PWM signal to analog signal to be fed to the motor. The current of the analog signal will be further boosted to drive the motor. For the ESC design, a single 8-pin processor ATtiny13A is utilized for speed controlling. The processor is responsible for decoding PWM signals fed from the control CPU (ATmega328P) and outputting a stable analog voltage. The current output of the analog port is, however, too low to drive the motor. Thus a single MOSFET chip is also included to feed the current from the aircraft’s power supply (battery) directly to the motor. Four ESCs and MOSFETs are needed for the quadrotor MAV design. 11.3.2.4 Radio-Frequency Receiver Conventionally, a radio-frequency (RF) receiver is used to receive and decode RF signals sent from a transmitter controlled by a ground pilot in radio-controlled (RC) flights. In general, a receiver is not needed in a fully autonomous flight control system as no remote pilot is needed. Most of the UAV designs, however, retain the receiver component for fail-safe purposes, where a ground pilot has higher authority to remotely control the UAV during emergencies such as controller failures. In the quadrotor MAV design, the receiver has a different function. Besides receiving the control signals from the remote pilot, the receiver is utilized to receive control signals or measurement values from the ground control station in autonomous mode. It is particularly useful when the MAV system is navigating in indoor environment with the aid of VICON motion technology, where the VICON system measures the position and velocity of the MAV, then sends the control signals or the measurement values to the aircraft’s onboard CPU via the transmitter to receiver link. This communication system could be realized with a PCTx cable, a product by Endurance R/C. The PCTx cable connects the ground station (desktop or laptop) to the transmitter, making use of the transmitter to send RF signals to the onboard receiver. Upon receiving signals from the ground, the receiver will then output the signals in the form of PPM signal to the onboard CPU for processing. In order to fulfill the mentioned requirement, a good candidate would be Rx31d from DelTang. It is chosen to be integrated in the avionic system due to its ultra tiny package of 10  10 mm with 0.21 g. It is able to provide up to seven channels of PPM signals.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

193

11.3.2.5 Data Logger In most UAV designs, important flight data such as state variables of the UAV model will be recorded for post-flight observation and investigation. In the design of the MAV, a logger is required to be small, light, and simple enough to work. An open source data logger, called OpenLog from SparkFun, is utilized in the design. OpenLog is presented in a tiny package with 1.70 g. It will start logging any serial data up to 115,200 baud rate to the micro SD card upon powered up. Optionally, as SparkFun provides OpenLog firmware and design files, it can be redesigned to the main PCB of the MAV.

11.3.3 Power Supply The main consideration in designing the power supply is to meet the overall system and flight duration requirements. The choice of power supply is important as it constitutes most (approx. 30 %) of the overall weight of the MAV, and the power needed to lift the MAV will be increased due to its own weight. As all onboard components can be powered up with 3.3 V, a single cell Lithium-Polymer (Li-Po) battery with current capacity of 360 mAh is utilized to power the avionics and to drive the motors. A 3.3 V regulator is included to provide a clean voltage to the components, as single cell Li-Po battery’s output varied from 4.2 V during full charged down to lower than 3.4 V when it is used up. The battery is as light as 10 g and is able to provide enough energy for an 8-min flight duration.

11.3.4 Avionics Design In this subsection, the design process of the PCB for the avionic system of the quadrotor MAV will be described in detail. Among the five components to be included to the avionic system, the IMU, flight control CPU, and four ESCs will be incorporated into the design, while the receiver and the logger will be attached to the designed PCB. A general guideline to design avionics PCB for quadrotor MAV with Altium Designer is as follows: 1. Schematic design – A schematic diagram of the design must be drawn in Altium Designer with all the components needed, including four status indication LEDs. A 3.3 V voltage regulator is also included. 2. Layout assignment – The layout of the components of the PCB is important so as to reduce the electromagnetic interference between the components. To satisfy the dimension and weight constraints, a maximum of 4  4 cm PCB layout is imposed. The IMU must be placed in the middle of the design to be as closed to the CG as possible with the correct orientation. Then, components such as flight control CPU and motor speed controllers are placed at the opposite side of the board. Lastly, the four LEDs are located in a way that they are clearly visible to the user during flight. 3. Routing – The final step of designing PCB is the routing to connect each component according to the connection assigned in the schematic phase. The routing could be easily done (see Fig. 11.5) in a 2-layer-PCB setup.

194

S.K. Phang et al.

Fig. 11.5 PCB layout

Once the design is done, it can be sent to PCB manufacturer to fabricate. With PCB thickness of 1 mm, the fabricated product is approximately 7 g including all components, within the dimension of 40  40  1 mm.

11.4

Dynamic Modeling and Control

In this section, a nonlinear mathematic model of the quadrotor MAV is derived. The methods of parameters identification will also be discussed, which include direct measurement, test bench experiments, as well as evaluation in virtual environment. This mathematic model is further verified with the real flight test data obtained by the VICON motion capture system. PID control law is first implemented in the virtual model in Simulink, then further fine-tuned on the real platform based on the flight tests.

11.4.1 Nonlinear Mathematics Model The mathematical models of quadrotor aircraft are quite well developed in the literature (Bouabdallah et al. 2004; Erginer and Altug 2007; Guerrero-Castellanos et al. 2011; Kim et al. 2007; Phang et al. 2012; Pounds et al. 2006). The overall structure view of this quadrotor platform is pictured in Fig. 11.6, where ıail , ıele , ıthr , and ırud represent the normalized input signals from aileron, elevator, throttle, and rudder channels, respectively. ı1 , ı2 , ı3 , and ı4 are the normalized input values to each motors. 1 , 2 , 3 , and 4 are motors’ individual rotational speeds. Linear velocity u, v, w and angular velocity p, q, r can be obtained by 6 degree-of-freedom (DOF) rigid body dynamics, and position in ground frame x, y, z and Euler angles , , can be calculated through kinematics equation.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

1

ail

195

Ω1

2

Ω2

F

3

Ω3

M

4

Ω4

ele thr rud

Fig. 11.6 Overall structure of the quadrotor

Two coordinate frames, north-east-down (NED) frame Œxi yi zi  and the body frame Œxb yb zb , will be considered (Wang et al. 2012). The NED frame is stationary with respect to a static observer on the ground, and the body frame is the coordinate frame with the origin located at the CG and orientation moving together with the aircraft fuselage. The origin of the body frame is located at the CG of the platform with x-axis parallel with rotor 2 and rotor 3 along with y-axis parallel with rotor 1 and rotor 2. Euler angles ‚ D Œ  T are the angles rotating about x-, y-, and z-axis, describing the roll, pitch, and yaw motions. T1 , T2 , T3 , and T4 are the lift force created by rotor 1, 2, 3, and 4. Rotors 1 and 3 rotate clockwise, while rotors 2 and 4 rotate counterclockwise.

11.4.1.1 Kinematics and 6 Degree-of-Freedom Rigid Body Dynamics At any time instant, the aircraft can be pictured as having a translational and rotational motion with respect to the stationary NED (or ground) frame. The following kinematics equations represent the navigation equations applicable to transformation between the two frames: PP n D Rn=b Vb ;

(11.1)

P D S1 ; ‚

(11.2)

where Rn=b represents the transformation matrix and S1 represents a lumped transformation matrix. They are given by:

Rn=b

2 3 c c s s c  c s c s c C s s D 4c s s s s C c c c s s  s c 5 ; s s c c c

2

S1

3 1 s t c t D 40 c s 5 ; 0 s =c c =c (11.3)

with s D sin ./, c D cos ./, t D tan ./.

196

S.K. Phang et al.

Based on Newton-Euler formalism describing the translational and rotational dynamics of a rigid body, the dynamic equations can be written into the following input-output form: P b C   .mVb / D F; mV P C   .J/ D M; J

(11.4) (11.5)

where F and M are the force and moment vectors, m is the mass of aircraft, and J is the tensor of inertia matrix.

11.4.1.2 Forces and Moments Generation Based on the working principle of the quadrotor (Cai et al. 2011), the forces and torques are mainly generated by the four rotors (Goel et al. 2009). The following equation represents the component part of the overall force and moment vector: ƒD

    Fg C Fm F D ; M Mm C Mgy C Mrt

(11.6)

where subscripts g, m, gy, and rt correspond to forces and moments generated by gravity, rotors, gyroscopic effect, and reactional torque, respectively. As the platform is four-way symmetric, the CG is located on the z-axis, thereby the gravitational force only contributes to the force vector. Considering the coordinate frames, the gravity only exists on the z-axis in NED frame and needs to be transformed to the body frame by the transformation matrix: 2

3 3 2 0 mgs Fg D Rn=b 1 4 0 5 D 4mgc s 5 : mg mgc c

(11.7)

It is the motor and propeller pairs which produce the main movements to generate the forces and moments. As in Pounds et al. (2004), let Ti and Qi be the thrust and torque created by i -th rotors (i D 1; 2; 3; 4); they can be expressed as below: Ti D CT A.i R/2 ; 2

Qi D CQ A.i R/ R;

(11.8) (11.9)

where CT and CQ are the propeller aerodynamic coefficient,  is the air density, A is the area of the propeller swept by the rotating rotor, and R is the radius of the rotor A. Assuming that the distortion of the propellers during high frequency rotation can be ignored, Eqs. (11.8) and (11.9) can be simplified as: Ti D kT 2i ;

(11.10)

Qi D kQ 2i ;

(11.11)

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

197

where these two coefficients kT and kQ can be obtained by a series of experiments. Thereby, the sum of these four thrusts will result in a total thrust in negative z-axis in the body frame, as below: 2

3 0 5: Fm D 4 0 .T1 C T2 C T3 C T4 /

(11.12)

The moments are generated when the four thrusts have different magnitudes, results in pitch, roll, and yaw movements, as shown below: 2p Mm D

2 l.T2 6 p2 6 2 6 l.T1 4 2

C T3  T1  T4 /

3

7 7 ; C T 2  T 3  T 4 /7 5

(11.13)

Q1  Q2 C Q3  Q4 where l is the distance from the center of the motor to the platform CG. Gyroscopic effect is caused by the combinations of rotations of four propellers and can be modeled as

Mgy

2 31 0 D Jr @  405A .1/i i i D1 1 2 3 Jr q.1  2 C 3  4 / D 4 Jr p.1  2 C 3  4 / 5 0 4 X

0

(11.14)

(11.15)

where Jr is the moment inertia of the rotor, including the propeller and motor shaft. The gyroscopic effects produced by the propellers are only related to the moments of the aircraft. The first-order derivative of rotational movement can cause the inertia counter torque, and this torque only affects the movement in yaw direction, which is 2

3 0 5: Mrt D 4 0 P P P P Jr .1  2 C 3  4 /

(11.16)

11.4.1.3 Motor Dynamics Each brushed motor is controlled by a single cell brushed ESC, and the steady-state value of rotor angular speed to brushed ESC input can be approximated as a linear

198

S.K. Phang et al.

process near the equilibrium (hovering) value. The transient property of this brushed motor can be approximated as a first-order process, as shown below: P i D 1 ŒKm .ıi  ıi /  i ;  Tm

(11.17)

where ıi is the normalized input to the i -th motor, Tm and Km are the time constant and process gain of the first-order model, and ıi is the offset to the normalized input values. Note that the offset values of each motor are determined in the hovering condition.

11.4.2 Parameters Identification Several parameters need to be identified to obtain the complete model. Here, several methods of identifying each parameters will be presented. They are all to be done by direct measurement, experiments, or computer simulation.

11.4.2.1 Direct Measurement Parameters that can be directly measured by a weighing balance and ruler are mass of each components and the length of quadrotor arms, as follows: m D 0:032 kg;

(11.18)

L D 0:058 m:

(11.19)

Gravitational acceleration, depends on the location, can be easily calculated given the latitude of the location. In Singapore in particular, g  9:781 m=s2:

(11.20)

11.4.2.2 Computer Simulation Several parameters can be estimated numerically with computer software. As mentioned in the previous section, mechanical parts of the quadrotor MAV are first designed in SolidWorks, with exact density and scale to the real physical parts. A few parameters, such as the tensor moment of inertia of the quadrotor MAV and the rotating moment of inertia of the propeller, are calculated with the mass property function of SolidWorks: 2 3 3:0738 0 0 J D 4 0 3:0849 0 5  105 kg m2 ; (11.21) 0 0 5:9680 Jr D 5:897  108 kg m2 :

(11.22)

Note that J is diagonal as the designed quadrotor structure is highly symmetric.

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

a

x 10−4

Torque Coefficient

3

Experiment test Trend line

0.08

Experiment test Trend line Torque produced (N)

0.07 Thrust produced (N)

b

Thrust Coefficient 0.09

199

0.06 0.05 0.04 0.03 0.02

2

1

0.01 0

0 0

0.5

1

1.5

2 2

2

Rotational velocity square (rad /s )

c

x 10

0

6

0.5

1

1.5

2

2.5

Rotational velocity square (rad2/s2)

x 106

FR motor Kv model 2200 Experiment test Trend line

2100 Angular Speed (rad/s)

2.5

2000 1900 1800 1700 1600 1500 1400 1300 1200

0

0.2

0.4

0.6

0.8

1

Normalized Input

Fig. 11.7 Measurement of thrust coefficient, torque coefficient, and motor’s steady-state gain

11.4.2.3 Test Bench Experiment The thrust and torque coefficient of the rotor can be measured by lever-balance setup. An infrared transceiver is used to measure the time interval between two adjacent cutting of the propeller, where the real-time angular velocities can be calculated. Thrust and torque produced by the rotor can be measured from the weighing scale. The results are plotted in Fig.11.7a, b. Thrust and torque coefficients kT and kQ can then be obtained as the gradient of the approximate line: kT D 3:334  108 N=.rad2 =s2 /; kQ D 1:058  10

10

(11.23)

Nm=.rad =s /: 2

2

(11.24)

The next experiment is conducted to estimate the brushed motor dynamics. Several different values of constant input are fed to the motor, and the resultant angular speeds of the motors are recorded with the similar setup as the experiment above. The results are plotted in Fig. 11.7c. Here the motor steady-state velocity is

200

S.K. Phang et al.

assumed to be proportional to the input value at equilibrium (during hover). Thus, the steady-state gain, Km , can be extracted as Km D 803:9:

(11.25)

Also, as presented in Eq. 11.17, the motor dynamics can be approximated as a first-order process: Km i .s/ D ; Ri .s/ Tm s C 1

(11.26)

where average time constant Tm is estimated from the transient response of a step input to the motor. The value obtained is Tm D 0:0821 s:

(11.27)

11.4.3 Controller Design Upon obtaining the mathematical model of the aircraft, a simple yet reliable PID controller is designed and simulated with the aid of Simulink in MATLAB. Generally, the dynamics of a quadrotor without orientation stabilizer is too fast even for a skilled human pilot. The fast dynamics is contributed by the roll and pitch movements, while the yaw movement exhibits much slower dynamics. In the initial stage of the PID gains tuning, the controller is first designed in simulation by utilizing the Ziegler-Nichols method. Simulated result in Fig. 11.8 has shown a stable quadrotor system, with all the Euler angles, and angular rate signals are attenuated to zero within 5 s. This set of PID gains is further fine-tuned with real flight tests. Finally, a set of PID gains for each axis in which the aircraft is able to stabilize horizontally on this setup was obtained, as shown in Table 11.6.

11.5

Flight Test Results

The PID controller designed with software simulator shown above is implemented to the manufactured quadrotor MAV code-named KayLion (Fig. 11.9). Flight tests are then carried out to test the endurance of the vehicle and to verify the mathematical model derived in the previous section. In particular, chirp-like oscillating inputs are sent to the MAV system, while its Euler angles and angular rates responses are recorded in the onboard logger. The recorded responses are then compared to the simulated responses by using

Euler Angles (rad)

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

201

Simulation result of angles 0.5 roll φ pitch θ

0

heading ψ −0.5

0

5

10 Time (sec)

15

20

PWM input (μ s)

Angular rate (rad/s)

Simulation result of angular rates 2 roll p pitch q

0

heading r −2

0

5

10 Time (sec)

15

20

Simulation input 1501 aileron elevator

1500

rudder 1499

0

5

10 Time (sec)

15

20

Fig. 11.8 Simulation results with PID gains

Table 11.6 PID gains implemented to the quadrotor MAV

Channel Pitch Roll Yaw

Kp 1.60 1.60 4.00

Ki 2.45 2.45 0.75

Kd 0.51 0.51 0.10

the exact inputs to the simulator. Figure 11.10a shows the perturbation signal on elevator channel to the MAV, while Fig. 11.10b shows the responses of the system in pitch direction plotted together with the simulated results. Besides some random low-amplitude oscillations caused by the disturbances from the air movement, the responses match fairly well. Beside the pitch direction, the roll direction is assumed to be similar to the pitch, as the quadrotor MAV is of four-way symmetry. In the other flight test, a chirp-like signal was injected to the throttle channel of the quadrotor MAV, resulting in agitated heave movement. In this experiment, a VICON Motion Tracking System is used to measure and compute the position of the quadrotor with reference to the start-up origin. Both the input signals and position measurement are logged and plotted in Fig. 11.11a, b. In the latter figure, it can be seen that the derived mathematical model on heave movement matches well with

202

S.K. Phang et al.

Fig. 11.9 A full working prototype quadrotor MAV, KayLion, designed and built by the National University of Singapore (NUS)

the experimental data to a certain perturbation frequency, which is approximately 1 Hz. The quadrotor MAV was unable to respond to the perturbation signal above this frequency, as it was observed with naked eyes during the flight tests.

11.6

Conclusions

This chapter summarizes the steps to design and implement a micro scale quadrotor MAV, with a target take-off weight of less than 50 g. It has covered the full design areas, including the structural design, platform design, avionics design, and controller design, in which all areas are supported and verified by simulation or experimental results. Structural analysis was first done in MSC Patran and Nastran to obtain suitable candidates for the platform design, where the shape of the beam with the highest stiffness against weight ratio was selected. The platform including the arms was then designed in SolidWorks to ensure proper placement of components. Next, the avionic system design was discussed, first with the selection of suitable electronics and constraints to the trade-off between weight and performance, followed by a detailed guideline to integrate these components to a single PCB. Once the platform was fabricated, a nonlinear mathematical model of the quadrotor MAV was derived. PID controller was implemented to the quadrotor, and flight tests were done to finetune the controller. The flight test results have verified the design of the quadrotor MAV, from structural development to the avionic system implementation, and it has further proven the accuracy of the derived mathematical model. With the availability of the mathematical model for this quadrotor MAV, it serves as a good platform to test and verify control law design and implementation.

a

1550

δail

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

1500 1450

203

0

5

10

15

20

25

30

0

5

10

15

20

25

30

0

5

10

15

20

25

30

0

5

10

15 Time (sec)

20

25

30

15 time (sec)

20

25

30

15 time (sec)

20

25

30

δele

1600 1500

δthr

1400

1550 1500 1450 1400

δrud

1700 1600 1500 1400

b

0.6

Pitch angle (rad)

0.4 0.2 0 −0.2 −0.4

Simulated data Measured data

−0.6 −0.8

0

5

10

3

Pitch rate (rad/s)

2 1 0 −1 Simulated data Measured data

−2 −3

0

5

10

Fig. 11.10 Pitch angle and angular rate of the system response together with simulated response

S.K. Phang et al.

a

1600

δail

204

1500 1400

0

2

4

6

8

10

12

14

16

18

20

0

2

4

6

8

10

12

14

16

18

20

12

14

16

18

10 12 Time (sec)

14

16

18

δele

1550 1500 1450

δthr

1600 1400 2

4

6

8

10

δrud

1520 1500 1480

b

0

2

4

6

8

20

−1200 Simulated data Measured data

−1400

z−measurement (mm)

−1600 −1800 −2000 −2200 −2400 −2600 −2800

4

6

8

10

12 time (sec)

14

16

Fig. 11.11 z-axis position of the system response together with simulated response

18

20

11 Systematic Design Methodology and Construction of Micro Aerial Quadrotor

205

References M. Achtelik, A. Bachrach, R. He, S. Prentice, N. Roy, Autonomous navigation and exploration of a quadrotor helicopter in GPS-denied indoor environments, in IEEE ICRA, Kobe, 2009 M.I. Al-Qadi, A.M. Al-Bahi, Micro aerial vehicles design challenges: state of the art review, in SSAS UAV Scientific Meeting & Exhibition, Jeddah, 2006 S. Bouabdallah, A. Noth, R. Siegwart, PID vs LQR control techniques applied to an indoor micro quadrotor, in IEEE International Conference on Intelligent Robots and Systems, Sendai, 2004, pp. 2451–2456 G. Cai, B.M. Chen, T.H. Lee, Unmanned Rotorcraft Systems (Springer, London/New York, 2011) A. Eresen, N. Imamoglu, M.O. Efe, Autonomous quadrotor flight with vision-based obstacle avoidance in virtual environment. Expert Syst. Appl. 39(1), 894–905 (2012) B. Erginer, E. Altug, Modeling and PD control of a quadrotor VTOL vehicle, in IEEE Intelligent Vehicles Symposium, Istanbul, June 2007 R. Goel, S.M. Shah, N.K. Gupta, N. Ananthkrishnan, Modeling, simulation and flight testing of an autonomous quadrotor, in Proceedings of International Conference on Environmental and Agriculture Engineering, Bangalore, 2009 J.M. Grasmeyer, M.T. Keennon, Development of the black widow micro air vehicle. Prog. Astronaut. Aeronaut. 195, 519–35 (2011) J.F. Guerrero-Castellanos, N. Marchand, A. Hably, S. Lesecq, J. Delamare, Bounded attitude control of rigid bodies: real-time experimentation to a quadrotor mini-helicopter. Control Eng. Pract. 19(8), 790–797 (2011) J.S. Jang, D. Liccardo, Small UAV automation using MEMS. IEEE Aerosp. Electron. Syst. Mag. 22, 30–34 (2007) T.S. Kim, K. Stol, V. Kecman, Control of 3 DOF quadrotor model, in Robot Motion and Control 2007, ed. by K. Kozlowski (Springer, London, 2007), pp. 29–38 I. Kroo, P. Kuns, Mesoscale flight and miniature rotorcraft development, in fixed and flapping wing aerodynamics for micro air vehicle applications, of progress in astronautics and aeronautics, 195, pp. 503–517 (2001) C. Lanczos, An iteration method for the solution of the eigenvalue problem linear differential and integral operators. J. Res. Natl. Bur. Stand. 45, 255–282 (1950) L. Meier, P. Tanskanen, F. Fraundorfer, M. Pollefeys, PIXHAWK: a system for autonomous flight using onboard computer vision, in IEEE International Conference on Robotics and Automation, Shanghai, 2011 D. Mellinger, M. Shomin, V. Kumar, Control of quadrotors for robust perching and landing, in International Powered Lift Conference, Philadelphia, Oct 2010 D. Mellinger, M. Nathan, V. Kumar, Trajectory generation and control for precise aggressive maneuvers with quadrotors. Int. J. Robot. Res. 31(5), 664–674 (2012) R.C. Michelson, Overview of micro air vehicle system design and integration issues, in Encyclopedia of Aerospace Engineering, ed. by R. Blockley, W. Shyy (Wiley, Chichester/Hoboken, 2010) R. Naldi, L. Gentili, L. Marconi, A. Sala, Design and experimental validation of a nonlinear control law for a ducted-fan miniature aerial vehicle. Control engineering practice, Special issue on aerial robotics, 18(7), pp. 747–760 (2010) L. Petricca, P. Ohlckers, C. Grinde, Micro- and nano-air vehicles: state of the art. Int. J. Aerosp. Eng. 2011, 1–17 (2011) S.K. Phang, J.J. Ong, R. Yeo, B.M. Chen, T.H. Lee, Autonomous mini-UAV for indoor flight with embedded on-board vision processing as navigation system, in IEEE Region 8 International Conference on Computational Technologies in Electrical and Electronics Engineering (SIBIRCON), Irkutsk Listvyanka, 2010 S.K. Phang, C. Cai, B.M. Chen, T.H. Lee, Design and mathematical modeling of a 4-standardpropeller (4SP) quadrotor, in World Congress on Intelligent Control and Automation, Beijing, 2012

206

S.K. Phang et al.

P. Pounds, R. Mahony, J. Gresham, P. Corke, J. Roberts, Towards dynamically-favourable quadrotor aerial robots, in Proceedings of the Australian Conference on Robotics and Automation, Canberra, 2004 P. Pounds, R. Mahony, P. Corke, Modelling and control of a quad-rotor robot, in Proceedings Australasian Conference on Robotics and Automation 2006, Auckland, 2006 F. Wang, S.K. Phang, J. Cui, B.M. Chen, T.H. Lee, Search and rescue: a UAV aiding approach, in Canadian Congress of Applied Mechanics, Vancouver, June 2011 F. Wang, S.K. Phang, J. Cui, G. Cai, B.M. Chen, T.H. Lee, Nonlinear modeling of a miniature fixed-pitch coaxial UAV, in American Control Conference, Montreal, 27–29 June 2012 R.J. Wood, Design, fabrication and analysis of a 3DOF, 3 cm flapping-wing MAV, in Proceedings of the 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, San Diego, 29 Oct–2 Nov 2007 B. Yun, K. Peng, B.M. Chen, Enhancement of GPS signals for automatic control of a UAV helicopter system, in Proceedings of IEEE International Conference on Control and Automation, Guangzhou, 2007, pp. 1185–1189

Dexterous UAVs for Precision Low-Altitude Flight

12

Guangying Jiang and Richard M. Voyles

Contents 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1.1 Force Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1.2 Aerial Mobile Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1.3 3-D Translational Flight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Dexterous Hexrotor Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.1 Dexterous Hexrotor Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.2 Force Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.3 Mapping from Actuator Space to Force/Torque Space . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.4 Dexterous Hexrotor Performance Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3 Dexterous Hexrotor Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3.1 Dexterous Hexrotor Control System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3.2 Hexrotor Attitude Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4 Dexterous Hexrotor Prototype. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 Experiments and Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5.1 6-DOF Force/Torque Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5.2 Exert Forces Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

208 209 210 212 213 215 216 219 221 225 225 227 228 231 231 232 236 236

Abstract

Low-altitude flight usually introduces ground-effect disturbances and other backwash issues. In the new field of aerial mobile manipulation, it often includes close operations to structures for either inspection or manipulation of

G. Jiang () • R.M. Voyles Department of Electrical and Computer Engineering, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 130, © Springer Science+Business Media Dordrecht 2015

207

208

G. Jiang and R.M. Voyles

the structures. Although there has been a fair amount of research work of freeflying satellites with graspers, the more recent trend has been to outfit UAVs with graspers to assist various manipulation tasks. While this recent work has yielded impressive results, it is hampered by a lack of appropriate test beds for aerial mobile manipulation, similar to the state of ground-based mobile manipulation a decade ago. Typical helicopters or quadrotors cannot instantaneously resist or apply an arbitrary force in the plane perpendicular to the rotor axis, which makes them inadequate for complex mobile manipulation tasks. Based on the concept of force closure (a term from the dexterous manipulation community), this chapter introduces a new type of dexterous, 6-DOF UAV which provides the unique capability of being able to resist any applied wrench, or generalized force/torque, providing more precise control during low-altitude flight.

12.1

Introduction

A multicopter is a UAV (unmanned aerial vehicle) that is lifted and manipulated by a number of rotors. Rotors of contemporary multicopters are mostly installed on the frame such that all their thrust vectors are parallel and vertical. Control of vehicle motion is achieved by alternating the pitch or roll rotation rates. Numerous multicopter systems have been developed as custom research platforms, custom teaching platforms, toys, and commercially available systems. There even exist a number of open-source projects. Most of these multicopters have a number of fixed-pitch propellers with parallel and vertical thrust vectors. Depending on how these N thrust actuators are physically arranged, various combinations of thrust magnitudes result in net force/torque vectors that span a subset of the 6-D Cartesian space of generalized forces. Coriolis forces induced by the spinning rotors can even be used to augment the net force vector on the multirotor body. Mathematically, a matrix can be constructed that provides a mapping from N-D actuator space to 6-D Cartesian force space, but this matrix can have rank no greater than four when the thrusters are all parallel. In fact, the standard quadrotor configuration results in rank of exactly four as the four thrusters provide linear force along Z axis, and torques around X and Y, while Coriolis effects provide torque around the Z axis. (This is what makes tri-rotors infeasible, and conventional single-rotor helicopters also have four actuators: main rotor, tail rotor, and two actuators on the swash plate.) Instantaneous exertion of linear forces along X and Y is impossible with these configurations; they can only lift, pitch, roll, and yaw. They cannot move laterally without first rolling or pitching an angle, so these systems are nonholonomic. The multicopter system is dynamically unstable but allows for high maneuverability. Some attempts have been made to change this system like quadrotors, to increase controllability, stability, or maneuverability, but they still only have four motors. This instability lends itself to surveillance where quick movement is key and inability to maintain a stable pose is not an issue, but not much else can be done with these platforms.

12 Dexterous UAVs for Precision Low-Altitude Flight

209

Several hexrotor platforms have been developed including a miniature version from Airbotix and robust six- and eight-rotor versions from Aerobot and Dragonfly. The U.S. Air Force announced a solicitation for hexrotor platforms for military procurement. All these known hexrotor configurations employ parallel thrust vectors (identical to the quadrotors) which result in rank-4 actuator-to-Cartesian mappings incapable of exerting arbitrary forces in 6-D space. These six-rotor helicopters have been created simply for the increased lift of six rotors, redundancy, and smoother operation. Manipulation is making a comeback in robotics. Through most of the 1990s and early 2000s, mobile robots dominated research and application, such as the DARPA grand challenge and autonomous helicopters. Now in kind of the mid-2000s and later up to now, there has been research in manipulation again. Manipulation is about agility, dexterity, and immobilizing a part when grasping. In harkens back in manufacturing, there is an emerging field of mobile manipulation, combining the growth of mobile robots with robots people can manipulate (Holmberg and Khatib 2000). The field of mobile manipulation combines two broad classes of robots, locomotors and manipulators, yet the potential impact is greater than the sum of the parts. Aerial mobile manipulation is an emerging field within mobile manipulation for which the locomotor is a UAV (Stramigioli et al. 2012). The popular quadrotor has become the main UAV of choice in robotics research, due to its ease of control and low cost. The added mobility and access that quadrotors provide brings a new dimension to the study of mobile manipulation and new challenges as well. One of the greatest challenges that UAVs, in general, introduce, and quadrotors in particular, is the inability to instantaneously exert forces in the plane. Quadrotors are nonholonomic; in order for them to move forward or sideways, they first have to pitch the entire body of the quadrotor to direct the thrust vector in the desired direction. What this means for aerial mobile manipulation is that the quadrotor cannot resist an arbitrary generalized force/torque. In the parlance of the dexterous manipulation community, it lacks “force closure.” In fact, in one of the first attempts to use a UAV to interact physically with its environment, Albers et al. (2010) had to add an auxiliary actuator to maintain contact, so Newton’s third law of equal and opposite reaction would not immediately push the UAV away.

12.1.1 Force Closure Force closure and form closure (Mason and Salisbury 1985) are concepts from dexterous manipulation that long predate the field of robotics. Reuleaux (1963) and his contemporaries analyzed mobility under constraint in the late 1800s. Since those early days, force closure and form closure have received significant attention, yet definitions have not always been consistent within the robotics literature. Historically, force closure has been the more mature research area with a welldefined theory and set of definitions revolving around screw theory. Form closure, on the other hand, historically, has been more imprecise. Rimon and Burdick

210

G. Jiang and R.M. Voyles

Fig. 12.1 A ski resort being assembled in Canada (Reprinted with permission from Judy Shellabarger Gosnell)

published a seminal work in robotics that rigorously defined first- and second-order force and form closure and showed their equivalence. Force closure, as defined by Rimon and Burdick (1996), is the ability of a mechanism to directly resist any arbitrary wrench (combination of force and torque) applied to it. A “force closure grasp” in the dexterous grasping literature (Nguyen 1988; Bicchi 1995; Pham et al. 2006) is a grasp of an object that can resist an arbitrary wrench applied to the object. This class of grasps is important to the dexterous manipulation community but is often ignored by the mobile manipulation community because of the large mass of the mobile base and other issues have greater priority.

12.1.2 Aerial Mobile Manipulation Aerial mobile manipulation is a newly emerging field even though it has existed for decades. Figure 12.1 shows a ski resort being assembled in Canada. There is a group of workers doing the adjustment as this big helicopter is doing the heavy lifting. However, the helicopter can’t actually do the assembly. Due to Newton’s third law of equal and opposite reaction, as soon as the manipulated part comes into contact with the environment, it would balance away. This is because the helicopter

12 Dexterous UAVs for Precision Low-Altitude Flight

211

cannot exert forces instantaneously in the plane. It can only exert forces up and down, in addition to some pitch and yaw. Through most of the 1990s and early 2000s, mobile robots dominated research and application. And one of those things was quadrotors, helicopters with four blades. So people have started to apply helicopters and quadrotors to mobile manipulation. Aaron Dollar has studied unstable dynamics of the vehicle and coupled objectaircraft motion while grasping objects during flight (Dollar et al. 2011). They also demonstrate grasping and retrieval of variety of objects while hovering, using Yale Aerial Manipulator. Paul Oh equipped the quadrotor with a gripper and studied about contact forces and external disturbances acted on the gripper and the entire manipulation system (Korpela and Oh 2011). Bernard et al. (2010) and the GRASP lab at the University of Pennsylvania (Mellinger et al. 2010) both have worked on using multiple collaborative UAVs in order to perform transportation tasks. They did research on the interactions between UAVs, physical couplings in the joint transportation of the payload, and stabilizing the payload along three-dimensional trajectories. The helicopter or quadrotor approach is limited though, only the top of objects can be grasped by the underhung grasper and oddly balanced objects must be lifted by multiple UAVs. The UAV teams rigidly clamp to the manipulated object, but, due to their design, they can only apply limited forces and torques. These piloted helicopters and quadrotors manipulate objects by hanging them from a line and/or employ fewer than six vehicles, severely limiting the range of wrenches that can be applied to the object. In generalized force space, they are effectively degenerate. The floating nature of aerial platforms and the highly compliant nature of their positioning bring the issue of force closure to the fore. With fixed-base manipulators, determining the degree of force closure of a manipulation system is simplified to determining the degree of force closure of the gripper or end effector. In mobile manipulation, the manipulator base is not fixed to the ground, so determining the set of wrenches that can be resisted is not strictly limited to the capabilities of the end effector. But due to the large differences in mass of the mobile base and end effector, it is generally safe to assume the degree of force closure is limited by the end effector and not by the ability of the mobile base to remain motionless. Therefore, for aerial mobile manipulation, the concept of force closure of the entire manipulation system needs to be considered. Conventional aerial platforms are not able to resist an arbitrary wrench, so an end effector carried by such a vehicle will not be able to exhibit force closure. Force closure for arbitrary rigid objects in 3-D space requires six controllable degrees of freedom in force/torque space to truly accomplish. Current quadrotors lack both the number of degrees of freedom and independence of the degrees of freedom due to the fact that the force vectors are all parallel. Furthermore, it is not sufficient to simply attach a 6-DOF manipulator to the bottom of a quadrotor or other degenerate aerial platform, as this does not guarantee force closure. While a 6-DOF manipulator can exert arbitrary wrenches when grounded, if the base upon which it is mounted cannot resist an arbitrary wrench, the combination remains degenerate.

212

G. Jiang and R.M. Voyles

Finally, with the multi-quadrotor approach, force closure has not been obtained; lateral forces must be compensated for by torques, not direct opposition. A single UAV can have more manipulation abilities, chiefly force closure; all forces put on the grasped object can be opposed directly. Also a single UAV has numerous advantages over a system that requires multiple UAVs including simplicity, communications, cost, and available poses.

12.1.3 3-D Translational Flight Besides force closure, an interesting side effect of a UAV with control over six degrees of freedom is the ability to accomplish 3-D translational flight (hover at any orientations and translate in any directions). It would be much easier for a 6-degrees of freedom UAV to achieve unusual body attitudes as Mellinger et al. did (2010). In their work, they developed elaborate dynamic control methodologies to achieve unusual body attitudes during aggressive flight maneuvers to allow their quadrotor UAVs to pass through narrow windows and other hazards. Using high precision external sensing of the pose of the vehicle and an accurate dynamic model, their quadrotor is capable of passing through the diagonal of a rectangular orifice, presumably to enter a damaged building through a narrow window in response to a disaster. The window presents an orifice of which the horizontal width is not sufficient to allow passage of the UAV, but the diagonal distance is. Since a quadrotor cannot hover at an arbitrary orientation (the result of a mapping from actuator space to Cartesian generalized force space with rank less than six, indicative of incomplete force closure), an aggressive dynamic maneuver is required to achieve entry. Therefore, as a way to gain full controllability over the 6-DOF robot pose and track an arbitrary trajectory over time (e.g., it can hover on one spot with an angled pose), some UAV platforms with tilting propellers have been designed. M. Ryll proposed a quadrotor design with tilting propellers (Ryll et al. 2012). In their work, to solve the problem of limited mobility of standard quadrotor, rather than four fixed rotors, four variable-pitch rotors are used to provide an additional set of control inputs. Because of standard quadrotor’s inherent underactuation for 6 degrees of freedom parameterizing the quadrotor position/orientation in space, Ryll claims that they can gain full controllability over the quadrotor position/orientation by means of these four additional actuated degrees of freedom. Developed about the same time as this Dexterous Hexrotor design, a similar nonplanar design was introduced by the University of Manchester (Langkamp et al. 2011). Using six fixed-pitch/variable-speed or variable-pitch/fix-speed rotors, the “Tumbleweed” is designed to achieve full flight envelope (their saying of 3-D translational flight). A prototype is proposed with fixed-pitch/variable-speed rotors, and all rotors are pitched at 45ı . But as they claimed, it cannot achieve full flight envelope because of limited forces it can generate in frame plane. So they abandoned this design and shifted to a variable-pitch/fix-speed rotor design. And by the use of high-thrust/weight electric variable-pitch units and a low airframe mass fraction,

12 Dexterous UAVs for Precision Low-Altitude Flight

213

the Tumbleweed can achieve full flight envelope if the robot can be lifted by only one pair of rotors. In these two works, the actuation concept of tilting propellers during flight actually makes it possible to access all 6 degrees of freedom of the robot. But for aerial manipulation tasks, the aerial platform needs to exert forces as fast as possible to resist any arbitrary wrench. Tilting propellers during flight using servos may not be fast enough for our purposes. So each rotor on Dexterous Hexrotor is pitched at an optimized angle based on specific task. This chapter proposes a new approach in aerial vehicles as Dexterous Hexrotor UAV that can instantaneously resist arbitrary forces – in other words, it provides force closure. To perform precise and effective mobile manipulation, this is a property that any locomotor must have, be it ground based, water based, or air based. To achieve this, the thrusters of Dexterous Hexrotor are canted so that the combined thrust vectors span the space of Cartesian forces and torques. This adds little cost or complexity to a conventional hexrotor. With decomposed forces and torques, a thrust mapping from actuator space to force/torque for Dexterous Hexrotor is derived. Since each rotor is rotated a cant angle around its radius, a metric for the optimization of Dexterous UAV performance for manipulation-based tasks based on the condition number of the conversion matrix in the thrust mapping is developed. With an attitude controller created to stabilize the Dexterous Hexrotor for humancontrolled flight, a flight-capable prototype with translational force control has been built and tested.

12.2

Dexterous Hexrotor Theory

The basic idea behind hexrotor is 6 actuators will be provided. So the input is 6 degrees of freedom. If it needs to control 6 degrees of freedom, it has to have a minimum of 6 actuators. Or else the matrix is going to deficient. If only 4 actuators are provided as a quadrotor, it cannot possibly have independent control over 6 degrees of freedom. There are many hexrotors in the commercial world today. They all have six parallel thrust propellers spaced evenly around the circumference of a circle. All its thrusters are vertical, just like typical quadrotors as in Fig. 12.2. Therefore, they still result in rank-deficient matrices; in other words, there are no components from parallel thrusts that act in the plane perpendicular to the rotor axis. In this kind of configuration, these hexrotors work like quadrotors. Because all these thrusters are parallel and vertical, they can only provide linear force along Z axis, and torques around X, Y axes. Torque around Z is achieved indirectly through Coriolis forces resulting from differential angular velocities of the counterrotating propellers. As shown in Fig. 12.3, if all the rotors spin at a particular same speed, the robot hovers. If the speed of these rotors varies, the robot can roll, pitch, and yaw.

214

G. Jiang and R.M. Voyles

Fig. 12.2 Typical quadrotor

This configuration only has controllability over 4 degrees of freedom parameterizing the robot position and orientation in space. But it still can move arbitrarily in 6 degrees of freedom space, because translational acceleration in the plane perpendicular to the rotor axis can be achieved by maintaining a nonzero pitch or roll angle. So with these four degrees of freedom, quadrotors can do hover, roll, pitch, and yaw. If the robot wants to move in the plane perpendicular to the rotor axis, it has to maintain a nonzero pitch or roll angle to get the translational acceleration. To achieve 6 degrees of freedom, 6 actuators are provided as a typical hexrotor design. Then these parallel thrusters are made nonparallel to explore full entirely 6-D space of forces and torques. As Fig. 12.4 shows, Dexterous Hexrotor has six independently controlled rotors arranged in pairs on three planes. Each rotor is rotated a cant angle around its radius. Therefore in-plane components result while still maintaining a symmetric basis of vectors, and forces and torques around each axis can be produced. Since Dexterous Hexrotor can span the force/torque space, by detaching and combining forces and torques produced by each rotor, forces and torques acting on the UAV around each axis can be accomplished, and controllability over full 6 degrees of freedom can be achieved. By varying the speed and choosing the direction of the rotation of the rotors, Dexterous Hexrotor can get not only forces along Z axis and torques around X, Y, Z axes but also forces along X, Y axes. As shown in Fig. 12.5, with the force along Z axis and torques around X, Y, Z axes, Dexterous Hexrotor can hover, roll, pitch, and yaw just like typical quadrotor does. But instead of pitching or rolling an angle like typical quadrotors, Dexterous Hexrotor can get translational acceleration by simply varying speed of these rotors. It can truly control its 6 degrees of freedom mobility.

12 Dexterous UAVs for Precision Low-Altitude Flight

215

Fig. 12.3 How quadrotor moves: (a) hover, (b) roll, (c) pitch, and (d) yaw

12.2.1 Dexterous Hexrotor Structure There are many ways to use six motors to create six independent degrees of freedom, but to create a design that was easy to fabricate, a disk design was used. To get linear independence for all six degrees of freedom, the thrust vectors could not be on the same axis nor pointed along one axis. This fact leads to an arrangement which these six rotors are put in pairs of three planes and each thruster is rotated a cant angle  around its radius. These rotors are laid out along the edge of the disk canted tangentially to the edge of the disk alternating clockwise, counterclockwise, clockwise, and so on. Position of six motors and their rotation are defined in Fig. 12.6. X configuration of multicopter configuration is chosen, so X axis aligns with Motor3 and Motor6. To keep the direction of each rotor’s torque same with its force’s in-plane components, Motor1, Motor3, and Motor5 are rotating clockwise, while Motor2, Motor4, and Motor6 are rotating counterclockwise. And to minimize the effects

216

G. Jiang and R.M. Voyles

Fig. 12.4 Dexterous Hexrotor: note the rotors are nonparallel

of the torques from the motors themselves, pusher and puller propellers are used; Motor1, Motor3, and Motor5 use one type, while Motor2, Motor4, and Motor6 use the other.

12.2.2 Force Decomposition Each rotor produces a force and a torque. The force is a thrust pointed out along the rotor axis. The torque is generated by the drag of the propellers and acts on the body of the robot. The direction of the torque will be in the opposite direction of the motion of the propeller. This torque serves as yawing torque in typical quadrotors. But for Dexterous Hexrotor, it will contribute to torque around all three axes. The lift and drag produced by the propellers is proportional to the square of angular velocity. And the square of angular velocity is proportional to the pulse width modulation command sent to the motors. Therefore, force and torque produced by each motor can be expressed as Eq. (12.1): Fmotor D K1  PWMmotor motor D K2  PWMmotor

(12.1)

where Fmotor and motor are force and torque produced by the motors. K1 and K2 are motor-dependent parameters and can be determined experimentally. PWMmotor is the pulse width modulation command sent to the motor.

12 Dexterous UAVs for Precision Low-Altitude Flight

217

Fig. 12.5 How Dexterous Hexrotor moves: (a) hover, (b) roll, (c) pitch, (d) yaw, (e) translational acceleration along X axis, and (f) translational acceleration along Y axis

218

G. Jiang and R.M. Voyles

Fig. 12.6 Motor definition of Dexterous Hexrotor, where n is the motor Number and  represents angular displacements of the motors

To compute the net force/torque acting on the UAV from all thrusters, first each motor’s thrust and torque is decomposed into X, Y, and Z components on the body frame. The components of Cartesian generalized forces from Motor1 can be put into a matrix as in Eqs. (12.2) and (12.3): 3 2 3 F1f x K1  PWM1  cos.1 /  sin./ 6 F 7 6 K  PWM  sin. /  sin./ 7 1 1 1 6 1f y 7 6 7 7 6 6 7 K1  PWM1  cos./ 6 F1f z 7 6 7 7D6 6 7 6 F1 x 7 6 d  K1  PWM1  cos.1 /  cos./ 7 7 6 6 7 4 F1y 5 4 d  K1  PWM1  cos.1 /  cos./ 5 d  K1  PWM1  sin./ F1 z 3 2 3 2 K2  PWM1  cos.1 /sin./ Ttx1 4 Tty1 5 D 4 K2  PWM1  cos.1 /sin./ 5 K2  PWM1  cos./ Tt z1 2

(12.2)

(12.3)

where K1 and K2 are the constants between force and torque produced by Motor1 and PWM1 (the pulse width modulation command sent to Motor1), ' is the cant angle from vertical,  represents the rotor’s position, and d is the distance from rotor 1 to the central axis (the radius of Dexterous Hexrotor).  T are forces and torques decomposed from F1 F1fx F1fy F1fz F1 x F1 y F1 z T  (the force produced by Motor1) and £1£x £1£y £1£z are torques decomposed from £1 (the torque produced by Motor1).  T Then the total force/torque F1x F1y F1z £1x £1y £1z acting on Dexterous Hexrotor from Motor1 can be derived as

12 Dexterous UAVs for Precision Low-Altitude Flight

219

3 2 3 F1f x F1x 6F 7 6 7 F1f y 6 1y 7 6 7 6 7 6 7 F1f z 6 F1z 7 6 7 6 7D6 7 6 1x 7 6 F1 x C £1£x 7 6 7 6 7 4 1y 5 4 F1y C £1£y 5 1z F1 z C £1£z 2 3 K1  PWM1  cos.1 /  sin./ 6 7 K1  PWM1 C sin.1 /  sin./ 6 7 6 7 K1  PWM1  cos./ 6 7 D6 7 6 PWM1  cos.1 /  .d  K1  cos./  K2  sin.// 7 6 7 4 PWM1  cos.1 /  .d  K1  cos./ C K2  sin.// 5 PWM1  .d  K1  sin./ C K2  sin.// 3 2 K1 C 1 S 7 6 K1 C 1 S 7 6 7 6 K1 C  7 6 (12.4) D PWM1  6 7 6 C 1 .dK1 C   K2 S/ 7 7 6 4 C 1 .dK1 C  C K2 S/ 5 dK1 S C K2 C  2

Once one rotor is decomposed, forces and torques produced by other motors can be decomposed in the same way.

12.2.3 Mapping from Actuator Space to Force/Torque Space After decomposition of force/torque produced by each motor, now the net force/torque of Dexterous Hexrotor can be computed as Eq. (12.5): 2

3 2 3 2 3 2 3 2 3 2 3 2 3 Fx F1x F1x F3x F4x F5x F6x 6F 7 6F 7 6F 7 6F 7 6F 7 6F 7 6F 7 6 y 7 6 1y 7 6 1y 7 6 3y 7 6 4y 7 6 5y 7 6 6y 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 Fz 7 6 F1z 7 6 F2z 7 6 F3z 7 6 F4z 7 6 F5z 7 6 F6z 7 6 7D6 7C6 7C6 7C6 7C6 7C6 7 (12.5) 6 x 7 6 1x 7 6 2x 7 6 3x 7 6 4x 7 6 5x 7 6 6x 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 4 y 5 4 1y 5 4 2y 5 4 3y 5 4 4y 5 4 5y 5 4 6y 5 z 1z 2z 3z 4z 5z 6z T  where Fx Fy Fz £x £y £z is the net force/torque acting on the body of Dexterous Hexrotor. From force decompositions, the right part of equation equals this 66 conversion  T matrix multiplied by PWM1 PWM2 PWM3 PWM4 PWM5 PWM6 .

220

G. Jiang and R.M. Voyles

2 K1 C1 S 6 6K1 S1 S 6 6 6K1 C 6 6C1 .dK1 C  K2 S/ 6 6 4S1 .dK1 C C K2 S/ dK1 S C K2 C

K1 C2 S

K1 C3 S

K1 C4 S

K1 C5 S

K1 C6 S

K1 S2 S

K1 S3 S

K1 S4 S

K1 S5 S

K1 S6 S

K1 C

K1 C

K1 C

K1 C

3

C2 .dK1 C  K2 S/

C3 .dK1 C C K2 S/

C4 .dK1 C  K2 S/

C5 .dK1 C  K2 S/

S2 .dK1 C  K2 S/

S3 .dK1 C C K2 S/

S4 .dK1 C C K2 S/

S5 .dK1 C C K2 S/

7 7 7 7 7 7 C6 .dK1 C C K2 S/ 7 7 7 S6 .dK1 C C K2 S/5

dK1 S  K2 C

dK1 S C K2 C

dK1 S  K2 C

dK1 S C K2 C

dK1 S  K2 C

K1 C

Express this matrix as M' . This equation can be concluded as 3 Fx 6F 7 6 y7 6 7 6 Fz 7 6 7 D M' 6 x 7 6 7 4 y 5 z 2

3 PWM1 6 PWM 7 27 6 7 6 6 PWM3 7 6 7 6 PWM4 7 7 6 4 PWM5 5 PWM6 2

(12.6)

 T In Dexterous Hexrotor, Fx Fy Fz £x £y £z is the desired force/torque vector. Values of this vector will decide the robot’s orientation and position. [PWM1 PWM2 PWM3 PWM4 PWM5 PWM6 ]T are six independent controlled inputs for the robot. Based on this equation, a relationship between inputs (PWM commands) and outputs (force/torque) of Dexterous Hexrotor can be established, and a mapping from UAV’s actuator space to Cartesian force/torque space can be built. Therefore, to control Dexterous Hexrotor and get desired force/torque vector, the inversed conversion matrix is multiplied by the desired force/torque vector, then PWM commands are calculated and sent to the motors. The force/torque control equation can be derived as 2 3 2 3 PWM1 Fx 6 PWM 7 6F 7 27 6 6 y7 6 7 6 7 6 PWM3 7  1 6 Fz 7 (12.7) 6 7 6 7 D M' 6 PWM4 7 6 x 7 6 7 6 7 4 PWM5 5 4 y 5 PWM6 z The conversion matrix is the mapping from actuator space to force/torque space. With K1 and K2 determined for motors, if the cant angle is zero, the conversion matrix is M0ı : 2

M0ı

0 60 6 6 6 5:7 D6 6 1:33 6 4 0:769 0:13

0 0 5:7 1:3 0:769 0:13

0 0 5:7 0 1:54 0:13

0 0 5:7 1:33 0:769 0:13

0 0 5:7 1:33 0:769 0:13

3 0 7 0 7 7 5:7 7 7 7 0 7 1:54 5 0:13

12 Dexterous UAVs for Precision Low-Altitude Flight

221

which would conform to a typical hexrotor design. M0ı only has a rank of 4. It has no ability to instantaneously control forces in X and Y axes through this matrix. If the thrusters are canted at an angle, for example, at 20ı , the thrust mapping becomes M20ı : 2

M20ı

1:69 6 0:975 6 6 6 5:36 D6 6 1:21 6 4 0:701 0:649

1:69 0:975 5:36 1:21 0:701 0:649

0 1:95 5:36 0 1:4 0:649

1:69 0:975 5:36 1:21 0:701 0:649

3 0 1:95 7 7 7 5:36 7 7 7 0 7 1:4 5 0:649

1:69 0:975 5:36 1:21 0:701 0:649

which would conform to a nonparallel design. This matrix provides a mapping from 6-D actuator space to 6-D force/torque space and has a rank of 6, indicating 6 independent controlled degrees of freedom in Cartesian force/torque space.

12.2.4 Dexterous Hexrotor Performance Optimization T T   The cant angle decides how much force it can get in Fz £x £y or Fx Fy £z . If ' D 0ı , the matrix becomes 2

M0ı

0 60 6 6 6 5:7 D6 6 1:33 6 4 0:769 0:13

0 0 5:7 1:3 0:769 0:13

0 0 5:7 0 1:54 0:13

0 0 5:7 1:33 0:769 0:13

0 0 5:7 1:33 0:769 0:13

3 0 7 0 7 7 5:7 7 7 7 0 7 1:54 5 0:13

Clearly the matrix becomes degenerate as there are no forces in X or Y from rotor force/torque decomposition; this would conform to a fairly typical quadrotor design. There’s no mapping from actuator space to forces in X and Y. It’s rank deficient. The opposite happens when ' is 90ı : 2

M90ı

4:94 6 2:85 6 6 60 D6 6 0:11 6 4 0:065 1:54

4:94 2:85 0 0:11 0:065 1:54

0 5:7 0 0 0:13 1:54

4:94 2:85 0 0:11 0:065 1:54

4:94 2:85 0 0:11 0:065 1:54

3 0 5:7 7 7 7 0 7 7 7 0 7 0:13 5 1:54

Force can be applied in X and Y but no lift and the matrix becomes deficient again. With these examples in mind, it is obvious that the closer to ' D 0ı , the more

222

G. Jiang and R.M. Voyles

Table 12.1 Parameters of Dexterous Hexrotor engineering prototype

K1 K2 Diameter Mass of UAV Mass of manipulator Desired payload

10

5 Fx Fy Fz

9 8

4 3.5

6

3 Tz/N.m

Forces/N

Tx Ty Tz

4.5

7

5

2.5

4

2

3

1.5

2

1

1

0.5

0

5.7 1.3 0.27 m 2 kg 0.5 kg 0.3 kg

0

5

10

15

20

Cant Angle/°

25

30

35

0

0

5

10

15

20

25

30

35

Cant Angle/°

Fig. 12.7 Maximum forces Dexterous Hexrotor can achieve at different cant angles

Dexterous Hexrotor can lift, but the closer to ' D 90ı , the more lifting force can be obtained while tilted, or the more Dexterous Hexrotor can tilt. A cant angle has to be chosen somewhere between 0ı and 90ı when a prototype of Dexterous Hexrotor is being built. So a metric for optimization of Dexterous Hexrotor UAV’s performance has been developed. The performance of Dexterous Hexrotor at different cant angles is affected by several variables, including the thrust of the motors, the diameter of the UAV, and the load it needs to carry. For manipulation tasks, the engineering requirement for Dexterous Hexrotor is the desired payload. So for Dexterous Hexrotor engineering prototype, the thrust of the motors and the diameter of the UAV are given based on the motors and frame. Given the system’s own mass and desired payload, the load it needs to carry is also set. The only thing undecided is the cant angle. With parameters in Table 12.1, maximum forces and torques Dexterous Hexrotor that can achieve near hover condition at different cant angles can be plotted in Fig. 12.7. Clearly at ' D 0ı , there are no forces in X or Y and not much torques around Z. The opposite happens when ' D 35ı , where all forces in Z are used to provide

12 Dexterous UAVs for Precision Low-Altitude Flight

223

lift, but more force can be applied in X and Y. This means Dexterous Hexrotor engineering prototype can operate from 0ı to 35ı . So a cant angle between 0ı and 35ı should be chosen. To optimize the cant angle for the performance of Dexterous Hexrotor, Yoshikawa’s concept of “manipulability” was adapted here. As defined by Yoshikawa, “manipulability measure” is a quantitative measure of manipulating ability of robot arms in positioning and orienting the end effectors, by looking at the isotropism of manipulator’s motion in linear dimensions X, Y, Z and angles roll, pitch, and yaw (Yoshikawa 1985). So for Dexterous Hexrotor the combination of forces and torques is considered as a similar measure of mobility. The isotropism of the forces and torques is going to be checked, not just how strong is it. To visualize how isotropic are the forces and torques, these force/torque ellipsoids are plotted in Fig. 12.8. The radius of these ellipsoids represents the magnitude of maximum force/torque Dexterous Hexrotor can achieve around X, Y, Z axes at near hover condition. When the cant angle is 0ı , it gets no ability to control Fx, Fy, but a lot of ability to control Fz. It also has ability to control £x £y £z, but not too much control over yaw because Coriolis effect is weak. Then the motors are canted a little bit; it gets a little bit control over Fx, Fy, but still very strong control in Fz. And likewise it gets a little more control over yaw. At 20ı , the force ellipsoid becomes almost equal, and the torque ellipsoid also gets more round. At 30ı , both force ellipsoid and torque ellipsoid start to squash down again. Therefore, from 0ı to 30ı and beyond, the force/torque ellipsoid can get very isotropic at some point, which is good in this mobility measure. The condition number of the conversion matrix is also a way to check the isotropism of the forces and torques. Condition number is a mathematic term from linear algebra, which is the ratio of maximum eigenvalue to the minimum eigenvalue of the matrix. It can be seen as the rate at which one side of the equation will change with respect to the other side. If the condition number is large, even a small error in one side may cause a large error in the other side. As shown in Fig. 12.9, at 0ı , when it has no control over Fx and Fy, the condition number would be infinite, because two of the eigenvalues are 0. Then it can get smaller and smaller when increasing the cant angle. It is hard getting to 1 since force and torque are measured under different scales. Eventually, it gets higher again and it will be infinite at 90ı , because one of the eigenvalues is 0. These two metrics can be combined on a same plot as shown in Fig. 12.10. The condition number of the conversion matrix with force/torque ellipsoids at different cant angles is plotted, giving a measure of the isotropism of Dexterous Hexrotor UAV. Therefore, dependent on the motors, particular load of the manipulator, and diameter of the UAV frame, the cant angle is optimized at 20ı for this Dexterous Hexrotor engineering prototype. If the load or any other parameters are changed, a different cant angle would probably result after the optimization. And for tasks other than manipulation, its performance can be optimized with other parameters as well as the cant angle.

Fig. 12.8 Force/torque ellipsoids at different cant angles

224 G. Jiang and R.M. Voyles

12 Dexterous UAVs for Precision Low-Altitude Flight

225

10.4 10.2

Condition Number

10 9.8 9.6 9.4 9.2 9 8.8 8.6 8.4 10

15

20

25

30

35

Cant Angle/°

Fig. 12.9 Condition number of the conversion matrix

12.3

Dexterous Hexrotor Control

12.3.1 Dexterous Hexrotor Control System For human-controlled flight, the pilot controls Dexterous Hexrotor’s orientation and position by manipulating the forces and torques acting on the platform. For Dexterous Hexrotor, there are six independent controlled forces and torques acting on itself. Therefore, control systems to help the pilot control Dexterous Hexrotor’s orientation and position are developed. Forces and torques acted on the Dexterous Hexrotor are generated by the thrust of each rotor. And the thrust is proportional to the PWM value put into the motor. Based on the force/torque control equation, to generate desired force/torque, PWM T  values which are calculated from desired force/torque vector Fx Fy Fz £x £y £z through thrust mapping need to be sent to each motor. Multicopters are inherently unstable in the air. Therefore, an attitude controller is needed to stabilize the UAV during flight. A common control system used in quadrotors is as shown in Fig. 12.11. This system is constructed by radio receiver, attitude controller, IMU (inertial measurement unit), thrust mapping, and motors. With the help of attitude controller stabilizing the UAV, the pilot only needs to control its attitude to change the orientation and position (e.g., pitching for moving forward). The difference between controlling Dexterous Hexrotor and quadrotor is the  T thrust mapping. Thrust mapping of quadrotor only maps Fz £x £y £z to motors. Dexterous Hexrotor’s thrust mapping needs to map all force/torque around three axes to motors. So the control system should be as it is shown in Fig. 12.12.

8

9

10

11

12

13

14

15

16

17

5

10

15

Fig. 12.10 Condition number with force/torque ellipsoids

Condition Number

18

20 Cant Angle/°

25

30

35

40

226 G. Jiang and R.M. Voyles

12 Dexterous UAVs for Precision Low-Altitude Flight

227

Fig. 12.11 Quadrotor control system

Fig. 12.12 Dexterous Hexrotor control system

T  Thrust mapping in this system maps Fx Fy Fz £x £y £z to T ŒPWM1 PWM2 PWM3 PWM4 PWM5 PWM6  . It is a mapping from 6-D actuator space to 6-D force/torque space. Control system of Dexterous Hexrotor has two more channels controlling Fx and Fy than quadrotor. These forces are in the plane perpendicular to the rotor’s axis. They can provide horizontal acceleration. These forces can be used to exert forces immediately in the plane, and they also give the pilot another option of moving Dexterous Hexrotor. When the pilot wants to move the quadrotor, it needs to pitch or roll an angle in order to get the translational acceleration. Dexterous Hexrotor can simply generate the translational acceleration by Fx or Fy. The difference between control system of Dexterous Hexrotor and typical quadrotor is as shown in Table 12.2.

12.3.2 Hexrotor Attitude Controller The attitude controller is designed to stabilize Dexterous Hexrotor for humancontrolled flight. It helps Dexterous Hexrotor maintaining desired attitude during flight by controlling Dexterous Hexrotor roll, pitch, and yaw. This is also a commonly used control structure in quadrotor community. For stability and also fast response, three double-loop PID controllers for roll, pitch, and yaw have been developed, and they share same structure as it is shown in Fig. 12.13.

228

G. Jiang and R.M. Voyles

Table 12.2 Difference between control system of Dexterous Hexrotor and typical quadrotor Quadrotor Dexterous Hexrotor Pilot controlled variables Number of pilot controlled variables Inputs to thrust mapping Number of inputs to thrust mapping Outputs of thrust mapping

Roll, pitch, yaw, Fz

Roll, pitch, yaw, Fx, Fy, Fz

4

6

Fz £x £y £z 4

Fx Fy Fz £x £y £z 6

PWM1 PWM2 PWM3 PWM4

PWM1 PWM2 PWM3 PWM4 PWM5 PWM6

Fig. 12.13 Double-loop PID controller

The outer loop is using a PI controller for controlling the angle. It takes desired angle as input, actual IMU angle as feedback, and outputs desired angular rate to the inner loop. The inner loop is using a PD controller for controlling angular rate. It takes angular rate as input, actual IMU angular rate as feedback, and outputs torques to adjust the attitude of Dexterous Hexrotor. The PI and PD combination is chosen for a good reason. A single traditional PID controller would work, but slow. An outer loop controlling angle and an inner loop controlling rate would be much better for its fast response. P item is necessary in both loops for fast response. A D term in the inner loop is for improving response time. An I term in the outer loop is for helping Dexterous Hexrotor dealing with persistent external forces, like wind or incorrect center of gravity.

12.4

Dexterous Hexrotor Prototype

A prototype that uses off-the-shelf components, preferably those commonly used by modern research quadrotors, has been built to improve cost, compatibility, and simplicity. The prototype of Dexterous Hexrotor design can be seen in Fig. 12.14. The motors are mounted on in-house designed and fabricated ABS plastic adapters that cant 20ı tangentially to the edge. These plastic adapters are then mounted on the end of each arm as shown in Fig. 12.15.

12 Dexterous UAVs for Precision Low-Altitude Flight

229

Fig. 12.14 Dexterous Hexrotor prototype built on a commercial frame

Fig. 12.15 Motor mounted on a 20ı ABS plastic adapter

Standard 11.1 V lithium polymer battery packs are used to power the 30 A ESCs (electronic speed controllers). Each ESC drives a 1,130 kV brushless DC motor with a three-phase signal converted from the PWM command. The motors are installed with 900  4:700 pusher or puller propellers. A 2.4 GHz radio transmitter and receiver set are used for human-controlled flight. Four main joysticks are used for roll, pitch, yaw, and throttle. Two side knobs are used to set Fx and Fy values. The controller is the RecoNode platform developed in collaborative mechatronics lab. A universal PCB board is used for routing signals between onboard electronics, the RecoNode Platform, IMU, radio receiver, and ESCs as shown in Fig. 12.16. A Sparkfun 9-DOF sensor stick is used as IMU. It includes an ADXL345 accelerometer, an HMC5883L magnetometer, and an ITG-3200 gyro. It communicates with CPU by I2C interface.

230

G. Jiang and R.M. Voyles

Fig. 12.16 Onboard electronics of Dexterous Hexrotor

Fig. 12.17 RecoNode stack with Virtex4 FX20 CPU and two single-width wedges

The RecoNode is a multiprocessor architecture based on the Virtex 4 FPGA with multiple, hardcore PowerPCs. Capable of up to 1,600 MIPS from four PowerPCs plus hundreds of additional MIPS of special-purpose coprocessing from the FPGA fabric itself, this computational node is very high in performance compared to conventional wireless sensor nodes – roughly 100 times greater computational throughput. The RecoNodes are dual-processor models running at 400 MIPS with a power budget of 0.9 mW/MIPS. The RecoNode used in Dexterous Hexrotor has two base boards and two same wedges. In Fig. 12.17, from bottom to top: DU105 power board, providing 1.2, 2.5, 3.3, and 5 V for the system; DU100 CPU board, carrying a Virtex4 FX20 CPU running at 100 MHz; DU121 servo wedge, transferring the PWM commands from CPU to ESCs and input signals from radio receiver and IMU to CPU.

12 Dexterous UAVs for Precision Low-Altitude Flight

12.5

231

Experiments and Results

12.5.1 6-DOF Force/Torque Test Force closure in 3-D space requires six controllable degrees of freedom in force/torque space to truly accomplish. To prove Dexterous Hexrotor can apply force/torque in all six degrees of freedom or only one degree of freedom without affecting the others, the Dexterous Hexrotor is bolted on an ATI 6-DOF force sensor and tried to produce forces and torques in all three dimensions. The test setup and ATI force sensor software are as shown in Fig. 12.18.  During the test, unit forces and torques in the force/torque vector Fx Fy Fz £x £y £z T were commanded sequentially for around 5 s each at near hover condition. For 5 s with a five-second off after, each of the script requests a force in X, then Y, and then Z axis. Then it requests a torque about X, then Y, and then Z axis. The result of the test is as shown in Fig. 12.19 and tabulated in Table 12.1. In this plot, the first 5 s, all motors are off. It is apparent from Fz that at roughly 5 s, Dexterous Hexrotor “takes off” and then the sequence begins. Data was recorded by an ATI 6-DOF force sensor at 10 kSamp/s. Time is in milliseconds. In Table 12.3, T1 is from 5 to 10 s, which shows half throttle. T2 is from 10 to 15 s, when Dexterous Hexrotor is generating a force in x axis. T3 is from 15 to 20 s, when Dexterous Hexrotor is generating a force in y axis. T4 is from 20 to 25 s, when Dexterous Hexrotor raises its throttle. T5 is from 25 to 30 s, when Dexterous Hexrotor is generating a torque around x axis. T6 is from 30 to 35 s, when Dexterous Hexrotor is generating a torque around y axis. T7 is from 35 to 40 s, when Dexterous Hexrotor is generating a torque around z axis. This data was produced by taking the time average of the central 3 s of each period. In Table 12.4, error of each force is compared to the magnitude of required force vector. For example, during T1, required force vector in Z axis has a magnitude of 11.56 N. Fx and Fy should be zero during T1, but they are not. So values of Fx and Fy are compared to the magnitude of the force vector and get percentage error of 0:08=11:56 D 0:69 % and 0:04=11:56 D 0:34 %. It’s same in T4, T5, T6, and T7, but different in T2 and T3. During T2 and T3, required force vector points along a direction between Z axis and X or Y axis. Since errors are no more than 1 %, it proves it can actually control force in each axis accurately with no coupling to the other axes. In Table 12.5, error of each torque compared to its peak value is checked. For example, Tx during T1 should be 0, as it is given a force/torque command of [0 0 Fz 0 0 0]T , producing only a force vector in Z axis. So comparing the value of Tx during this period to its peak value happened in T5, its error is 0:01=1:13 D 0:8 %. Values of torques are much smaller than the magnitude of force vector, so their percentage error looks much larger. But still an error of torques at most 6.6 % can be observed, proving that it can actually control torque in each axis with no coupling to the other axes.

232

G. Jiang and R.M. Voyles

Fig. 12.18 Force/torque test setup: (a) Dexterous Hexrotor bolted on an ATI 6-DOF force sensor. (b) ATI 6-DOF force sensor software

With the ability of controlling force/torque in each axis with no coupling to the other axes, it shows full controllability over six degree of freedoms of Dexterous Hexrotor and linear independence between them.

12.5.2 Exert Forces Test To test the response time of the Dexterous Hexrotor corresponding to external forces, a staged peg-in-hole task is presented with Dexterous Hexrotor and a normal quadrotor. A frame diagram in Fig.12.20 shows the experiment’s setup. The peg was held rigidly by the UAV. With the peg trapped halfway in a hole which is attached to the 6-DOF ATI force sensor, the UAV starts to take off and increases its throttle until it can carry its own weight. Then it exerts a horizontal force perpendicular to the axis of the peg. The result is shown in Fig. 12.21. Changes in force sensor measurements of Fy are direct measurements of the force which Dexterous Hexrotor and quadrotor applied on the hole. A positive pulse and a negative pulse are detected around 500 and 1,500 ms. At the meantime, the attitude of Dexterous Hexrotor was recorded and plotted in the same figure. When Dexterous Hexrotor is exerting a force in Y axis, its attitude, especially pitch angle, does not change more than 1ı . This proves Dexterous Hexrotor is exerting a force in Y axis without pitching. When the quadrotor is trying to apply the same force, it pitches about 20ı and ı 20 . This is because quadrotor can only generate a horizontal force by tilting an angle and Dexterous Hexrotor can do it by simply changing the rotational velocity of the rotors without tilting.

Fx/N

Fy/N

Fz/N

2 1 0 −1

500

500

500

500

500

500

1000

1000

1000

1000

1000

1000

1500

1500

1500

1500

1500

1500

2000

2000

2000

2000

2000

2000

Fig. 12.19 Force/torque plot recorded by an ATI 6-DOF force sensor

2 1 0 −1

2 1 0 −1

20 10 0

3 2 1 0 −1

Ty/N*m

Tz/N*m

Tx/N*m

3 2 1 0 −1

2500 time

2500

2500

2500

2500

2500

3000

3000

3000

3000

3000

3000

3500

3500

3500

3500

3500

3500

4000

4000

4000

4000

4000

4000

4500

4500

4500

4500

4500

4500

12 Dexterous UAVs for Precision Low-Altitude Flight 233

234

G. Jiang and R.M. Voyles

Table 12.3 Average value of each period in the force/torque plot T1 T2 T3 T4 T5 Fx .N/ Fy .N/ Fz .N/ Tx .N  m/ Ty .N  m/ Tz .N  m/

0:08 0:04 11:56 0:01 0:04 0:04

1:59 0:02 11:96 0 0:08 0:04

0:07 1:30 11:93 0:06 0:08 0:04

0:09 0:06 19:94 0 0:08 0:03

0:06 0:05 11:67 1:13 0:09 0:03

T6 0:03 0:03 11:43 0:07 1:44 0:05

T7 0:03 0:04 11:59 0:05 0:07 0:75

Table 12.4 Percentage error of each force compared to the magnitude of force vector (%) Fx Fy Fz

T1 0.69 0.34 –

T2 – 0.17 –

T6 0.26 0.26 –

T7 0.26 0.35 –

Table 12.5 Percentage error of each torque compared to its peak value (%) T1 T2 T3 T4 T5

T6

T7

Tx Ty Tz

6.2 – 6.6

4.4 4.8 –

0.8 2.7 5.3

0 5.4 5.3

T3 0.58 – –

5.3 5.4 5.3

T4 0.45 0.30 –

0 5.4 4.0

T5 0.51 0.43 –

– 6.2 4.0

Hexrotor↵

Hexmanipulator↵ Force Sensor↵

x

Peg↵

Hole↵

z y Fig. 12.20 Peg-in-hole setup diagram, with force sensor’s coordinate system indicated

12 Dexterous UAVs for Precision Low-Altitude Flight

235

Hexrotor Fy/N

1 0

Pitch/°

−1

0

500

1000

0

500

1000

1500

2000

2500

1500

2000

2500

1500

2000

2500

1500

2000

2500

10 0 −10

Pitch/°

Pitch/°

Quadrotor 1 0 −1

0

500

1000

0

500

1000

20 0 −20 times/ms

Fig. 12.21 Dexterous Hexrotor and quadrotor measurements of Fy and pitch

100 ms

1 Dexterous Hexrotor

Fy/N

0.5 0 −0.5 −1

500

1000

1500

2000

2500

500

1000

1500

2000

2500

1 Fy/N

0.5 Quadrotor

0 −0.5 −1

0

300 ms

Fig. 12.22 Dexterous Hexrotor and quadrotor measurements of Fy

Therefore, as shown in Fig. 12.22, when it needs to exert an external force, it is obvious that the lag varying speeds of propellers of Dexterous Hexrotor will be much smaller than that caused by tilting an angle of a quadrotor. This is Dexterous Hexrotor’s advantage to typical quadrotors.

236

G. Jiang and R.M. Voyles

Table 12.6 Difference between Dexterous Hexrotor and typical quadrotor Quadrotor

Dexterous Hexrotor

Number of actuators Cant angle of each actuator DOFs in force/torque space Controllable DOFs Ability to achieve force closure Mobility in 3-D space Response of exert horizontal force

6 0ı –90ı 6 6 Yes Holonomic Fast

12.6

4 0ı 4 4 No Nonholonomic Slow

Conclusion

This chapter introduces a truly holonomic aerial rotorcraft that provides force closure for controlled interaction with external structures. The difference between Dexterous Hexrotor and quadrotor is listed in Table 12.5. The key contribution of this chapter is to derive the mapping from actuator space to force/torque space based on decomposed net force/torque and develop a metric for the optimization of Dexterous UAV’s performance for manipulation-based tasks. A flight-capable prototype with translational force control has been built and tested (Table 12.6). It should be noted that Dexterous Hexrotor can “instantaneously resist arbitrary forces” is not strictly true as the Dexterous Hexrotor can only change the torque of its motors instantaneously. The required change in the thrust magnitude is not dependent on torque, for a propeller, but on speed. Therefore, the independent thrust magnitudes and the resulting net force/torque experience a lag due to the inertia of the thrusters. The lag due to the rotor inertia is much smaller than that due to pitching the entire vehicle, as for a conventional quadrotor, and is smaller than the pitching of the variable cant angle. To achieve force closure, we cant all thrusters of Dexterous Hexrotor at an angle and make all thrust vectors nonparallel and vertical. This would definitely influence the power efficiency of the UAV. When typical quadrotor hovers, all its power is used to combat gravity. But for Dexterous Hexrotor, some of its power will be used to cancel out each rotor’s thrust. This is the main drawback of nonparallel design.

References A. Albers, S. Trautmann, T. Howard, T. Nguyen, M. Frietsch, C. Sauter, CIS&RAM 2010: semiautonomous flying robot for physical interaction with environment, in IEEE International Conference on Robotics Automation and Mechatronics, Singapore, June 2010, pp. 441–446 M. Bernard, A. Ollero, I. Maza, K. Kondak, J. Intell. Robot. Syst. 57, 417–449 (2010) A. Bicchi, Int. J. Robot. Res. 14, 319–334 (1995) A.M. Dollar, D.R. Bersak, P.E. Pounds, ICRA 2011: grasping from the air: hovering capture and load stability, in IEEE International Conference on Robotics and Automation, Shanghai, May 2011, pp. 2491–2498 R. Holmberg, O. Khatib, J. Robot. Res. 19, 1066–1074 (2000)

12 Dexterous UAVs for Precision Low-Altitude Flight

237

C. Korpela, P.Y. Oh, TePRA 2011: designing a mobile manipulator using an unmanned aerial vehicle, in IEEE International Conference on Technologies for Practical Robot Applications, Boston, Apr 2011 D. Langkamp, G. Roberts, A. Scillitoe, I. Lunnon, A. Llopis-Pascual, J. Zamecnik, S. Proctor, M. Rodriguez-Frias, M. Turner, A. Lanzon, W. Crowther, IMAV 2011: an engineering development of a novel hexrotor vehicle for 3D applications, in International Micro Air Vehicle Conference and Competitions, ’t Harde, Sept 2011 M.T. Mason, J.K. Salisbury, Robot Hands and the Mechanics of Manipulation (MIT, Cambridge, 1985) D. Mellinger, M. Shomin, N. Michael, V. Kumar, DARS 2010: cooperative grasping and transport using multiple quadrotors, in International Symposium on Distributed Autonomous Robotic Systems, Lausanne, Nov 2010. (Springer, Heidelberg, 2013), pp. 545–558 V.-D. Nguyen, Int. J. Robot. Res. 7(3), 3–16 (1988) C.B. Pham, S.H. Yeo, G. Yang, M.S. Kurbanhusen, I.-M. Chen, Mech. Mach. Theory 41, 53–69 (2006) F. Reuleaux, The Kinematics of Machinery (Dover, New York, 1963) E. Rimon, J. Burdick, ICRA ’96: on force and form closure, in IEEE International Conference on Robotics and Automation, Minneapolis, vol. 2, Apr 1996, pp. 1795–1800 M. Ryll, H.H. B¨ulthoff, P.R. Giordano, ICRA 2012: modeling and control of a quadrotor UAV with tilting propellers, in IEEE International Conference on Robotics and Automation, Saint Paul, May 2012, pp. 4606–4613 S. Stramigioli, A. Keemink, M. Fumagalli, R. Carloni, ICRA 2012: mechanical design of a manipulation system for unmanned aerial vehicles, in IEEE International Conference on Robotics and Automation, Saint Paul, May 2012, pp. 3147–3152 T. Yoshikawa, J. Robot. Res. 4(2), 3–9 (1985)

Section III UAV Fundamentals Isaac I. Kaminer

UAV Fundamentals: Introduction

13

Kimon P. Valavanis and George J. Vachtsevanos

The first two sections of the handbook have introduced the reader to UAVs in general and in particular to their design principles, paving the way for this third section, UAV Fundamentals, which details those fundamentals that constitute the foundational elements of UAV design technologies. This section covers kinematics and dynamics for fixed-wing, rotorcraft, quadrotors and flapping-wing MAVs, as well as principles of UAV guidance, navigation, and control. Kinematics and Dynamics of Fixed-Wing UAVs by Dobrokhodov provides a thorough review of the fundamentals required for accurate mathematical modeling of flight of a fixed-wing UAV, including kinematics and dynamics of motion and transformation of forces and moments acting on the airplane. The main objective is to familiarize the reader with the “kinematics–dynamics–actions” triad as it applies to a generic fixed-wing UAV. Emphasis is given to the understanding of reference frames and their dynamics as this is essential for the design of the UAV guidance, navigation, and control systems. Dynamic Model for a Miniature Aerobatic Helicopter by Gavrilets presents a nonlinear dynamic model of a miniature aerobatic helicopter following a component buildup to derive the model using simplified analytical expressions for the component forces and moments. Key parameters are estimated based on flighttest experiments. The derived model is used to design control logic for aerobatic maneuvers performed entirely under computer control.

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 134, © Springer Science+Business Media Dordrecht 2015

241

242

K.P. Valavanis and G.J. Vachtsevanos

Quadrotor Kinematics and Dynamics by Powers, Mellinger, and Kumar presents an overview of the rigid body dynamics of a quadrotor, starting from describing the Newton–Euler equations that govern the quadrotor motion. This is followed by the derivation of a linear controller based on a linearized model of the dynamics and a nonlinear controller derived from the original dynamic model. Experimental results that illustrate the dynamics and control of small quadrotors are also presented. Dynamics and Control of Flapping Wing MAVs by Doman, Oppenheimer, and Sigthorsson focuses on presenting methods suitable for hover-capable flappingwing vehicles. This is a challenging problem as flapping-wing motion is characterized by seven wing motion parameters like wing stroke amplitude, symmetric and asymmetric frequency, bias, upstroke and downstroke angle-of-attack limits, and wing stroke plane inclination. A method is presented that enables the designer to estimate how the seven wing motion parameters change in response to changes in the control mechanism position, followed by a method for estimating the combined effect of fuselage motion and flapping-wing motion. Control design considerations and simulation methods are also discussed. Principles of Guidance, Navigation, and Control of UAVs by Elkaim, Pradipta Lie, and Gebre-Egziabher presents two easily reconfigurable system architectures for guidance, navigation, and control of small UAVs. The presented system architectures integrate a low-cost inertial measurement unit, a GPS receiver, and a triad of magnetometers to generate a navigation solution which, in turn, is used in the guidance and control algorithms. The full system architecture – the hardware, software, and algorithms – is included for completeness. Hardware in the loop simulation and flight-test results documenting the performance of these two systems is given. With this background, the reader, novice or expert, is now well prepared to study “specifics” and those constituent technologies, such as sensors and sensing strategies, propulsion, control, and communications, to name a few, which will enable him/her to gain a thorough understanding of the subject or explore key areas of interest.

Kinematics and Dynamics of Fixed-Wing UAVs

14

Vladimir Dobrokhodov

Contents 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2 Reference Frames and Coordinate Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2.1 Kinematics of Moving Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2.2 Generalized Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2.3 Coordinate Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3 Rigid Body Dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.1 Conservation of Linear Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.2 Conservation of Angular Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.3 Complete Set of 6DoF Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4 Forces and Moments Acting on the Airplane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4.1 Gravitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4.2 Propulsion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4.3 Unsteady Atmosphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4.4 Aerodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.5 Accounting for the Earth Rotation Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

244 245 246 251 253 261 261 264 267 268 268 269 270 270 274 277 277

Abstract

This chapter provides a review of basic knowledge required for accurate mathematical modeling of flight of a fixed-wing UAV. These include the kinematics and dynamics of motion, and the transformation of forces and moments acting on the airplane. The detailed discussion of the “kinematics–dynamics–actions” triad in application to a generic fixed-wing UAV is the main objective of this chapter. Therefore, the presentation starts with an introduction to the coordinate frames,

V. Dobrokhodov Mechanical and Aerospace Engineering Department, Naval Postgraduate School, Monterey, CA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 53, © Springer Science+Business Media Dordrecht 2015

243

244

V. Dobrokhodov

their transformations, and differential rotations. Kinematics of the coordinate frames is what connects states of a fixed-wing UAV and transforms forces and moments acting in different coordinate frames. Understanding of reference frames and their dynamics is essential for the guidance, navigation, and control systems design. Next, the chapter provides a detailed derivation of the equations of motion using the Newtonian approach. Assuming that a fixed-wing UAV can be represented as a rigid body moving in an inertial space allows for the derivation of the linear and angular momentum equations. Starting in an inertial frame, it is shown how the final form of translational and rotational equations of motion becomes written in a body-fixed coordinate frame. The development of both the kinematic and dynamic equations is carried out first in a general vector form; then, using simplifying assumptions applicable to a generic fixedwing symmetric UAV, the vector equations are expanded into a scalar form to better represent the resulting simplification and the physical meaning of the remaining components. Finally, the chapter presents the principles of defining the external forces and moments acting on a generic fixed-wing airplane. Since the forces and moments found on an airplane act in a number of coordinate frames including inertial, body-fixed, and wind frames, the chapter utilizes the concepts and tools built in the kinematics section to transform the forces and moments into the body-fixed frame. Such transformations complete the presentation of the “kinematics–dynamics–actions” triad.

14.1

Introduction

The chapter objective is to provide an overview of the necessary theoretical material to enable a reliable mathematical modeling of the free and controlled motion of a generic fixed-wing UAV. Although the subject is not new and is well presented in existing literature, the rapid advancements of the last decade in research and development of fixed-wing UAV technologies open new applications that require understanding and a careful application of the existing assumptions. New materials, novel structural designs, new aerodynamic configurations, and advanced onboard instrumentation including miniature sensors, actuators, and tremendous onboard processing power enable much wider operational envelop of fixed-wing UAVs and significantly higher utility of their payloads. Depending on the UAV configuration and its intended operational use, the standard 12 equations of motion might not suffice for the task at hand and require deeper consideration of the UAV components interaction. This chapter starts with some preliminaries required to describe kinematics of a rigid body motion in three-dimensional (3D) space. Thus, the kinematics of 3D rotation is introduced first. The most commonly used coordinate frames that are utilized in the description of UAV states are presented next. Applying the kinematics of rotating frames to a set of specific coordinate frames builds the basis for a convenient description of the external forces and moments acting on a fixed-wing airplane. Understanding of reference frames and their rotations is essential for the eventual development of the guidance, navigation, and control systems architecture.

14 Kinematics and Dynamics of Fixed-Wing UAVs

245

Next, the chapter provides a detailed derivation of the equations of motion using the classical Newtonian approach. Assuming that a fixed-wing UAV can be represented as a rigid body moving in an inertial space allows for the derivation of the linear and angular momentum equations. Starting in an inertial frame, it is shown how the final form of translational and rotational equations of motion can be written in a body-fixed coordinate frame. The development of both the kinematic and dynamic equations is carried out first in a general vector form, and then, using simplifying assumptions applicable to a generic fixed-wing symmetric UAV, the vector equations are expanded into a scalar form to better highlight the physical meaning of their components. Since the external forces and moments found on an airplane act in a number of coordinate frames including inertial, body-fixed, and wind frames, the chapter utilizes the concepts and tools built in the kinematics description to transform the forces and moments into the body-fixed frame. Thus, the complete derivation of linear and angular momentum equations, along with the forces and moments acting on a rigid body, results in the generalized set of 6 degrees of freedom (6DoF) equations of motion.

14.2

Reference Frames and Coordinate Transformations

In order to accurately describe a body motion, it is required to define (i) the forces and moments acting on the body and thus resulting in the body motion and (ii) the coordinate system that can be used as a reference for the motion states definition. It is important to note that there are two types of forces acting on a body in free motion. First, the inertial forces and moments that depend on the velocities and accelerations relative to an inertial reference frame; the classical Newtonian dynamics equations hold only in the inertial frame. The second group consists of the aerodynamic forces and moments resulting from interaction of the body with the surrounding airflow and therefore relative to the air. Since the airflow might not be stationary, it is therefore convenient to describe the resulting aerodynamics in the coordinate frames connected to the body and to the surrounding air. The resulting motion can be conveniently described in terms of the position, velocity, acceleration, and attitude coordinates which comprise the states of the moving body. Some of these states, in turn, need to be defined with respect to a reference frame which choice is defined by the specifics of the UAV application. Thus, the information carried by various reference frames is what facilitates the complete and convenient definition of the free body motion. Therefore, this section starts with a definition of a coordinate frame and the description of the coordinate frame rotation. The reference frames required to represent the aerodynamic forces and moments and facilitating the solution of the navigation states are introduced next. Communication of the states information occurring during the coordinate frame transformation is presented for the major coordinate frames typical for UAV applications. The section ends with a set of kinematic equations required to represent the transition of linear and angular accelerations.

246 Fig. 14.1 The same plane rotation considered with respect to two and three axes. (a) Two-dimensional coordinates. (b) Three-dimensional coordinates

V. Dobrokhodov

a

y0 y1

a Rotated frame

x1

p

a Original frame

x0 Two dimensional coordinates

b

y0 y1

a Rotated frame

p

x1 a Original frame

x0

z0 Axis of rotation z1 Three dimensional coordinates

14.2.1 Kinematics of Moving Frames The objective of this subsection is to define a coordinate frame transformation and the associated mathematical formalism. Namely, the direct cosine matrix (DCM) is introduced, and its key properties are presented. The DCM matrix formalism is then followed by a differential rotation that defines the rate of change of the DCM matrix. A fundamental property of the simple summation of angular rates is introduced next. The section ends with a detailed presentation of the coordinate frames used to describe the 6DoF motion of a rigid body. The results of this development are heavily utilized throughout the entire chapter. An arbitrary motion of a rigid body can be described by a transformation that consists of (Goldstein 1980) translational and rotational components. First, a pure rotation of a rigid body is addressed. Consider a vector p defined in two orthogonal coordinate frames rotated with respect to their mutual origin by angle ˛, as shown in Fig. 14.1a. From this geometrical setup, it can be demonstrated that vector p D Œx0 ; y0  can be uniquely defined in both frames as follows: x1 D x0 cos ˛ C y0 sin ˛ ; y1 D x0 sin ˛ C y0 cos ˛

(14.1)

14 Kinematics and Dynamics of Fixed-Wing UAVs

247

where 0 and 1 subscripts refer to the coordinates of p in the original and rotated frames correspondingly. Introducing the matrix notation for the linear transformation above results in a simple form that relates the vector p components in .x0 ,y0 / frame to the corresponding components in .x1 ,y1 / frame: 

x1 y1

 D

R01



  x0 cos ˛ 1 ; R0 D  sin ˛ y0

 sin ˛ : cos ˛

(14.2)

The resulting rotation matrix is called a DCM matrix. The DCM matrix R01 consists of the cosine and sine functions which are the direction cosines between the matching axes of the new and old coordinate systems denoted in the superscript and the subscript correspondingly. Following the same approach, it can be demonstrated that for the case of right-handed coordinate system represented by three orthogonal axes (see Fig. 14.1b), the same right-hand rotation results in transformation 2

cos ˛ 1 z R0 D 4 sin ˛ 0

sin ˛ cos ˛ 0

3 0 05 ; 1

(14.3)

where for clarity, the subscript z denotes the axes of rotation. Proceeding similarly, right-handed rotations of the coordinate frame about the y0 and x0 axis give 2

cos ˛ 1 4 R D 0 y 0 sin ˛

3 2 0  sin ˛ 1 1 0 5 ; x R01 D 40 0 cos ˛ 0

0 cos ˛  sin ˛

3 0 sin ˛ 5 : cos ˛

(14.4)

It is worth noting that the DCM transformation has the following easy-to-remember properties that simplify its application (see more details in Rogers 2003): 1. The transformed vector components along the axis of rotation remain unchanged with the rotation about that axis; elements of DCM are either 0 or 1. 2. The remaining elements of DCM are either sin or cos functions of the angle of rotation. 3. The cos elements are on the main diagonal with sin elements on off-diagonal. 4. The negative sin component corresponds to the component rotated “outside” of the quadrant formed by the original frames. 5. Columns (rows) of a DCM matrix form an orthonormal set. It is straightforward to verify that a DCM matrix corresponding to the right-handed frames has the following properties:  T



T

det.R/ D 1I R D R I R R D I I R D Œc1 ; c2 ; c3  ) ci  cj D

0; i ¤ j 1; i D j (14.5)

248

V. Dobrokhodov 1 →y

Fig. 14.2 Three consecutive rotations

y0 y1 y2

q y 3 →f

j x2

q

x1

y

z0

x0

y j

z z1 z2

2 →q

and therefore it belongs to a general class of orthonormal transformation matrices. For a sequence of rotations performed with respect to each orthogonal axis, the resulting transformation can be obtained by a matrix composed of three sequential rotations, called Euler angle rotations, starting from the original frame of reference; see Fig. 14.2. Formally, this transformation is accomplished by rotating through the ordered sequence of Euler angles Π; ; , where the numerical indexes define the ordered sequence of rotations and the corresponding axis of rotations: Rxy0 D Rxy2 Rxx12 Rxx01

(14.6)

It is worth mentioning that the corresponding Euler angles are also widely used to express elementary rotation matrices so that in (14.6), the following notation y R D Rx2 ; R D Rxx12 ; R D Rxx01 is possible. As an example, R D Rxx01 defines a rotation of the axis x0 to x1 and z0 to z1 performed with respect to the axis y0 by the angle . From now, the same approach to denoting the rotations and vectors is used throughout this chapter. In the case of rotations, the subscript refers to the original frame while the superscript refers to the rotated frame; in the case of vectors denoted in bold face, the subscript defines the frame where the vector is resolved and the superscript refers to a specific meaning of the vector when necessary, and the frames are indicated by the brace as in fframeg. Therefore, a vector p0 D Œx0 ; y0 ; z0 T given in f0g coordinate frame can be resolved in another coordinate frame f1g of arbitrary orientation with respect to the original frame by a transformation matrix R01 composed of three sequential rotations as follows:

14 Kinematics and Dynamics of Fixed-Wing UAVs

32 3 2 cos  1 0 0 x1 4 y1 5 D 4 0 cos  sin  5 4 0 sin  0  sin  cos  z1 „ 2

2

cos  cos R01 D 4 cos  sin C sin  sin  cos sin  sin C cos  sin  cos

32 cos 0  sin  5 4  sin 1 0 0 0 cos  ƒ‚

249

sin cos 0

R01

cos  sin cos  cos C sin  sin  sin  sin  cos C cos  sin  sin

32 3 0 x0 5 4 0 y0 5 1 z0 … (14.7) 3  sin  sin  cos  5 cos  cos  (14.8)

This matrix, which represents a transformation resulting from three sequential Euler angle rotations, will be used throughout the chapter. Overall, any rotation matrix has a number of properties. They are summarized here for completeness; an interested reader is referred to references (Goldstein 1980; Murray et al. 1994) for thorough details: • Rotation matrices are orthogonal. • The determinant of a rotation matrix is unity. • Successive rotations can be represented by the ordered product of the individual rotation matrices. • Rotation matrices are not commutative; hence, in general case, Rbc Rab ¤ Rab Rbc . • A nontrivial rotation matrix has only one eigenvalue equal to unity with other two, being a complex conjugate pair with unity magnitude; a trivial rotation is described by an identity matrix. The time rate of change of the DCM matrix that defines the dynamics of the attitude states is important in derivation of the kinematic equations of motion. As it will be shown shortly, it enables relating the sensor (e.g., given by the rate gyros) measurements obtained in a body-fixed frame to the time derivatives of the Euler angles describing the attitude of a body in an inertial frame. In general case, the time derivative of a rotation matrix that is considered as a function of time can be obtained based on its key properties. Let R.t/ D R01 .t/ be a rotation matrix given as a function of time. Since R  RT D I , taking the time derivative of both sides yields   T  d R  RT D RP  RT C R  RP T D RP  RT C RP  RT D 0: dt First, it can be observed that RP  RT D S is a skew symmetric matrix. Next, let p1 D R.t/p0 , where p0 is a vector of constant length rotated over time with an angular velocity vector . Comparing two expressions of the absolute time derivatives of p1 P pP 1 D R.t/p 0 D S.t/R.t/  p0 D S.t/  p1 ; pP 1 D   p1 D S..t//  p1

250

V. Dobrokhodov

leads to

P T: RP D S./R; S./ D RR

(14.9)

Thus, the skew symmetric matrix S./ in (14.9) is used to represent the vector cross product between the vectors  and the R.t/p0 . The matrix S./, where vector  D Œ!x ; !y ; !z T is represented by its components, can be written in the form 3 0 !z !y S./ D 4 !z 0 !x 5 : !y !x 0 2

Another useful general property of angular velocities is called the angular velocities addition theorem (Rogers 2003). The theorem states that for angular velocity vectors coordinated in a common frame, the resulting angular velocity of the cumulative rotation is a plain sum of the contributing rotations. Now, if a rotating frame is given by a set of time varying Euler angles Œ ; ;  defined with respect to a stationary frame, then it is straightforward to determine the components of the angular velocity vector  D Œ!x ; !y ; !z T as though measured in the rotating frame. Starting from an initial stationary frame (see Fig. 14.2) and using two intermediate frames whose P and utilizing relative angular velocities are defined by the Euler angle rates Œ P ; P ;  the angular velocities addition theorem, the following kinematic equation can be obtained: 2 3 2 3 2 3 2 3 !x 'P 0 0 4 !y 5 D R' R 4 0 5 C R' 4 P 5 C 4 0 5 (14.10) P 0 : 0 !z Substituting the corresponding DCM matrices from (14.7) results in 3 2 1 !x 4 !y 5 D 4 0 0 !z 2

0 cos '  sin '

32 3 'P  sin  5 4 cos  sin ' P 5 P cos  cos ' :

(14.11)

3 sin  3 2 !x cos ' cos   sin ' 5 4 !y 5 cos ' cos1  !z

(14.12)

Inverting the last equation results in equation 3 2 1 'P 4 P 5 D 4 0 P 0 2

sin  sin ' cos  cos ' sin ' cos1 

that defines the derivatives of the Euler angles in terms of the angles themselves and the angular velocity vector  D Œ!x ; !y ; !z T as it was measured in the rotated frame. These equations define the rotational kinematics of a rigid body; they contribute to the final set of 6DoF equations of motion. Analysis of the (14.12) shows that four elements of the inverted matrix become singular when the second rotation angle  approaches =2. This problem is usually called a kinematic singularity or a gimbal lock in navigation and is one of the

14 Kinematics and Dynamics of Fixed-Wing UAVs

251

issues associated with the use of Euler angles for the attitude determination. For differently ordered Euler rotation sequences, the kinematic singularity will occur at a different point. Therefore, one way to avoid the singularity is to switch or change the Euler angle sequences when approaching the singularity. Next, depending on the available computing power, the integration of the kinematic equation (14.12) can be computationally expensive because it involves calculation of trigonometric functions. Furthermore, it can be observed that the Euler angle-based DCM matrix is redundant; it requires only 3 out of 9 elements of the DCM matrix to uniquely define the Euler angles. These shortcomings usually result in applying different parameters describing the attitude and its dynamic transformation. In applications to the fixed-wing UAV attitude determination, the Rodrigues– Hamilton parameter, or quaternion, is one of the most widely used alternatives (Goldstein 1980). Utilizing the quaternion approach is very powerful because it gives a singularity-free attitude determination at any orientation of a rigid body. Next, since it can be shown that the equations of motion of a rigid body are linear differential equations in the components of quaternion, then it (linearity) is a desirable property, especially when developing estimation and control algorithms. Furthermore, the quaternion is a relatively computationally efficient approach since it does not involve trigonometric functions to compute the attitude matrix and has only one redundant parameter, as opposed to the six redundant elements of the attitude matrix. However, it is also worth noting that the quaternion and Euler angle techniques are both widely used in various UAV applications; the equations connecting both representations are well developed, thus enabling complementary definition of the DCM matrix and Euler angles through the parameters of quaternion and vice versa. An interested reader is referred to an extensive historical survey of attitude representations (Shuster 1993) and references (Rogers 2003; Goldstein 1980; Murray et al. 1994) for more details in the alternative methods of attitude determination.

14.2.2 Generalized Motion In the development of dynamic equations of motion, it will be necessary to calculate the absolute time derivative of a vector defined in coordinate frames that rotate and move with respect to each other. In an application to the UAV kinematics, this can be justified by a necessity to relate the absolute time derivative of a position vector in inertial space (inertial velocity) that is defined based on the measurements taken in a body frame. Similarly, the second time derivative defines the body inertial acceleration. Consider two coordinate frames fFi g and fFr g, where i stands for an inertial not rotating frame and r stands for the rotating frame. The first objective is to calculate the derivative of a unity vector rr defined in fFr g attached to a rigid body rotating with respect to the fFi g with angular speed ¨; see Fig. 14.3. Denote the rotation from fFr g to fFi g as R: ri D Rrr

252

V. Dobrokhodov

Fig. 14.3 Deriving the time derivative of a vector

zi

Rigid body

zr

r

yr

w xi

O

xr yi

particle

Rotating frame

Non-rotating frame

Taking the derivative results in P r C RPrr D Rr P r D S.¨/rr D ¨  rr ; rP i D Rr

(14.13)

where the time derivative rP r is zero due to the rigid body assumption. Next, using the same setup, calculate the absolute time derivative of an arbitrary time varying vector r defined in fFr g. Defining the vector in terms of its components in both frames and taking its time derivative in the inertial frame result in r D rxi i C ryi j C rzi k D rxr l C ryr m C rzr n: Taking the absolute time derivative of both expressions gives dr D rPxi i C rPyi j C rPzi k dt : dr D rPxr l C rPyr m C rPzr n C rxr lP C ryr m P C rzr nP dt Applying (14.13) allows rewriting the last equation as ır dr D C ¨  r D rP r C ¨  r; dt ıt

(14.14)

which expresses the derivative of the vector r in the inertial frame fFi g in terms of its change (Prr ) calculated in a rotating frame fFr g and its relative rotation defined by the angular velocity ¨. The result in (14.14) is known as the Coriolis theorem. The second derivative of r defines the body acceleration and is used in the development of the dynamic equations. It is obtained in a similar manner by recursively applying the Coriolis theorem, thus leading to

14 Kinematics and Dynamics of Fixed-Wing UAVs

ı .Prr C ¨  r/ C ¨  .Prr C ¨  r/ ıt ır ı¨ D rR r C rC¨ C ¨  rP r C ¨  .¨  r/; ıt ıt rR D rR r C ¨ P  r C 2¨  rP r C ¨  .¨  r/:

253

rR D

(14.15)

Similarly to (14.14) where rP r denotes the local derivative taken in a rotating frame fFr g, the ı¨ refers to the derivative of angular velocity ¨ taken in fFr g frame. ıt However, it can be easily demonstrated ı¨ ı¨ d¨ D C¨¨ D dt ıt ıt

(14.16)

that the derivative of ¨ is independent on the coordinate frame. In turn, this justifies omitting the subscript in ¨ referring to fFr g. The last two terms in ( 14.15) are commonly known as Coriolis and the centripetal accelerations correspondingly. The following chapter heavily relies on the results in (14.14) and (14.15) when it develops the dynamic equations of motion.

14.2.3 Coordinate Frames Deriving equations of motion of a fixed-wing UAV requires a definition of coordinate frames where forces and moments acting on the airplane can be conveniently defined and where the states including the position, velocity, acceleration, and attitude can be suitably described. It is worth noting that with the latest advances in power technologies and novel materials, a mission of long duration becomes a reality. As an example, solar power technology is one of the alternatives that can make the 24/7 flight of a fixed-wing solar powered autonomous soaring gliders feasible. Thus, the long duration and great operational distances might require considering the UAV flight operations with respect to the rotating Earth. This consideration would extend the set of coordinate frames used in long-endurance UAV applications. The primary reason for such an extension would lie in the necessity to resolve the inertial angular velocity of rotation of a body frame in a “true” inertial frame where the Earth rotation can be resolved. Those frames would include the Earth-centered inertial, Earth-centered Earth-fixed, and a geodetic coordinate frames. Consequently, the angular velocities addition theorem mention above would be used to resolve the angular velocity vector of body rotation with respect to the inertial frame as a vector sum of angular velocities of the intermediate frames. An interested reader is referred to (Rogers 2003) for more details in the coordinate frames and transformations used in inertial navigation. Therefore, in addition to the body and wind frames that define dynamics of the body–fluid interaction, this subsection defines the following set of coordinate frames used in various UAV applications: 1. Earth-centered inertial frame fi g 2. Earth-centered Earth-fixed frame feg

254

V. Dobrokhodov

Geodetic coordinate system f, ', hg Tangent plane coordinate system fug Body-carried frame fng Body-fixed frame fbg Wind frame fwg Depending on the duration of flight and operational range, both dictated by a specific UAV application, the first four frames can be considered inertial frames, with the remaining three frames body fixed. The positioning of inertial and body frames is related by a plain translation, while the orientation of body frames relates to each other by pure rotations. Details of the frames definition and their relations are the subject of this section. 3. 4. 5. 6. 7.

14.2.3.1 Earth-Centered and Geodetic Coordinate Frames It is convenient to consider two coordinate frames connected to the Earth. The Earth-centered Earth-fixed (ECEF) coordinate system is fixed to the Earth, and therefore, it rotates at the Earth’s sidereal rate with respect to the Earth-centered inertial (ECI) frame that represents nonrotating inertial frame. The ECI frame is usually denoted fi g while the ECEF frame is denoted feg. Both frames are righthanded orthogonal and have their origins at the center of the Earth. The ECI frame has its zi axis aligned with the direction of the Earth’s rotation vector and xi and yi axes placed in the equatorial plane with xi fixed in some celestial reference direction, for example, a line connecting the Sun’s center and the Earth’s position at vernal equinox (Kaplan 1981). The ECEF has xe and ye axes placed in the equatorial plane and ze axis aligned with zi axis; see Fig. 14.4 where the Earth is modeled as a spheroid. The xe axis is usually attached to the intersection of the Greenwich meridian and the equator, and the ye axis completes the right-hand system. It is worth noting that the ECEF axes definition may vary; however, the definition always states the attachment of two vectors to the direction of the Earth rotation and the Greenwich meridian as the inherent Earth properties. The sidereal rate ei is the vector defining the ECEF rotation with respect to the ECI; the latter one is often called the true inertial frame. If necessary, for the purpose of UAV flight description, the magnitude ei of the rate can be approximated by one full rotation in 23h5604.09900, thus resulting in ei D 15:04106718 deg/h. Therefore, the transformation from ECI to ECEF frame is a plain rotation around the zi axis defined by a single rotation by an angle ei  t, where t is the time interval. The local geodetic f; '; h g frame is usually associated with the ECEF frame, see Fig. 14.4. It has the same origin placed at the center of the Earth. The frame defines the orientation of a line normal to Earth’s surface and passing through the point of interest. The orientation of the line is defined by two angles,  (geodetic latitude) and ' (geodetic longitude), with the height h above the Earth’s surface; eventually these three parameters, along with the components of the UAV velocity vector, become the major navigation states. For most UAV applications, it is sufficiently accurate to model Earth’s surface as an oblate spheroid with given re (equatorial) and rp (polar radiuses) or one of the radiuses and the e (ellipticity) (Kaplan 1981). Last revisited in 2004, the datum of World Geodetic System (WGS-84) provides

14 Kinematics and Dynamics of Fixed-Wing UAVs zi, ze

255

Ωi

e

(xe, ye, ze)

h

Local tangent plane

j ye l

xi Celestial reference

Equatorial plane

yi xe Greenwich meridian

Fig. 14.4 ECI, ECEF, and geodetic coordinate frames

the following parameters for the oblate spheroid modeling: re D 6,378,137:00 m, rp D 6,356,752:314 m. The resulting transformation from the geodetic f; '; hg to the ECEF frame is as follows: xe D .r C h/ cos ' cos  ye D .r C h/ cos ' sin  ze D ..1  "2 /r C h/ sin '

(14.17)

where " the eccentricity of oblate ellipsoid is defined as re I r D q 1  "2 sin2 ' s rp2 " D 1 2 re :

14.2.3.2 Local Tangent Plane Coordinate System The origin of the local tangent plane (LTP) is fixed to the surface of the Earth with two of its axes attached to the plane tangent to the surface; see Fig. 14.5. The frame is usually marked with the subscript fug and serves the purpose of an inertial frame in most short-duration low-speed UAV applications. The frame’s xu and yu axes are in the tangent plane and most often aligned with the north and east directions correspondingly; the zu axis completes the right-hand coordinate system, thus pointing down. Quite often the order and alignment of the LTP frame principal

256

V. Dobrokhodov

Fig. 14.5 Definition of the local tangent plane; NED

xu yu

North

East

Local tangent plane at

l, j, h zu Down, toward the Earth center

axes change. In such cases, the LTP coordinate system explicitly specifies its type; in the nominal case presented above, it can be also defined as an NED frame indicating the north–east down alignment of the coordinate axes. When the origin of the LTP frame is defined in terms of its geodetic latitude, longitude, and altitude above the ground surface, then the equations (14.17) can be applied to define the kinematics of navigation states.

14.2.3.3 Body-Carried and Fixed Frames In flight dynamics, the body-attached reference frames usually have their origin at the center of gravity (CG) of an airplane; therefore, these frames are moving. The body-carried frame fng is an orthogonal frame originated at the CG of the UAV. All its axes are permanently stabilized and aligned with the LTP frame axes as it was connected to the CG; see Fig. 14.6; the frame is usually utilized in defining the navigation equations thus assigning its subscript fng. This frame is connected to the LTP frame by means of a plain translation r D Œrn ; re ; rd T . The body-fixed frame is an orthogonal frame defined with respect to the bodycarried frame. Its origin is at the CG of UAV, and its axes are rigidly connected to the body; therefore, the frame also rotates with the body. The frame is usually marked with the subscript fbg. The alignment of the fbg frame axes is a degree of freedom that can exploit the body symmetry. It can be proven (Goldstein 1980) that for every rigid body, there is always an orthogonal coordinate system, usually called principal, in which the cross products of inertia terms are zero. Assuming that a typical UAV has at least one plane of symmetry (geometric and mass symmetry) results in two of the body-fixed axes laying in the plane of symmetry. When the axes are aligned along the principal axes of inertia of the body, as will be shown in the following chapter, the dynamic equations of motion become significantly simpler. In a majority of fixed-wing UAV configurations, the axes of fbg frame match the principal axes of inertia. The typical orientation of the body-fixed axes is as follows (see Fig. 14.6): if the UAV has a vertical plane of symmetry, then xb and zb lie in that plane of symmetry; xb points toward the direction of flight and zb points downward and yb points right, thus completing the right-hand system.

14 Kinematics and Dynamics of Fixed-Wing UAVs Fig. 14.6 Definition of the body-fixed frame with respect to LTP frame

257

yb

xn

yn CG

xb

zb

r

zn North

xu yu

East

NED {u}

zu

Down, towards the Earth center

As the body moves, its attitude is defined with reference to the body-carried frame fng by three consecutive Euler rotations by (yaw),  (pitch), and  (roll) angles. See their graphical illustration in Fig. 14.2 where frames f0g and f1g relate to the frames fng and fbg correspondingly. The formal definition of the Euler angles in the application to an airplane attitude specification is presented here for completeness: • – yaw is the angle between xn and the projection of xb on the local horizontal plane. •  – pitch is the angle between the local horizon and the xb axis measured in the vertical plane of symmetry of the UAV. •  – roll is the angle between the body-fixed yb axis and the local horizon measured in the plane yb zb . As it follows from the attitude representation section, the DCM matrix Rub transforming the body-carried fng to the body-fixed fbg frame can be constructed as in (14.8). Now the subscripts (u ! b) denote the rotation from the LTP fug to the bodyfixed frame; fug and fng frames are always aligned by the definition of body-carried frame. The application of the rotation matrix ( 14.8) immediately follows from the need to describe the UAV translational motion in an inertial frame of reference by utilizing the inertial velocity measurements taken in the body-fixed frame at

258

V. Dobrokhodov g

CG – Vb D Œu; ; wT . To this end, consider Fig. 14.6 where vector r D Œrn ; re ; rd T denotes the position of an airplane CG with respect to the LTP (NED) frame attached to the Earth. Relating the translational velocity and position and accounting for the fact that body-carried frame fng is stabilized with respect to the nonrotating fug frame results in dr g D Rbu Vb dt 2 3 2 3 (14.18) u : r d 4 n5 u4 5 re D Rb  dt w rd The relation between the Euler angles defining the relation between the stabilized fng frame and the body-fixed frame fbg was already derived in (14.11) and (14.12). They define the dynamics of Euler angles defined in an inertial frame with respect to the rates measured in the body-fixed frame. Thus, the kinematic equations (14.12) and (14.18) represent the dynamics of translational and rotational coordinates and therefore are part of the final set of equations of motion.

14.2.3.4 Wind Frame Aerodynamic forces and moments resulting from the body–air interaction as the airframe moves through the air depend on the body orientation with respect to the surrounding air. In other words, they depend on the vector representing the wind. The velocity vector of the possibly moving air (wind) resolved in the inertial frame fug is denoted Vau ; see Fig. 14.7. The magnitude of Vau is called an airspeed Va , as opposed to the velocity vector defined in LTP with respect to the ground – ground g speed vector Vu . The orientation of the wind frame fwg defined by the direction of Vau with respect to the body–fixed fbg is defined by two angles. To generate the lift force in flight, the wing of the UAV must be oriented at a usually positive angle ˛ with respect to the Vau vector. This angle is called the angle of attack. The angle of attack ˛ is also one of the key parameters that define

CG

yb Fig. 14.7 Wind frame and body-fixed frames. Definition of the angle of attack and the sideslip

a

Pitot tube

b

zb

xw

Vua

xb

14 Kinematics and Dynamics of Fixed-Wing UAVs

259

the longitudinal stability of an airplane. Therefore, quite often, the coordinate frame that results from a single rotation from the body-fixed fbg frame on angle ˛ is called a stability frame (Beard and McLain 2012; Etkin and Reid 1995). As illustrated in Fig.14.7, the angle of attack is defined by the projection of Vau into a vertical plane of symmetry of the UAV (spanned by axes xb ; zb in frame fbg) and the longitudinal axis xb of the UAV. It is positive when a leading edge of the wing rotates upward with respect to the Vau . In turn, the angle between the velocity vector Vau projected into the “wing-level” plane (spanned by axes xb ; yb in frame fbg) and the longitudinal axis xb of UAV is called the sideslip angle. It is denoted by ˇ. Then the transformation from the body-fixed frame fbg to the wind frame fwg is given by 3 32 2 cos ˛ 0 sin ˛ cos ˇ sin ˇ 0 Rbw D 4 sin ˇ cos ˇ 05 4 0 1 0 5  sin ˛ 0 cos ˛ 0 0 1 3 2 cos ˛ cos ˇ sin ˇ sin ˛ cos ˇ (14.19) D 4 cos ˛ sin ˇ cos ˇ  sin ˛ sin ˇ 5 :  sin ˛ 0 cos ˛ The inverse transformation from the wind frame fwg to the body-fixed frame fbg is  T the transpose of (14.19): Rwb D Rbw . The importance of the wind frame in application to the UAVs flying in wind conditions that might contribute up to 100 % of the nominal airplane speed cannot be overestimated. As an example, imagine an autonomous glider that is designed to utilize the wind energy to sustain the long duration of flight. Therefore, it is necessary to understand the difference between the airspeed, represented by the air g velocity vector Vau and the ground speed Vu , both resolved with respect to the LTP frame. Consider the graphical representation of the relation between these vectors in Fig. 14.8. In the presence of constant wind, these velocities are related by the equation that is often called the wind triangle: Vau D Vgu  Vwu

(14.20)

where Vwu is the wind velocity defined in the LTP frame. The objective of the following development is to define the relations among these velocities defined in three different frames, while being measured or estimated by the sensors installed in the body-fixed and in the inertial frame. First, define the components of all three vectors in the body-fixed frame fbg. Let the UAV velocity g in the LTP (inertial) frame expressed in the body frame be Vb D Œu; ; wT , and let the wind velocity in the LTP frame expressed in body frame be Vwb D Œuw ; w ; ww T . Observe that Vau can be expressed in fwg frame as Vaw D ŒVa 0 0T , and let Vab D Œua a wa T be its components expressed in the body frame. Utilizing the definition of the angles of attack and sideslip relating the wind frame to the body-fixed frame and the “wind triangle” (14.20) expressed in the body frame results in the following:

260

V. Dobrokhodov

Fig. 14.8 Wind triangle in 2D plane and the yaw ( ), sideslip (ˇ), and course ( ) angles

xb

North

xu

χc

χ ψ

Vua Vuw Vug

β

CG

Ground track

2 3 2 3 uw u g Vab D Vb  Vwb D 4  5  4 w 5 w ww 2

3 2 3 2 32 3 ua Va Va cos ˛ cos ˇ sin ˇ sin ˛ cos ˇ Vab D 4 a 5 D Rwb 4 0 5 D 4  cos ˛ sin ˇ cos ˇ  sin ˛ sin ˇ 5 4 0 5 :  sin ˛ 0 cos ˛ 0 0 wa 3 3 2 cos ˛ cos ˇ ua 4 a 5 D Va 4 sin ˇ 5 sin ˛ cos ˇ wa 2

(14.21)

This last equation relates the airspeed components of Vab resolved in the body frame with the airspeed and the angles of attack and sideslip. In turn, if the wind components resolved in the body frame are known, then inverting the last equation allows for calculation of the airspeed Va and the ˛; ˇ angles: Va D

q

 u2a C a2 C w2a ; ˛ D a tan

wa ua



 ; ˇ D a sin

 a : Va

(14.22)

14.2.3.5 Summary of Kinematics This section developed the fundamental kinematic equations that not only define the kinematics of states and contribute to the final set of 6DoF equations of motion but also serve as the basis for the design of the guidance and navigation tasks. There are numerous publications describing the kinematics of moving frames. Most of the publications originate in the area of classical mechanics and rigid body

14 Kinematics and Dynamics of Fixed-Wing UAVs

261

dynamics (Murray et al. 1994; Goldstein 1980). The publications in the area of flight dynamics and control always contain material addressing the attitude representation techniques and differential rotations and thus can be a good source of reference information. The most recent and thorough presentation of these topics can be found in (Beard and McLain 2012) where the authors specifically address the kinematics and dynamics of small UAVs.

14.3

Rigid Body Dynamics

This section addresses the development of the dynamics of a rigid body. The discussion is based on the application of the Newton’s laws in the cases of linear and angular motion. In particular, the Newton’s second law of motion states that the sum of all external forces acting on a body in an inertial frame must be equal to the time rate of change of its linear momentum. On the other hand, the sum of the external moments acting on a body must be equal to the time rate of change of its angular momentum. The application of these laws is the objective of this chapter. Thus, a fixed-wing UAV is considered as the rigid body, and its dynamics is defined with respect to the body-fixed coordinate system. The approach considers a typical fixed-wing UAV operating in a small region of the Earth thus justifying the assumption of LTP frame as an inertial frame. Relations necessary to translate the inertial forces to the body-fixed frame are also presented. Before proceeding to the derivation, it is necessary to present some assumptions typical for the fixed-wing UAVs: A1. The mass of the UAV remains constant during flight. A2. The UAV is a rigid body. The relations derived in this chapter are general and can be applied to any rigid body; however, the treatment of the aerodynamic forces and moments acting on the body will be specific to the aerodynamically controlled fixed-wing UAVs.

14.3.1 Conservation of Linear Momentum First, assume that a rigid body consists of a set of i “isolated” elementary particles with mass mi exposed to the external force Fiu while being connected together by the internal forces Riu . Since the set of N particles comprises a rigid body structure N P Riu D 0. The set of (see A2), the net force exerted by all the particles is i

external forces acting on the body is a combination of the gravity force acting in an inertial frame fug and the aerodynamic and propulsion forces defined with respect to the body-fixed frame fbg but expressed in the inertial frame fug. Thus, the linear momentum of a single particle expressed in an inertial frame obeys the equality Fiu C Riu D

d .mi Viu /; dt

(14.23)

262

V. Dobrokhodov

where the time derivative is taken in an inertial frame. Summing up all the N particles comprising the body gives the linear momentum equation of the entire body N X

Fiu D

i D1

N X d .mi Viu /: dt i D1

(14.24)

The left-hand side of this equation represents the sum of all forces (gravitational, propulsion, and aerodynamic) expressed in an inertial frame, with the right side depending on the velocity of the body defined in an inertial frame. First, observe that the individual inertial velocities Viu are not independent. Since they comprise g a rigid body, their vectors can be represented as a sum of Vb the inertial velocity vector resolved in fbg frame with respect to CG plus the velocity vector induced by the body rotation with respect to the CG and defined by the radius vector rib of the g i th particle in fbg frame; thus Viu D Vb C ¨  rib . Next, using an assumption A1 and utilizing the result in (14.14) for the total velocity of the i th particle in an inertial frame, allows for the calculation of the absolute time derivatives in an inertial frame in the following form: N X i D1

Fiu

N N X X  d d   g i .mi Vu / D mi Vb C ¨  rib D dt dt i D1 i D1 N X

g

N

X d Vb d C mi .¨  rib / dt dt i D1 i D1 # " N N g X X d Vb d i C ¨ D mi m i rb : dt dt i D1 i D1

D

mi

(14.25)

Here ¨ represents the angular velocity of the UAV body resolved in the inertial cg cg frame; see (14.14). Defining rb the vector of CG location in fbg frame as mrb D N N P P mi rib , where m D mi is the total mass of the body, simplifies the linear

i D1

momentum equation:

i D1

N X i D1

Fiu D m

g d Vb d cg

Cm ¨  rb : dt dt

Resolving all external forces in fbg frame and assuming that the location of CG does not change with time and applying the result in (14.14) to the absolute derivatives g of the vectors Vb and ¨ result in cg g cg

Pg C¨ P  rb C ¨  Vb C ¨  ¨  rb /; Fb D m.V b where the vector quantities can be represented in the following scalar form:

(14.26)

14 Kinematics and Dynamics of Fixed-Wing UAVs

263

Fb D ŒX; Y; ZT – the externally applied forces expressed in the body frame g Vb D Œu; v; wT – the inertial velocity components defined in the body frame ¨ D Œp; q; rT – the inertial angular velocity resolved in the body frame cg rb D Œxcg ; ycg ; zcg T – the body referenced location of the center of gravity. The translation of the inertial forces to the body frame is justified by the g convenience of calculating the local body frame derivatives of the Vb and ¨ P g , while the derivative of expressed in the body frame; the first one results in V b ¨ is independent on the coordinate frame; see (14.16).  cg

cg  Utilizing the double vector product identity ¨  ¨  rb D ¨  rb  ¨  cg .¨  ¨/  rb allows for the expansion of the linear momentum equation in the scalar form as follows:

X D m uP C qw  r C qz P cg  ry P cg C .qycg C rzcg /p  .q 2 C r 2 /xcg

P cg C .rzcg C pxcg /q  .r 2 C p 2 /ycg Y D m P C ru  pw C rP xcg  pz

P cg C .pxcg C qycg /r  .p 2 C q 2 /zcg : Z D m wP C p  qu C py P cg  qx (14.27) The last set of equations allows for an arbitrary choice of the body frame origin fbg with respect to the CG. However, if the origin of the body-fixed frame fbg is chosen at the CG, the last set of equations can be significantly simplified by substituting cg rb D Œ0; 0; 0T , thus leading to   g P C ¨  Vg : Fb D m V b b

(14.28)

Expanding the cross product results in the following simplified form of the linear momentum equation: X D m ŒPu C qw  r Y D m ŒP C ru  pw

(14.29)

Z D m ŒwP C p  qu: Resolving equations (14.29) with respect to the derivatives (accelerations in the body frame) leads to the standard form of differential equations suitable for immediate mathematical modeling: uP D X m C Œr  qw P D

Y m

C Œpw  ru

wP D

Z m

C Œqu  p :

(14.30)

264

V. Dobrokhodov

14.3.2 Conservation of Angular Momentum Applying the law of conservation of angular momentum to an i th particle in a moving frame is very similar to the approach used above. Consider an i th particle subjected to the internal (Miu ) and external (riu  Fiu ) moments acting on the body in inertial frame. Similar to the linear momentum case, the sum of internal  N moments  P i acting on the particle of a rigid body (A2) should be equal to zero Mu D 0 , i D1

while the external moments arise from the inertial gravity and the body-attached forces, such as aerodynamic and propulsion. Thus, the conservation of angular momentum calculated across the entire rigid body results in N X

.Miu C rib Fiu / D

i D1

N X

Miu C

i D1

N X

rib  Fiu D

i D1

N X

rib 

i D1

d .mi Viu /; dt

(14.31)

where the sum of internal moments cancel each other out. Then, applying the Coriolis theorem (14.14) leads to N X i D1

rib 

N

X d d .mi Viu / D mi rib  .Viu / dt dt i D1 D

N X

mi rib

 

i D1

D

N X

mi rib 

i D1

C

N X



g

ıVb ı¨ g C ¨  Vb C  rib C ¨  ¨  rib ıt ıt



 X   N g ıVb ı¨ g C ¨  Vb C  rib mi rib  ıt ıt i D1



mi rib  ¨  Œ¨  rib  :

(14.32)

i D1

The first term can be expanded by utilizing the definition of the CG: N X



g

ıVb g C ¨  Vb ıt





g

ıVb g C ¨  Vb ıt i D1 2

3 m ycg .wP C p  qu/  zcg .P C ru  pw/ 7 6 6

7 7 6 P C p  qu/ 7 6 m zcg .Pu C qw  r/  xcg .w D6 7: 7 6

7 6 4 m xcg .P C ru  pw/  ycg .Pu C qw  r/ 5 mi rib 



D mrcg 

(14.33)

14 Kinematics and Dynamics of Fixed-Wing UAVs

265

Utilizing the double vector product identity allows for the expansion of the second term as follows: N X

mi rib

 

i D1

2

ı¨  rib ıt

 D

N X

 mi

i D1

ı¨  i i  r  r  rib ıt b b



ı¨ i  rb ıt



3 2 2 2 3 m ..y C z / p P  .y q P C z r P /x / i i i i 7 i i 6 Ixx pP C Ixy qP C Ixz rP 6 i D1 7 6 6 N 7 7 6 6P 7 7 2 2 6 6 7 7 m ..z C x / q P  .z r P C x p/y P / i i i 7 D 6 I pP C I q i D 6 i D1 i i P C I r P yy yz 7 6 yx 6 7 7 6P 7 4 5 6 N 7 2 2 mi ..xi C yi /Pr  .xi pP C yi q/z P i/5 4 Izx pP C Izy qP C Izz rP N P

i D1

2

32 3 pP 76 7 76 7 76 7 P Iyz 7 6 qP 7 D I  ¨: 76 7 54 5 Izz rP

Ixx Ixy Ixz

6 6 6 D 6 Iyx Iyy 6 4 Izx Izy

(14.34)

The equation (14.34) is obtained by recognizing the moments of inertia and their symmetrical properties:

Ixx D

N X

N N X X  2  2    2 2 Iyy D Izz D mi yi C zi mi zi C xi mi xi2 C yi2

i D1

Ixy D Iyx D 

i D1 N X i D1

mi xi yi

Ixz D Izx D 

i D1 N X i D1

mi xi zi

Iyz D Izy D 

N X

mi yi zi

i D1

Combining them into a matrix form defines the inertia tensor I that allows the conversion of the entire double vector product into a very compact form as in ( 14.34). The diagonal terms of I are called the moments of inertia. The offdiagonal terms are called the products of inertia; they define the inertia cross coupling. The moments of inertia are directly proportional to the UAV’s tendency to resist angular acceleration with respect to a specific axis of rotation. For a body with axes of symmetry, the inertia tensor can be resolved (Goldstein 1980) with zero offdiagonal terms that significantly simplify its form and the final equations of angular momentum.

266

V. Dobrokhodov

The last term in (14.32) utilizes twice the same double cross product expansion, thus leading to N P i D1

N

P    mi rib  ¨  ¨  rib D mi rib  ¨  rib ¨  .¨  ¨/ rib

2

i D1

2

2

Iyz .q  r / C Ixz pq  Ixy pr

3

2

.Izz  Iyy /rq

3

6 7 6 7 6 7 6 7 6 6 7 7 D 6 Ixz .r 2  p 2 / C Ixy rq  Iyz pq 7 C 6 .Ixx  Izz /rp 7 : 6 7 6 7 4 5 4 5 Ixy .p 2  q 2 / C Iyz pr  Ixz qr .Iyy  Ixx /qp

(14.35)

Resolving the total inertial moment acting on the UAV in body frame and denoting the components as Mb D ŒL; M; N T and combining the results in (14.33)–(14.35) lead to the following complete angular momentum equations in an expanded form: L D Ixx pP C Ixy qP C Ixz rP CIyz .q 2  r 2 / C Ixz pq  Ixy pr C .Izz  Iyy /rq

P C p  qu/  zcg .P C ru  pw/ Cm ycg .w M D Iyx pP C Iyy qP C Iyz rP CIxz .r 2  p 2 / C Ixy rq  Iyz pq C .Ixx  Izz /rp

Cm zcg .Pu C qw  r/  xcg .wP C p  qu/

(14.36)

N D Izx pP C Izy qP C Izz rP CIxy .p 2  q 2 / C Iyz pr  Ixz qr C .Iyy  Ixx /qp

Cm xcg .P C ru  pw/  ycg .Pu C qw  r/ : cg

If the origin of the body-fixed frame fbg is chosen at the CG (rb D Œ0; 0; 0T ), then in the case of a typical UAV with a vertical plane of symmetry spanned by the body-fixed axes xb ; zb , the two pairs of the off-diagonal terms of I matrix become zero, namely, Ixy D Iyx D 0 and Iyz D Izy D 0. This significantly simplifies the above equations: L D Ixx pP C .Izz  Iyy /rq C Ixz .Pr C pq/ M D Iyy qP C .Ixx  Izz /rp C Ixz .r 2  p 2 /

(14.37)

N D Izz rP C .Iyy  Ixx /qp C Izx .pP  qr/:

These equations represent the complete rotational dynamics of a typical fixed-wing UAV modeled as a rigid body with a longitudinal plane of symmetry.

14 Kinematics and Dynamics of Fixed-Wing UAVs

267

14.3.3 Complete Set of 6DoF Equations of Motion The final set of 6DoF equations of motion describing the kinematics and dynamics of a generic UAV with a longitudinal plane of symmetry modeled as a rigid body can be summarized as follows: X D m ŒPu C qw  r Y D m ŒP C ru  pw

(14.38)

Z D m ŒwP C p  qu L D Ixx pP C .Izz  Iyy /rq C Ixz .Pr C pq/ M D Iyy qP C .Ixx  Izz /rp C Ixz .r 2  p 2 /

(14.39)

N D Izz rP C .Iyy  Ixx /qp C Izx .pP  qr/ g

rP D Rbu Vb D 2 cos  cos 4cos  sin  sin 

 cos  sin C sin  sin  cos cos  cos C sin  sin  sin sin  cos 

3 sin  sin C cos  sin  cos g  sin  cos C cos  sin  sin 5Vb cos  cos  (14.40)

2

3

2

'P 1 6 7 6 6 7 6 6 P7 6 6  7 D 60 6 7 6 4 5 4 P 0

sin  sin ' cos 

cos ' sin ' cos1 

32 3 sin  cos ' cos p  76 7 76 7 76 7  sin ' 7 6 q 7 : 76 7 54 5 1 cos ' cos  r

(14.41)

Analysis of the above differential equations shows that these equations are nonlinear and coupled, that is, each differential equation depends upon variables which are described by other differential equations. In general case, their analytical solutions are not known, and they can only be solved numerically. There are 12 states describing the free motion of a rigid body subject to external forces (Fb D ŒX; Y; ZT ) and moments (Mb D ŒL; M; N T ). In the control system design, these variables are called state variables because they completely define the state of a rigid body at any instance of time. The state variables are summarized in Table 14.1 for completeness. What remains in the description of 6DoF equations of motion is to define the external forces and moments acting on the airplane. This will be the objective of the next section.

268

V. Dobrokhodov

Table 14.1 State variables of the 6DoF equations of motion State variable Definition r D Œrx ; ry ; rz T

Vector of inertial position of the UAV and its components Vector of inertial velocity components resolved in the body-fixed frame Euler angles that define the attitude of body-fixed frame with respect to the inertial frame The inertial angular rates resolved in the body-fixed frame

g

Vb D Œu; ; wT Œ; ;  ¨ D Œp; q; rT

14.4

Forces and Moments Acting on the Airplane

The objective of this section is to present a generalized approach to defining the external forces and moments acting on a fixed-wing UAV as functions of its states. The primary forces and moments acting on an airplane are the gravitational, thrust of the propulsion system, aerodynamic, and disturbances due to the flight in unsteady atmosphere. The most challenging task here is in defining the aerodynamic forces and moments resulting from the air–body interaction. Although the aerodynamic description of airfoils defining a fixed wing is not a new subject, the varieties of shapes, aspect ratios, and aerodynamic configurations of modern fixed-wing UAVs do not allow thorough presentation of all configurations. As an example, possible aerodynamic configurations of aerodynamic surfaces include tandem, variable span wings, joined wings, twin boom, and V-tail configuration, just to name a few. However, a generalization is possible. An interested reader is referred to the most relevant survey (Mueller and DeLaurier 2003) of aerodynamics of small UAV that describes the modeling approaches and their limitations.

14.4.1 Gravitation Assuming that the flight altitude is negligible in comparison to the radius of the Earth, it is sufficient to consider the gravity’s magnitude constant. Then, the effect of the Earth’s gravitation can be naturally modeled in the body-carried frame by the force applied to the CG of the UAV of mass m; the gravitational force is proportional to the gravitational constant g and is called the weight of the UAV 3 0 D 4 0 5: mg 2

Fgr u

(14.42)

Before substituting this force expression into the equations of motion (14.39), it needs to be resolved in the body frame. The inertial to body rotation Rub enables this transformation:

14 Kinematics and Dynamics of Fixed-Wing UAVs

3 2 3  sin  0 D Rub 4 0 5 D mg 4 sin  cos  5 : cos  cos  mg

269

2

gr Fb

(14.43)

Assuming that the frame fbg origin is chosen at the CG and since the gravitational force acts through the CG of the airplane, the corresponding moment contribution gr is zero, Mb D Œ0; 0; 0T .

14.4.2 Propulsion The configuration of the propulsion system of modern fixed-wing UAVs varies greatly. The architectures can be categorized by the number of engines, their type, and their installation arrangement in the airframe. A thorough review of the existing configurations along with some future projections and trends in the modern and future UAV systems can be found in reference (OSD 2001). However, what is common across all possible configurations is that the vector of thrust in all systems is set parallel to the existing axes of symmetry; the thrust vectoring is not a common feature of fixed-wing UAVs yet. The thrust is naturally represented in the body-fixed reference system. The direction of the thrust vector Ftbr is usually fixed and lies in the plane of symmetry or is parallel to it; however, it may not be aligned with the longitudinal xb -axis. If the orientation of the thrust vector Ftbr varies in its reference to the airframe, then a separate coordinate system analogous to the wind axes should be defined, thus introducing the required rotation of the thrust vector to the body-fixed coordinate system. It is a common design requirement that the installation of multiple engines should not introduce any unbalanced moments, thus not inducing any loss of control efforts for the UAV stabilization. For the analysis of a nominal flight regime, the thrust vector Ftbr is considered fixed with respect to the body-fixed frame. For the sake of simplicity, consider a typical fixed-wing UAV architecture where the installation of one or multiple engines results in the cumulative thrust vector Ftbr passing through the CG and the only moment being the torque generated primarily by the reactive force from the rotating propeller; depending on the type and the power of the propulsion system, there might be three more components (Illman 1999) of the torque, namely, the spiraling slipstream, the gyroscopic precession, and the asymmetric propeller loading (“P-factor”). Thus, in the case of a typical UAV, the net force Xt r of thrust in xb direction and the moment L around xb axis can be considered proportional to the thrust control command ıt r . Moreover, thrust characteristics of most conventional engines are always functions of the air density and the airspeed. Thus, the contributing force and moment resulting from the propulsion system can be presented as follows: 2

2 3 3 Ftr .Va ; h; ıtr / Mtr .Va ; h; ıtr / 5 I Mtrb D 4 5: Ftrb D 4 0 0 0 0

(14.44)

270

V. Dobrokhodov

A particular example of modeling the propulsion force for the case of a micro UAV can be found in Beard and McLain (2012).

14.4.3 Unsteady Atmosphere In the previous discussion of the wind frame, it was assumed that the wind Vwu defined in the LTP frame is constant; thus, the velocities are related by the “wind g triangle” equation Vau D Vu  Vwu . The most common approach (McRuer et al. 1999) in wind modeling is to wst eady consider two components contributing to the wind. The first component Vu defines the steady wind resolved in the inertial frame, and therefore it can be wgust presented by the measurements in the LTP frame. The second component Vb is stochastic, which represents the short-period disturbances or gusts resolved in the body-fixed frame. Since the equations of motion are written in the body-fixed frame, then wgust eady Vwb D Rub Vwst C Vb : (14.45) u From the components of the wind and the UAV velocity, both resolved in the body frame, it is therefore possible to find the body frame components of the air velocity as 3 2 3 2 wst eady 3 2 wgust 3 u u u ua Vab D 4 a 5 D 4  5  Rub 4  wst eady 5  4  wgust 5 : w wa wwst eady wwgust 2

(14.46)

These body frame components of the air velocity enable straightforward calculation of the airspeed and the angles of attack and sideslip as in (14.22). Modeling of the stochastic and steady components of wind is based primarily on a history of experimental observations expressed using linear filters. The most widely used techniques are represented by von Karman and Dryden wind turbulence models (Hoblit 2001). Both methods are well supported with their numerical implementations.

14.4.4 Aerodynamics Aerodynamic forces and moments depend on the interaction of an aircraft with the airflow, which may also be in motion relative to the Earth. However, for the purpose of representing the nominal aerodynamic effects, the large-scale motion of the atmosphere is not critical and therefore will be considered constant; in fact, it will only affect the navigation of the UAV. The small perturbation theory (Ashley and Landahl 1985) is one of the approaches used in describing the aerodynamic interaction of a given aerodynamic shape with airflow. The perturbation in aerodynamic forces and moments are

14 Kinematics and Dynamics of Fixed-Wing UAVs

271

functions of variations in state variables and control inputs. The control inputs here are the deflections of the control surfaces of an airplane that modify the airflow around the body, thus generating the desired aerodynamic effects. The nomenclature of the control surfaces and their control mechanization depends on the particular aerodynamic composition of the airplane. Nevertheless, the principles describing the effects of the control surface deflection on the generated forces and moments are the same. Consider the following control effectors of a classical aerodynamic configuration: the elevator, the aileron, and the rudder (see Fig. 14.9). In this configuration, the ailerons are used to control the roll angle , the elevator is used to control the pitch angle , and the rudder controls the yaw angle . Their deflections are denoted as ıa for the aileron, ıe for the elevator, and ır for the rudder. The positive deflection of a control surface is defined by applying the right-hand rule to the rotation axis of the surface. The positive direction of the aileron, elevator, and rudder deflections is also depicted in Fig. 14.9. Deflection of the control surfaces modifies the pressure distribution around the control surfaces or the entire body thus producing corresponding forces. The forces acting with respect to the CG of the body result in aerodynamic moments. For example, deflecting the elevator primarily changes the pitching moment acting on the airplane. In turn, this results in changing the angle of attack of the wing that increases the lifting power of the airplane. The calculation of aerodynamic characteristics of one or more lifting surfaces with variable deflections of the control surfaces at various attitudes with respect to the airflow can be accomplished by utilizing well-developed linear panel methods (Hess 1990; Henne 1990) conveniently implemented in various software packages (Fearn 2008; Kroo 2012). The panel methods capture the effect of pressure distribution in the form of parameterized forces and moments versus the angles of attack and sideslip, and airspeed; they play a role of states here. For example, considering the longitudinal plane, the effect of aerodynamic pressure acting on a fixed wing can be modeled † using a total force F† b and pitching moment Mb acting on the wing both resolved in the body frame. It is common to project the total force to the wind axes, thus resulting in the lift Flif t and drag Fdrag force components. Figure 14.10

Left aileron

+da Rudder

+dr

Elevator

Fig. 14.9 Control surfaces of a classical aerodynamic configuration

+de

Right aileron

+da

272 Fig. 14.10 Definition of the lift and drag forces and the pitching moment in the wind frame

V. Dobrokhodov FbΣ

Flift M bΣ

xb α

V∞ xw

Fdrag

Va

c 4 zb

b 2

zw

c

demonstrates the approach to modeling aerodynamic effects in the wind and bodyfixed frames with respect to the vector of free airstream V1 . As shown in Fig. 14.10, the lift Flif t and drag Fdrag forces act in the wind frame and are applied at the aerodynamic center of the lifting surface that is located at the quarter-chord point (c is the length of the mean aerodynamic chord). The pitching component M of moment M† b acts around the aerodynamic center. Then, the values of forces and moments are represented in a form connecting a number of surfacespecific parameters and the states in the following form: Flif t D 1=2 Va2 S CL Fdrag D 1=2 Va2 S CD ;

(14.47)

M D 1=2 Va2 ScCm where CL ; CD ; Cm are the nondimensional aerodynamic coefficients (to be parameterized), S is the planform area of the wing surface, and c and b are the mean aerodynamic chord and the wing span. The same approach is applied to each of the aerodynamic surfaces comprising the airplane. Then by using the parallel axis theorem (Goldstein 1980), the elementary moments from all the lifting and control surfaces can be transferred to the CG of the rigid body, thus resolving the actions with respect to one unifying center. It is common practice to consider the total aerodynamic forces and moments in projections to the longitudinal and lateral planes of the airplane. The benefit of this approach is in the simplicity of representing the aerodynamic effects and in providing a natural ground for the nonlinear model decomposition at the next step of the control system design. Thus, the longitudinal forces and moments (14.47) consist of lift, drag, and pitching moment acting in the vertical plane of symmetry. The lateral side force Fside , yawing N , and rolling L moments are caused by the asymmetric airflow around the airplane and control surfaces deflection; the asymmetry can be caused by the side wind or intentional deflection of the rudder. For the majority of fixed-wing UAVs, the key states that define the parameterization

14 Kinematics and Dynamics of Fixed-Wing UAVs Table 14.2 Parameterization of longitudinal and lateral aerodynamics

273

Longitudinal channel

Lateral channel

Fdrag D 1=2 Va2 SCD .˛; q; ıe /

Fside D 1=2 Va2 SCY .ˇ; p; r; ır ; ıa /

Flift D 1=2 Va2 SCL .˛; q; ıe /

L D 1=2 Va2 SbCl .ˇ; p; r; ır ; ıa /

M D 1=2 Va2 Sc Cm .˛; q; ıe /

N D 1=2 Va2 SbCn .ˇ; p; r; ır ; ıa /

of the aerodynamic coefficient are the angle of attack ˛, the sideslip ˇ, body rates Œp; q; r, and the controls which are the surface deflections Œıe ; ır ; ıa . The most general functional form of the longitudinal and lateral aerodynamics can be presented as follows (Table 14.2): Without delving deeper into the intricacies of aerodynamic parameterization, it is sufficient to demonstrate the final form of forces and moments defined in the wind coordinate frame: • Longitudinal plane   q c Fdrag D 1=2 Va2 S CD0 C CD˛ ˛ C CD  q C CDıe  ıe 2Va   q c ıe 2 ˛ 1 Flift D =2 Va S CL0 C CL ˛ C CL  q C CL  ıe (14.48) 2Va   q c ıe ˛ M D 1=2 Va2 Sc CM0 C CM ˛ C CM  q C CM  ıe 2Va • Lateral plane   ˇ p b ır ıa 2 r b 1 Fside D =2 Va S CY0 C CY ˇ C CY  p C CY  r C CY  ır C CY  ıa 2Va 2Va   b ˇ p b L D 1=2 Va2 S b Cl0 C Cl ˇ C Cl  p C Clr  r C Clır  ır C Clıa  ıa 2Va 2Va   b b ˇ p N D 1=2 Va2 S b Cn0 C Cn ˇ C Cn  p C Cnr  r C Cnır  ır C Cnıa  ıa 2Va 2Va (14.49) The presented parameterization is a simple linear approximation of the aerodynamics given by the Taylor series expansion taken with respect to the given trim at e conditions. The coefficients Cfst=m are the nondimensional partial derivatives of the corresponding forces and moments (denoted in the subscript) defined with respect to the corresponding state or control (denoted in the superscript). The coefficients with zero in the subscript denote the forces and moments calculated when all states, including the control surface deflection, are zero; for example, Cl0 denotes the roll moment coefficient estimated at ˇ D p D r D 0 and ır D ıa D 0. The common naming convention suggests that those derivatives defined with respect to states Œ˛; ˇ; p; q; r are called the stability derivatives and those with respect to controls Œıe ; ır ; ıa  are called the control derivatives. The static stability of an aircraft

274

V. Dobrokhodov

with respect to disturbances in some variable is directly reflected in the sign of a ˛ particular derivative. For example, the sign of CM should be negative to guarantee ˇ static stability in pitching motion, while the sign of CN should be positive for the directional static stability. Each of the presented coefficients is usually a function of states. The precision requirement of the linear parameterization greatly depends on the operational envelop of the UAV and its intended use; the higher the maneuverability of the UAV, the more terms necessary to accurately represent the aerodynamics. Each of the coefficients has very intuitive physical meaning and is usually studied separately. An interested reader is referred to Beard and McLain (2012) for a detailed discussion of the aerodynamic coefficients of small and micro fixed-wing UAVs. One last step needs to be performed before the aerodynamics (14.48) and (14.49) defined in the wind coordinates can be plugged into the equations of motion (14.38) and (14.39) resolved in the body-fixed frame. The transformation (14.19) from the wind to the body frame serves this purpose. Therefore, the total forces and moments acting on the fixed-wing UAV can be presented as follows: 3 3 2 2 2 3 3 X Ftr .Va ; h; ıtr / 0 CD .˛; q; ıe / 4 Y 5 D Rub 4 0 5 C 4 5 C 1= Va2 S  Rwb 4 CY .ˇ; p; r; ır ; ıa / 5 0 2 Z mg CL .˛; q; ıe / 0 2

(14.50) 2

3 2 3 2 3 L Mtr .Va ; h; ıtr / bCl .ˇ; p; r; ır ; ıa / 4M 5 D 4 5 C 1= Va2 S  4 cCM .˛; q; ıe / 5 : 0 2 N bCn .ˇ; p; r; ır ; ıa / 0

14.5

(14.51)

Accounting for the Earth Rotation Rate

The complete set of 6DoF equations of motion presented above is an approximation of the rigid body kinematics and dynamics and is valid as long as the assumption of the flat Earth model satisfies the task at hand. During the high-speed flight or in longduration and extended range missions, the precision of the derived states will suffer from omitting the sidereal rate of the rotating Earth. The key reason for the error is in the accumulation over time of the Coriolis and centripetal accelerations induced by the rotating Earth. Thus, the following derivation outlines how the Earth rotation can be accounted for in the definition of the inertial velocity and acceleration vectors. First, consider the ECI as the true inertial frame fi g. Next, by using the simplifying properties of defining the free motion of a rigid body with respect to the CG and utilizing the Coriolis theorem, resolve the absolute time derivative of cg the CG position vector ri in the true inertial frame as follows: cg

cg

cg

e Vi D rP i D Vcg e C  i  ri :

(14.52)

14 Kinematics and Dynamics of Fixed-Wing UAVs

275

Taking the second time derivative and assuming that that the sidereal rate of the P ei D 0) results in Earth rotation is constant ( cg cg cg cg e e e P cg P cg P cg C ¨  Vcg V  rP i D V e C i  .Ve C i  ri / i D Vb C ¨ b  i cg   Ve e C : cg cg e e P C ¨ C i  Ve C i  .i  ri / DV b (14.53)

In (14.53), the ¨ denotes the vector of inertial angular velocity resolved in the body cg cg g frame, and Vb is the same as Vb D Vb . The equation (14.52) updates the kinematic dead reckoning equation in ( 14.39), while the vector of inertial acceleration in (14.53) should be used in the application of the second Newtonian law. Applying the angular velocities addition theorem, the vector ¨ can be represented as a sum of the angular velocity vector ¨bn of body frame fbg resolved in the body-carried frame fng, the angular velocity vector ¨ne of body-carried frame fng resolved in ECEF frame feg, and the sidereal rate of the Earth rotation vector ei resolved in the true inertial frame fi g. Thus, the last equation can be also written as   cg e e P cg D V P cg C ¨bn C ¨ne C 2ei  Vcg V e C i  .i  ri /: i b

(14.54)

What remains is to define the elements of ( 14.54) that enable calculation of the vector cross products. cg cg As before, the term 2ei  Ve is the Coriolis, and the term ei  .ei  ri /, is n the centripetal accelerations. The angular velocity vector ¨e , can be obtained from P rates, which in turn can be calculated the geodetic latitude (') P and longitude () cg from the NED components of Ve D ŒVN ; VE ; VD T . The transformation of rates of P to the body-carried stabilized frame fng can be obtained P ) the geodetic system (, similarly to (14.10) by a left-handed rotation around the east axis through the latitude angle ' 2 3 2 32 P 3 0 cos  0 sin   ¨ne D 4 P 5 C 4 0 1 0 5 4 0 5 : (14.55) 0  sin  0 cos  0 The rate of change of latitude and longitude can be calculated from the Vn northern and Ve eastern components of the velocity vector as follows: P D

Vn ; rm C h

P D

VE ; .rn C h/ cos '

where h is the height of CG above the reference oblate spheroid and rm D 

re .1  "2 / 1  "2 sin2 '

 32 ;

re rn D q 1  "2 sin2 '

(14.56)

276

V. Dobrokhodov

are the estimates of the reference spheroid radius in the meridian and normal directions at given latitude and longitude. Substituting (14.56) into the (14.55) results in the estimate of ¨ne as follows: ¨ne



VN VE VE ;  ;  tan  D rn C h rm C h rn C h

T

:

(14.57)

Observe that the Earth sidereal rotation vector ei has only one component in ECEF

T frame ei D 0; 0; ei . Resolving for convenience ei in the fng frame by a single ' rotation produces

T ni D ei cos ; 0;  ei sin  ;

(14.58)

thus completing the definition of all terms in (14.54). Obviously, the result of substituting of all the vectors into (14.54) is cumbersome; however it demonstrates how the Earth sidereal rate can be accounted for. The corresponding linear and angular momentum equations can be obtained by applying the second Newtonian law; the procedure is similar to the simplified case presented above and resulted in the (14.26) and (14.32). Utilizing the same set of assumptions (A1–A2) and resolving all the external forces and moments with respect to the CG in the body frame results in the same angular momentum equation; however, the kinematic and the linear momentum equations need to be modified. Applying the second Newtonian law to the linear motion of the CG and accounting for a new result in (14.52) and (14.53) gives cg

cg

cg

Vi D Rbe Vb C ei  ri

 cg 

P C ¨ C eb  Rbe Vcg C eb  .eb  Rib rcg Fb D m V i / ; b b

(14.59) (14.60)

where Fb , as before, is the sum of all externally applied forces applied at CG resolved in the body frame. Equations (14.59) and (14.60) are the new relations derived in a true inertial frame fi g, thus accounting for the rotating Earth. To give a reader a sense of numerical significance of the resulting acceleration, the following numerical example compares the contribution of the Coriolis and the centripetal terms with an assumption that a UAV is at the constant altitude in the wing-level flight due east and is not maneuvering; therefore, ¨bn D 0 and cg Ve D Œ0; VE ; 0T . In these conditions, the centripetal term becomes equal to the Coriolis term at the speed of 914 m/s. In turn, when at the equator latitude, the third vertical component of the Coriolis acceleration is about 0.27 m/s2 , that is, 2.7 % of the acceleration due to gravity (9.8 m/s2 ). Thus, the applicability of the simplifying flat Earth assumption becomes justified for a case of a short-duration and relatively low-speed flight of an airplane. Therefore, this new set of equations should be used when accurate modeling is required for a UAV moving faster than 600 m/s over the Earth or when long distance and duration navigation is considered.

14 Kinematics and Dynamics of Fixed-Wing UAVs

14.6

277

Conclusion

The objective of this chapter was to provide a review of the theoretical material required to enable accurate mathematical representation of the free and controlled motion of a generic fixed-wing UAV modeled as a rigid body. The key building blocks presented were the coordinate frames and their transformations, kinematics of rotation, dynamics of motion, and the definition of forces and moments acting on the airplane. The kinematics of spatial rotation is what connects the three building blocks of the “kinematics–dynamics–actions” triad. In addition to the 6DoF equations of motion describing the kinematics and dynamics of a rigid body motion, the tools and methods developed in this chapter contribute significantly into the UAV flight dynamics, system identification, control, guidance, and navigation.

References H. Ashley, M. Landahl, Aerodynamics of Wings and Bodies (Dover Books on Aeronautical Engineering) (Dover Publications, New York, 1985) R. Beard, T. McLain, Small Unmanned Aircraft: Theory and Practice (Princeton University Press, Princeton, 2012) B. Etkin, L.D. Reid, Dynamics of Flight: Stability and Control, 3rd edn. (Wiley, New York, 1995) R.L. Fearn, Airfoil aerodynamics using panel methods. Math. J. (Wolfram Research) 10(4), 15 (2008) H. Goldstein, Classical Mechanics, 2nd edn. (Addison-Wesley, Reading, 1980) P. Henne, Applied Computational Aerodynamics (Progress in Astronautics and Aeronautics) (AIAA, Washington, DC, 1990) J.L. Hess, Panel methods in computational fluid dynamics. Ann. Rev. Fluid Mech. 22, 255–274 (1990) F. Hoblit, Gust Loads on Aircraft: Concepts and Applications (AIAA Education Series, Washington, DC, 2001) P.E. Illman, The Pilot’s Handbook of Aeronautical Knowledge (McGraw-Hill Professional, New York, 1999) G.H. Kaplan, The IAU Resolutions on Astronomical Constants, Time Scales, and the Fundamental Reference Frames, vol. circular no. 163 (United States Naval Observatory, Washington, DC, 1981) I. Kroo, LinAir 4. A nonplanar, multiple lifting surface aerodynamics program. Desktop Aeronautics (2012). http://www.desktop.aero/linair.php. Accessed 10 Apr 2012 D. McRuer, I. Ashkenas, D. Graham. Aircraft Dynamics and Automatic Control (Princeton University Press, Princeton, 1999) T.J. Mueller, J.D. DeLaurier, Aerodynamics of small vehicles. Ann. Rev. Fluid Mech 35, 89–111 (2003) R. Murray, Z. Li, S. Sastry, A Mathematical Introduction to Robotic Manipulation, vol. 1 (CRC Press, Boca Raton, 1994) OSD, Unmanned Aerial Vehicles Roadmap 2000–2025 (Office of the Secretary of Defence, Washington, DC, 2001) R.M. Rogers, Applied Mathematics in Integrated navigation Systems, 2nd edn. (AIAA, Reston, 2003) M.D. Shuster, A survey of attitude representations. J. Astronaut. Sci. 41(4), 439–517 (1993)

Dynamic Model for a Miniature Aerobatic Helicopter

15

Vladislav Gavrilets

Contents 15.1 15.2 15.3 15.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Helicopter Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Component Forces and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.1 Main Rotor Forces and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.2 Engine, Governor, and Rotor Speed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.3 Fuselage Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.4 Vertical Fin Forces and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.5 Horizontal Stabilizer Forces and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.4.6 Tail Rotor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.5 Actuator Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

280 282 284 285 285 295 298 299 300 300 304 305

Abstract

A nonlinear dynamic model of a miniature aerobatic helicopter is presented in detail. A component buildup was applied in devising the model, using simplified analytical expressions for the component forces and moments. Flighttest experiments were used to estimate several key parameters, such as the equivalent torsional stiffness in the hub. The model was used to design control logic for aerobatic maneuvers performed entirely under computer control.

V. Gavrilets Flight Control and Navigation Group, Rockwell Collins Control Technologies, Warrenton, VA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 54, © Springer Science+Business Media Dordrecht 2015

279

280

15.1

V. Gavrilets

Introduction

Miniature helicopters with hingeless rotors are extremely agile compared to their full-scale counterparts due to fast shrinking of moments of inertia with size and therefore growth of control authority Mettler (2002). These vehicles can be used for tasks requiring aggressive maneuvering, such as pursuit and evasion in urban or mountainous environment, or aerial stunts for movie industry. Precise autonomous execution of aggressive maneuvers, coupled with online motion planning, can further expand utility of agile miniature rotorcraft. This created the need for an adequate nonlinear model of an aerobatic miniature helicopter. An extensive body of literature has been devoted to the dynamics of full-scale helicopters, and step-by-step procedures for creating a first-principles dynamic model were outlined (Padfield 1996; Bramwell 2001; Talbot et al. 1982). Frequencydomain system identification techniques are widely used to evaluate accuracy of linearized models at different operating conditions (NASA Ames Research Center 2000; Tischler 1996). Due to a wide range of flow conditions and their complex interaction with the wake, cross-axis coupling in the hub, and sometimes interaction of the structural modes with the rigid body modes, the models of full-scale helicopters used in simulators tend to be complex and contain a large number of states. Recent years have seen an increase in modeling efforts for small-scale helicopters. Mettler (2002) performed a comprehensive study of the features common to the dynamics of small-scale helicopters and applied frequency-domain methods for identification of linearized models. Most of the small-scale helicopters have BellHiller stabilizer bars, which slow down rotor dynamics and mitigate gust response. It was shown (Mettler 2002) that this creates relatively low-frequency lateral and longitudinal pendulum-like flapping modes, which involve fuselage, rotor, and stabilizer bar. The same effect was observed on a full-size UH-1H helicopter, equipped with a stabilizer bar with mechanical dampers instead of aerodynamic surfaces. The modes are lightly damped and must be accounted for to develop a high-bandwidth control system required for autonomous aggressive maneuvering. LaCivita et al. (2002b) developed a technique for integrating first-principles modeling and frequency-domain system identification to yield a nonlinear model for small-scale helicopters. The resulting model of the Yamaha R-50 helicopter employed 30 states and was used for H1 -based control system design for the helicopter in steady-state hovering or slow forward and backward flight (LaCivita et al. 2002a). The authors further compared accuracy of full linearized models and reduced-order linear models and came to similar conclusions with those previously noted by Mettler (2002), which advocated the use of first-order models for tip-pathplane flapping dynamics. Abbeel et al. (2006) built a nonlinear model of a small-scale aerobatic helicopter, using a combination of first-principles modeling and data fitting based on reinforcement learning. The model used fewer states than the one described in this chapter: it included rigid body states and main rotor speed only. The authors used the model and apprenticeship learning algorithms to develop control laws that enabled tracking

15 Dynamic Model for a Miniature Aerobatic Helicopter

281

of highly aggressive aerobatic trajectories (Abbeel et al. 2010) and autonomous autorotation landing (Abbeel et al. 2008). Under this approach, time histories from aerobatic flights conducted by an expert pilot were used to refine the model and tailor control law parameters to a particular task. This chapter presents a development of a first-principles-based nonlinear dynamic model of a miniature helicopter, which proved sufficiently accurate to predict helicopter dynamics in a wide range of conditions, including aerobatic flight. This work is closely based on a modeling chapter in the author’s doctoral dissertation (Gavrilets 2003), devoted to human-inspired approach to autonomous helicopter aerobatics. Most of the model parameters were measured directly, and several were estimated using data from flight-test experiments. Analytical linearization of the model with respect to forward speed yields simple models, directly used for design of feedback control laws for autonomous execution of aerobatic maneuvers. The helicopter used in the study is X-Cell .60 SE by Miniature Aircraft (1999), equipped with a 0.90 size engine and an electronic governor. This helicopter features a particularly stiff hub, which gives large control authority to cyclic actuators. The coupling in the hub for this helicopter was also shown to be negligible, which further simplified model development. The helicopter flew a fully autonomous mission that included an axial roll, a hammerhead, and a Split-S and sequences of these maneuvers (Gavrilets et al. 2002). An artist representation of a Split-S maneuver is shown in Fig. 15.1. The state trajectories in both open loop and closed loop flights were well predicted by the nonlinear model that used a single set of parameters, invariant of the task being performed. The work on the dynamic model was performed using a custom-designed and built avionics suite, which was concurrently developed. The avionics suite is described in detail in the author’s Ph.D. dissertation (Gavrilets 2003). The model has typical rigid body states with the quaternion attitude representation used in order to enable simulation of extreme attitudes (Rolfe and Staples 1986), two states for the lateral and longitudinal flapping angles, one for the rotor speed, and one for the integral of the rotor speed tracking error. This last state comes from the governor action, modeled with a proportional-integral feedback from the rotor speed tracking error to the throttle command. The model covers a large portion of the X-Cell’s natural flight envelope: from hover flight to about 20 m/s forward flight. The maximum forward speed corresponds to an advance ratio  D 0:15, which is considered as relatively low (Padfield 1996), and permits a number of assumptions (e.g., thrust perpendicular to the rotor disk; see Chen (1979)). The cross-coupling effects in the rotor hub were also shown to be negligible for this helicopter, which further simplified model development. The mathematical model was developed using basic helicopter theory, accounting for the particular characteristics of a miniature helicopter. Most of the parameters were measured directly, and several were estimated using data collected from simple flight-test experiments, involving step and pulse responses in various actuator inputs. No formal system identification procedures are required for the proposed model structure. The model’s accuracy was verified using comparison between model predicted responses and responses

282

V. Gavrilets

Fig. 15.1 Miniature helicopter undergoing a Split-S maneuver. The same maneuver was performed fully autonomously under control laws, developed based on the dynamic model described in this chapter (Drawing courtesy of Popular Mechanics magazine)

collected during flight-test data. The model was also “flown” in simulator by an expert RC pilot to determine how well it reproduces the piloted flying qualities. Analytical linearization of the model with respect to forward speed was used to derive simple linear models. These were subsequently used for a model-based design of the controllers used for the automatic execution of aerobatic maneuvers. The actual aggressive trajectories flown by the helicopter were adequately predicted by the simulation based on the developed nonlinear model. The remainder of the chapter contains a full list of model parameters with the numerical values, dynamics equations of motion, and expressions for forces and moments exerted on the helicopter by its components. Flight-test data used to validate the model for various flight regimes, including aerobatics, is provided throughout the chapter.

15.2

Helicopter Parameters

The physical helicopter parameters used for the model are given in Table 15.1. The moments of inertia around the aircraft body axes passing through the vehicle center of gravity were determined using torsional pendulum tests (Harris 1996). The crossaxis moments of inertia are hard to measure without a balancing device, and since they are usually small, they were neglected. The X-Cell main and tail rotors, as well as the stabilizer bar, have symmetric airfoils. The lift curve slopes of these surfaces were estimated according to their respective aspect ratios (Kuethe and Chow 1986).

15 Dynamic Model for a Miniature Aerobatic Helicopter

283

Table 15.1 Parameters of MIT instrumented X-Cell 60 SE helicopter Parameter Description m D 8:2 kg Ixx D 0:18 kg m2 Iyy D 0:34 kg m2 Izz D 0:28 kg m2 Kˇ D 54 Nm/rad f b D 0:8 Bınom D 4:2 rad/rad lat Anom ılon D 4:2 rad/rad K D 0:2 nom D 167 rad/s Rmr D 0:775 m cmr D 0:058 m amr D 5:5 rad1 CDmr0 D 0:024 CTmrmax D 0:0055 Iˇmr D 0:038 kg m2 Rtr D 0:13 m ctr D 0:029 m atr D 5:0 rad1 CDtr0 D 0:024 CTtrmax D 0:05 ntr D 4:66 nes D 9:0 ırtrim D 0:1 rad Svf D 0:012 m2 CLvf˛ D 2:0 rad1 tr vf D 0:2 Sht D 0:01 m2 CLht˛ D 3:0 rad1 idle Peng D 0:0 W max Peng D 2;000:0 W Kp D 0:01 s/rad Ki D 0:02 1/rad Sxfus D 0:1 m2 Syfus D 0:22 m2 Szfus D 0:15 m2 hmr D 0:235 m ltr D 0:91 m htr D 0:08 m lht D 0:71 m

Helicopter mass Rolling moment of inertia Pitching moment of inertia Yawing moment of inertia Hub torsional stiffness Stabilizer bar Lock number Lateral cyclic to flap gain at nominal rpm Longitudinal cyclic to flap gain at nominal rpm Scaling of flap response to speed variation Nominal m.r. speed m.r. radius m.r. chord m.r. blade lift curve slope m.r. blade zero lift drag coefficient m.r. max thrust coefficient m.r. blade flapping inertia t.r. radius t.r. chord t.r. blade lift curve slope t.r. blade zero lift drag coefficient t.r. max thrust coefficient Gear ratio of t.r. to m. r. Gear ratio of engine shaft to m. r. t.r. pitch trim offset Effective vertical fin area Vertical fin lift curve slope Fraction of vertical fin area exposed to t.r. induced velocity Horizontal fin area Horizontal tail lift curve slope Engine idle power Engine max power Proportional governor gain Integral governor gain Frontal fuselage drag area Side fuselage drag area Vertical fuselage drag area m.r. hub height above c.g. t.r. hub location behind c.g. t.r. height above c.g. Stabilizer location behind c.g.

284

V. Gavrilets

The effective torsional stiffness in the hub was estimated from angular rate responses to step commands in cyclic, as described in Sect. 15.4.1.3. Note that in the table, “m.r.” stands for the main rotor, and “t.r.” stands for the tail rotor.

15.3

Equations of Motion

The rigid body equations of motion for a helicopter are given by the Newton-Euler equations shown below. Here the cross products of inertia are neglected: uP D vr  wq  g sin  C .Xmr C Xfus / =m vP D wp  ur C g sin  cos  C .Ymr C Yfus C Ytr C Yvf / =m wP D uq  vp C g cos  cos  C .Zmr C Zfus C Zht / =m pP D qr.Iyy  Izz /=Ixx C .Lmr C Lvf C Ltr / =Ixx qP D pr.Izz  Ixx /=Iyy C .Mmr C Mht / =Iyy rP D pq.Ixx  Iyy /=Izz C .Qe C Nvf C Ntr / =Izz The set of forces and moments acting on the helicopter are organized by components: ./mr for the main rotor, ./tr for the tail rotor, ./fus for the fuselage (includes fuselage aerodynamic effects), ./vf for the vertical fin, and ./ht for the horizontal stabilizer. These forces and moments are shown along with the main helicopter variables in Fig. 15.2. Qe is the torque produced by the engine to counteract the aerodynamic torque on the main rotor blades. The helicopter blades rotate clockwise when viewed from above; therefore, Qe  0. In the above equations, it was assumed that the fuselage center of pressure coincides with the c.g.; therefore, the moments created by the fuselage aerodynamic forces were neglected.

Fig. 15.2 Moments and forces acting on helicopter

15 Dynamic Model for a Miniature Aerobatic Helicopter

285

The rotational kinematic equations were mechanized using quaternions (Rolfe and Staples 1986). The inertial velocities are derived from the body-axis velocities by a coordinate transformation (flat-Earth equations are used) and integrated to obtain inertial position. A fourth-order Runge-Kutta integration method is used, with an integration step of 0.01 s.

15.4

Component Forces and Moments

15.4.1 Main Rotor Forces and Moments 15.4.1.1 Thrust For the main rotor thrust, it was assumed that the inflow is steady and uniform. According to Padfield (1996, p. 126), the time constant for settling of the inflow transients at hover is given by  D

0:849 4 trim mr

(15.1)

Induced velocity at hover trim condition can be determined from simple momentum theory: r mg Vimr D D 4:2 m/s (15.2) 2 2 Rmr tip

The tip speed of the main rotor is Vmr D mr Rmr D 125:7 m/s, from which tip the inflow ratio is imr D Vimr =Vmr D 0:033. Therefore, the time it takes for the inflow to settle is  D 0:038 s, which is significantly faster than the rigid body dynamics. During the maneuvers requiring large thrust variations, the time constant may change substantially. However, as shown in the section on the main rotor flapping dynamics, the X-Cell cyclic control authority is dominated by the hub torsional stiffness, which makes the modeling of the inflow transients less critical. A momentum theory-based iterative scheme given by Padfield (1996, p. 123) was adapted to compute the thrust coefficient and inflow ratio as a function of airspeed, rotor speed, and collective setting. Flapping angles were neglected in the computation of the rotor thrust. The blades of the main rotor have no twist. The influence of the cyclics and the roll rate on thrust are of second order for advance ratio range  < 0:15 and were neglected as well. An empirically determined maximum thrust coefficient was introduced, since momentum theory does not take into account the effect of blade stall. The thrust coefficient is given by (omitting the “mr” index) CT D

T

.R/2 R2

(15.3)

286

V. Gavrilets

where T is the main rotor thrust. Then, the following system of equations can be solved iteratively: 0 D

CTideal

q

CT

2 w 2 C . 0  z /2

a D 2

    2 z  0 1 C 0 C 3 2 2

8 ideal if  CTmax  CTideal  CTmax ˆ < CT CT D C max if CTideal < CTmax T ˆ : max CT if CTmax < CTideal CTmax D

T max

.R/2 R2

(15.4)

(15.5)

(15.6)

(15.7)

Here p .u  uwind /2 C .v  vwind /2 D – advance ratio R w  wwind z D – normal airflow component R 2c – solidity ratio R a  lift curve slope

D

0  commanded collective angle w  coefficient of nonideal wake contraction T

max

D 2:5 mg – maximum rotor thrust

Based on momentum theory, the rotor wake far downstream contracts by a factor of 2 (Padfield 1996, p. 116). A coefficient w was introduced to account for nonideal wake contraction and the power lost due to the nonuniform velocity and pressure distribution in the wake. An approximate value for this coefficient was determined to be w D 0:9. Hence, the iterative scheme given in Padfield (1996, p. 123) is modified as follows. First, define the zero function: g0 D 0 

CT , where 2 w ƒ1=2

ƒ D 2 C . 0  z /2

15 Dynamic Model for a Miniature Aerobatic Helicopter

287

and thrust coefficient CT is given by Eq. (15.5). Apply Newton’s iterative scheme:   0j C1 D 0j C fj hj 0j   g0 hj D  dg0 =d 0 0 D 0

j

An explicit expression for hj :   2 w 0j ƒ1=2  CT ƒ   hj D  2 w ƒ3=2 C a 4 ƒ  CT z  0j

Padfield (1996, p. 123) suggests a constant value of the convergence rate coefficient fj D 0:6. Note that at hover the denominator of Eq. (15.4) is zero when the vertical velocity is equal to the inflow velocity. This condition corresponds to a vortexring state, which cannot be modeled adequately by the momentum theory. Instead, the denominator is numerically separated from zero. In general, this condition is avoided in flight because it leads to a loss of control. One has to keep in mind that the simulation does not adequately represent the helicopter dynamics when vortex-ring conditions exist on either the main or the tail rotor. Furthermore, strictly speaking, the momentum theory applies only to a fully developed steadystate flow in ascending flight. Empirical corrections for descending flight, cited by Padfield (1996, p. 118), could be used to make thrust prediction somewhat more accurate. The momentum theory approach was previously shown to be adequate for estimating steady-state main rotor thrust both at hover and in fast forward flight. Results of the wind tunnel tests with a 5-ft-diameter rotor are given by Harris (1972) and summarized by Bramwell (2001, pp. 109–114). It was shown that the blade lift curve slope coefficient a can be determined from experiments such that the momentum theory accurately predicts thrust for a wide range of advance ratios and collective pitch angles. To test the applicability of momentum theory-based thrust calculation to transient response, flight data was gathered for collective pitch pulse responses at hover. At hover, the vertical acceleration can be represented by a linear relation: az D Zw w C Zcol ıcol (15.8) The vertical speed damping stability derivative Zw and the collective pitch control derivative Zcol can be obtained analytically by linearization of the momentum theory equations (Padfield 1996, pp. 219, 229). For hover,

288

V. Gavrilets

Filtered az, m/sec2

4 2 0 2 Measured Model

4 6 436

438

440 442 time, sec

444

446

Fig. 15.3 Modeling vertical acceleration response at hover

Zw D 

Zcol D 

.R/ R2 2a 0 m 16 0 C a

.R/2 R2 8 a 0 m 3 16 0 C a

(15.9)

(15.10)

The same value for the blade lift curve slope was used as the one determined for the particular airfoil used in the tests summarized by Harris (1972), a D 5:5. This value is consistent with the high aspect ratio of the main rotor blades, if the blades are considered as a wing (Kuethe and Chow 1986). The mean values were subtracted from the collective command and the vertical acceleration measurement; the signals were filtered with a first-order low-pass filter with time constant of 0.2 s. The digital models of the analog low-pass filter and the servomechanism were applied to the collective command for consistency. Figure 15.3 shows the comparison of the computed vertical acceleration from Eq. (15.8) and actual acceleration. As can be seen, the model, based on linearization of the momentum theory, agrees well with flight data.

15.4.1.2 Torque The main rotor torque can be approximated as a sum of induced torque due to generated thrust and torque due to profile drag on the blades (Padfield 1996, p. 116): CD0 CQ D D CT . 0  z / C 2 3 8

.R/ R Q

  7 2 1C  3

(15.11)

15 Dynamic Model for a Miniature Aerobatic Helicopter

289

where CQ is the torque coefficient and CD0 is the profile drag coefficient of the main rotor blade. The profile drag is not significantly affected by changes in the collective setting. Thus, the yawing moment produced by the main rotor is given by Qmr D CQ .R/2 R3

(15.12)

Profile drag coefficient of the main rotor blade was estimated as CD0 D 0:024. Underestimation of the profile drag coefficient would lead to overprediction of the main rotor speed in windmilling flight conditions, which occur during autorotation and some agile maneuvers. Except for the hover condition, the rotor in-plane force, which contributes to the drag and side force, is substantially smaller than the drag provided by the fuselage and the side force from the fuselage and the empennage. This force was neglected in the calculations; effectively it was lumped with the fuselage forces, estimated in Sect. 15.4.3. The moments due to the in-plane force are much smaller than those due to the blade flapping, since the torsional stiffness of the hub retention is high on the X-Cell.

15.4.1.3 Main Rotor Moments and Flapping Dynamics The main rotor flapping angle ˇ can be represented as a Fourier series of the blade azimuth angle , with only the first three coefficients retained (Padfield 1996, p. 32): ˇ. / D a0 C a1 cos

C b1 sin

(15.13)

Flapping of the teetering stabilizer bar can be represented by a similar equation without the constant term since no coning takes place: ˇs . / D a1s sin

C b1s cos

(15.14)

Stabilizer bar flapping contributes to the change of the main rotor blade pitch angle through a mechanical linkage: . / D 0 C lon sin

C lat cos

C ks ˇs

(15.15)

The swashplate deflections change the cyclic pitch angle of both the main rotor and the stabilizer bar. Coupled second-order differential equations can be developed for Fourier coefficients of the main rotor and stabilizer bar flapping. It can be shown (Padfield 1996, pp. 33–35) that the undamped natural frequency of the flapping motion is close to the rotor speed mr and the damping ratio can be approximated by =8, where  is the Lock number of the blades being considered (main rotor or stabilizer bar). The Lock number represents the ratio of aerodynamic to inertial forces and is defined as D

caR4 Iˇ

(15.16)

290

V. Gavrilets

For the main rotor blades, the Lock number is relatively high, mr  3:7; therefore, the flapping motion is well damped. For a step response, this corresponds to the settling time (to within 5 % of the steady-state value) of 24=  D 0:039 s. For the stabilizer bar, with its small aerodynamic surfaces, the Lock number is low, fb  0:8, and the corresponding settling time is 0.144 s. Earlier work on modeling of small-scale rotorcraft with Bell-Hiller stabilizer bars (Mettler et al. 2002b; Mettler 2002; LaCivita et al. 2002b) showed that the main rotor and stabilizer bar flapping dynamics can be lumped and represented by tip-path plane (TPP) flapping dynamics with only two states. This result was based on frequency-domain identification and comparison of reduced and full order transfer functions for attitude dynamics. Furthermore, coupling of the lumped flapping dynamics and rigid body pitch and roll motions leads to pronounced second-order characteristics (von Grunhagen et al. 1996; Talbot et al. 1982; Mettler 2002; LaCivita et al. 2002b). These modes are lightly damped and should be explicitly accounted for in designing high-bandwidth attitude or rate control systems (Mettler et al. 2000, 2002a). The lateral and longitudinal flapping dynamics were represented by the first-order equations: Bı b1 1 @b1 v  vw bP1 D p  C lat ılat  e e @v R e aP 1 D q 

a1 1 C e e



@a1 w  ww @a1 u  uw C @ R @z R

(15.17)  C

Aılon ılon e

(15.18)

where Bılat and Aılon are effective steady-state lateral and longitudinal gains from the cyclic inputs to the main rotor flap angles; ılat and ılon are the lateral and longitudinal cyclic control inputs (pilot stick or control system outputs); uw , vw , and ww are the wind components along, respectively, X, Y, and Z helicopter body axes; and e is the effective rotor time constant for a rotor with the stabilizer bar. Frequency-domain identification showed that the pitch and roll cross-coupling flapping coefficients are approximately an order of magnitude less than the direct coefficients for the X-Cell (Mettler 2002) and were neglected. This result holds for large-amplitude inputs as well. For example, during an axial roll maneuver, the pitch rate remains close to zero with no pilot compensation, similarly to the roll rate during a loop. This natural decoupling of the cyclic responses makes the X-Cell a particularly attractive helicopter for aerobatics. The dominant rotor moments are the control moments produced by the rotor flapping. In the following pages, the moments in the roll direction are described (resulting from the lateral TPP flapping b1 ). Figure 15.4 shows the rotor moments that are acting on the fuselage. The first contribution results from the restraint in the blade attachment to the rotor head. The restraint can be approximated using a linear torsional spring with a constant stiffness coefficient Kˇ , resulting in a roll moment Mk;lat D Kˇ b1 . The second contribution results from the tilting of the thrust vector. Assuming that the thrust vector is perpendicular to the TPP, the thrust vector will tilt proportionally to the rotor flapping angles. The moment arm is the distance hmr

15 Dynamic Model for a Miniature Aerobatic Helicopter

291

Fig. 15.4 Rotor moments acting on the helicopter fuselage

between the rotor head and the helicopter center of gravity, resulting in a lateral moment Mh;lat D T hmr b1 . The total main rotor rolling moment, entering the rigid body equations of motion, is represented by Eq. (15.19):   Lmr D Kˇ C T hmr b1

(15.19)

Similarly, the pitching moment is given by Eq. (15.20):   Mmr D Kˇ C T hmr a1

(15.20)

To determine the parameters entering the flapping equations, one can split up the problem into “slow” and “fast” dynamics. First, one can notice that the dihedral derivatives are important only at the low-frequency spectrum of the dynamics. At high frequencies (above 0.5 Hz), the transfer functions from cyclic inputs to angular rates can be approximated by second-order transfer functions, derived by omitting the translational flapping derivatives in Eqs. (15.17) and (15.18) and combining it with Eqs. (15.19) and (15.20): 2 !nq q Aı  lon 2 2 ılon e s C 1=e s C !nq

(15.21)

2 !np p Bı  lat 2 2 ılat e s C 1=e s C !np

(15.22)

Pitching dynamics in fast forward flight is significantly influenced by the horizontal tail, which provides a stabilizing effect, and the main rotor flapping due to vertical speed, which provides a destabilizing effect. Therefore, the longitudinal cyclic to pitch rate transfer function given in Eq. (15.21) is valid in low-speed flight only. Here the undamped natural frequencies of the longitudinal and lateral fuselage-rotor modes are

292

V. Gavrilets

s !nq D s !np D

Tmr hmr C Kˇ Iyy Tmr hmr C Kˇ Ixx

Note that for hover and straight and level flight, Tmr  mg. The distance between the main rotor hub and the helicopter center of gravity can be measured. The moments of inertia were determined with the torsional pendulum tests (Harris 1996). The natural frequencies of the lightly damped second-order systems can be easily determined by counting oscillation periods in a recorded step response, thereby providing an estimate of the hub torsional stiffness. These parameters are given in Table 15.1. An approximate value of the damping time constant for the flapping motion (Mettler 2002) is given in Eq. (15.23): e D

16  0:1 s fb mr

(15.23)

Note that the damping is proportional to the stabilizer bar Lock number, making it small. The steady-state cyclic to rate gains depend on the swashplate gearing. Experiments have also indicated that the values of Bılat and Aılon grow with the rotor speed. This effect was approximated as a function of effective dynamic pressure, or square of the rotor speed:  Bılat D

Bınom lat

Aılon D

Anom ılon



 nom  nom

2 rad/rad 2 rad/rad

nom D 167 rad/s These gains were determined by matching DC gain of the angular rate responses to steps in cyclics. Final verification of the derived parameters is provided by simulation of the linear systems described in Eqs. (15.21) and (15.22) and comparison with the flight-test data. Figure 15.5 shows the actual and simulated roll rates for a segment including an axial roll maneuver. The helicopter undergoes negative rotor loading during inverted portion of the maneuver, leading to 15 % lower roll rate in that segment than predicted by the simplified linear model. Figure 15.6 shows the actual and simulated pitch rates for a segment of 15 m/s forward flight with pulse commands on longitudinal cyclic. Note that the linearized models of angular rate dynamics are adequate for both small- and large-amplitude motion. The flapping due to translational velocity is described by the flapping derivatives @a1 =@ and @b1 =@v . The longitudinal flapping due to the forward speed increase

15 Dynamic Model for a Miniature Aerobatic Helicopter Fig. 15.5 Actual and model roll rate response during axial roll maneuver

293

200 Measured Model Output

deg/sec

150

100

50

0

50 282

283

284

285

286 287 time, sec

288

289

290

100

Q, deg/sec

50

0

50 Measured Model Output

100

Fig. 15.6 Actual and model pitch rate response in low-speed flight

150

0

1

2

3

4

5

time, sec

is caused by an increased lift on the advancing blade with respect to the retreating blade, which turns into a flap-back moment on the main rotor due to the 90ı gyroscopic phase lag. A theoretical value for the steady-state longitudinal flapping for a rotor without a stabilizer bar is given by Bramwell (2001, p. 107) a1 D

2 .4ıcol =3  0 /  2 .4ıcol =3  0 / 1 C 32 =2

(15.24)

While this expression is valid for a teetering rotor at hover, theoretical approximation for a hingeless rotor without a stabilizer bar is very close (Padfield 1996). The stabilizer bar dramatically reduces flapping response to gusts, and

294

V. Gavrilets

Eq. (15.24) cannot be used for predicting dihedral effect on a rotor equipped with one. Since this derivative plays a primary role in the frequency and damping of the phugoid mode, which is very slow, it is difficult to estimate with the frequencydomain identification methods (Mettler 2002). An open loop excitation would have to last for much longer than it takes for the helicopter to diverge; therefore, the necessary pilot’s feedback would bias an estimate of the derivative (Ljung 1999). A scaling coefficient was introduced in Eq. (15.24), which was linearized to yield   4ıcol @a1 D 2K  0 (15.25) @ 3 A rough estimate for the scaling coefficient K can be obtained by matching the steady-state cyclic input in forward flight at constant speed (maintained with the velocity-tracking feedback controller) with that predicted by the simulation in the same conditions. An estimate for X-Cell 60 yielded K D 0:2, which implies that the stabilizer bar reduces steady-state flapping response to forward speed by a factor of 5. From Eq. (15.24) and the rotor symmetry, one can conclude that the longitudinal and lateral dihedral derivatives are equal in magnitude and, in both cases, cause the rotor to flap away from the incoming air: @b1 @a1 D @v @

(15.26)

Positive Z-axis velocity causes higher lift on advancing blade, which results in a flap back of the rotor; this effect is captured by the stability derivative @a1 =@z in Eq. (15.17). An analytical estimate of the derivative is adapted (Bramwell 2001, p. 159) to accommodate backward flight and scaled by the same coefficient K to reflect the effect of the stabilizer bar: @a1 162 162 sign   K sign  D K  @z .1  2 =2/ .8 jj C a / 8 jj C a

(15.27)

15.4.1.4 Rotor Forces For small advance ration flight ( < 0:15), one can assume that the thrust vector is perpendicular to the TPP. The small flapping angles (below 10ı ) allow one to use linear approximation for the main rotor force components along the helicopter body axes. As was stated above, the in-plane rotor force was lumped with the fuselage forces and is not accounted for in the equations below: Xmr D Tmr a1 Ymr D Tmr b1 Zmr D Tmr

15 Dynamic Model for a Miniature Aerobatic Helicopter

295

15.4.2 Engine, Governor, and Rotor Speed Model The rotor speed dynamics is modeled by the following equation: P D rP C 1 ŒQe  Qmr  ntr Qtr   Irot

(15.28)

where Qe is the engine torque (positive clockwise), Qmr D CQ .R/2 R3 is the main rotor torque (positive counterclockwise), Qtr is the tail rotor torque, ntr is the tail rotor gear ratio, Irot is the total rotating inertia referenced to the main rotor speed, and  is the rotor speed. The engine torque depends on the throttle setting ıt and rotor speed and is usually represented by engine maps or lookup tables. The maps for the engine were not available, and a simplified representation of the engine torque is suggested. Assume that engine power is proportional to the throttle setting: Pe D Pemax ıt (15.29) where 0 < ıt < 1. Then, the torque is Qe D

Pe 

(15.30)

The engine torque response to throttle changes can be considered instantaneous, since the time lags associated with air intake, fuel flow, and combustion are very small compared to vehicle dynamics. In the absence of manufacturer data, the governor can be modeled as a proportional-integral feedback controller, maintaining commanded rotor speed by changing the throttle: ıt D Kp  .c  / C Ki  !i

(15.31)

!P i D c   where c is the rotor speed command and Kp and Ki are proportional and integral feedback gains. The anti-windup logic resets the integrator state value !i in case computed throttle command is saturated. Throttle servo dynamics is much faster than the rotor speed dynamics and was neglected in the model. To determine the parameters of the given engine/governor model, time response to the rotor speed step command was analyzed. Consider linearization of Eqs. (15.28)–(15.31) around a nominal operating point, for example, hovering flight, neglecting the yawing acceleration and the tail rotor torque. The states of the linear system will be the rotor speed deviation from the nominal ! and an integral of the rotor speed tracking error !i . The inputs are !c , a variation of the rotor speed command, and ıc , a variation of the collective angle from the trim setting. The resulting linear system is given in Eq. (15.32):

296

V. Gavrilets

d dt



! !i



 0  Pemax Ki #   C Pemax Kp I  I1rot 3Qmr ! rot D !i 1 0 " P max K #  @CQ p 1 0 e  Q !c mr Irot Irot CQ @ıc C ıc 1 0 "

(15.32)

0 can be computed, and for the From Eq. (15.11), the main rotor torque at hover Qmr 0 X-Cell with the parameters given in Table 15.1, Qmr  6:3 Nm. The characteristic polynomial of the system is given in Eq. (15.33):

. / D 2 C

0 3Qmr C Pemax Kp P max Ki C e Irot Irot

(15.33)

To estimate the coefficients of the characteristic polynomial, the following test was performed (Sprague et al. 2001). The helicopter was kept at hover at 1,600 rpm, and a 100 rpm step input in rotor speed was commanded to the governor from a remote control. The rotor speed measurement was not available directly in the instrumentation package. Instead, the sound of the engine was recorded with a handheld camcorder. Next, a time-frequency decomposition analysis (Feron et al. 1998) was applied to determine frequency content of the engine noise as a function of time. In such an analysis, a signal of limited duration and frequency, called a wavelet, defined by its central frequency and width, is convoluted in the time domain with the data. The output in the time domain will have a larger magnitude when the wavelet’s central frequency is present in the signal than when the central frequency is missing. The spectral content of the signal as a function of time can then be determined by repeating this computation over a range of frequencies. Many frequency bands appear in the sound spectrum of the engine noise, representing harmonics. Figure 15.7 (top) shows the result of the Morlet wavelet time-frequency analysis performed on the engine noise during a step change in RPM setting; two harmonics are indicated. The same behavior appears on each harmonic, providing the opportunity to fine-tune the model to a number of step responses, thereby reducing measurement error due to external noise. The governor/engine system was approximated with a second-order system, whose response to a commanded RPM step input appears at the bottom of Fig. 15.7. The system’s damping ratio is  D 0:63, and its natural frequency is !n D 1:3 rad/s. By matching the coefficients of the characteristic polynomials, one can obtain the expressions given in Eq. (15.34): 2!n D !n2 D

0 3Qmr C Pemax Kp Irot

Pemax Ki Irot

(15.34)

15 Dynamic Model for a Miniature Aerobatic Helicopter

297

Frequency (Hz)

1400

1350

1300

1250

1200 0

2

4

0

2

4

6

8

10

12

6

8

10

12

Amplitude

1

0 Time (sec)

Fig. 15.7 (top) Two frequency bands in the engine noise spectrum; (bottom) simulated response to rotor speed step command

A total kinetic energy of all rotating components is 2Iˇmr 2 C Ies .nes /2 C 2Iˇtr .ntr /2 D .2Iˇmr C Ies n2es C 2Iˇtr n2tr /2 , where Iˇmr and Iˇtr are, respectively, the main and the tail rotor blade inertias, Ies is the inertia of the engine shaft and all components rotating at the engine speed, ntr is the tail rotor gear ratio, and nes is the engine gear ratio. Therefore, the rotating inertia referenced to the main rotor speed can be represented as Irot D 2Iˇmr C Ies n2es C 2Iˇtr n2tr . The most important contribution comes from the main rotor blades. The tail rotor inertia, after scaling with the gear ratio squared, amounts to 5 % of the main rotor inertia. The rotating inertia referenced to the engine speed is harder to estimate, but an upper bound can be found by estimating the total mass of rotating components (0.2 kg) and its effective radius of inertia (0.04 m). One thus arrives at an estimate for Irot equal to 2.5 inertias of the main rotor blade. Using this value, and matching coefficients of the characteristic polynomial according to Eq. (15.34), estimates for the proportional and integral governor gains were obtained. The model could be further refined if real-time rotor speed data were available, or the actual governor gain and the engine maps were available from the manufacturer datasheets.

298

V. Gavrilets

This engine/governor model is simplified. However, it reflects the trends which are important in some very aggressive maneuvers which involve large and rapid variation of the aerodynamic torque on the rotor. First, tight governor feedback keeps the rotor speed close to the nominal setting. Second, an increase in the aerodynamic torque leads to a temporary decrease in rotor speed and a lagged application of the yawing torque to the airframe. The reverse is true for windmilling flight, in which the rotor extracts energy from the air, and leads to an increase in rotor speed, and lagged decrease in torque applied to the airframe. The inaccuracies in the model can make feedforward compensation of the main rotor torque with the tail rotor difficult to tune; a tight yaw rate feedback to the tail rotor pitch is much more effective and is routinely used by R/C pilots in the form of a yaw rate gyro.

15.4.3 Fuselage Forces For hover flight and forward speeds well below the induced velocity at hover (4.5 m/s for X-Cell), the rotor downwash is deflected by the forward and side velocity. This deflection creates a force opposing the movement. One can express the X and Y drag forces created by the fuselage in this flight regime by 1 2 u Xfus D Sxfus Vimr 2 Vimr 1 2 v Yfus D Syfus Vimr 2 Vimr where Sxfus and Syfus are effective drag areas of the fuselage in the X and Y directions. When the forward speed is higher than the rotor induced velocity, the fuselage drag can be modeled as the drag of a flat plate exposed to dynamic pressure. In this case, the perturbations to the fuselage forces can be expressed as 1 u Xfus D Sxfus Ue2 2 Ue 1 v Yfus D Syfus Ue2 2 Ue where Ue is the trim airspeed. Considering the above equations, fuselage forces can be approximated by V1 D

q

u2a C v2a C .wa C Vimr /2

Xfus D 0:5 Sxfus ua V1

15 Dynamic Model for a Miniature Aerobatic Helicopter

299

Yfus D 0:5 Syfus va V1 Zfus D 0:5 Szfus .wa C Vimr / V1 where Sxfus , Syfus , and Szfus are effective frontal, side, and vertical drag areas of the fuselage and ua , va , and wa are fuselage center of pressure velocities with respect to air (i.e., ua D u  uw , where uw is the projection of wind velocity vector on the X body axis). One can neglect small moments generated by the fuselage and assume that the fuselage center of pressure coincides with the helicopter center of gravity. Based on the fuselage projection areas, one can assume that Syfus  2:2Sxfus , Szfus  1:5Sxfus . Effective frontal drag area can be determined from the average pitch angle required to maintain a certain forward speed. This is best done under automatic control in velocity hold mode. In a steady trimmed flight, mg  0:5 U 2 Sxfus . A pitch angle of 10ı was required to maintain 14.5 m/s forward speed, which resulted in the estimate Sxfus D 0:1 m2 .

15.4.4 Vertical Fin Forces and Moments The side force generated by the vertical fin can be approximated as follows:   tr C jvvf j vvf Yvf D 0:5 Svf CLvf˛ V1

(15.35)

where Svf is the vertical fin area, CLvf˛ is its lift curve slope, p tr D ua  ua C wtr  wtr is the axial velocity at the location of the tail rotor hub, V1 vvf is the side velocity relative to air at the location of the vertical fin, and wt r is the vertical velocity (same as for the tail rotor): tr Vitr  ltr r vvf D va  vf

(15.36)

wtr D wa C ltr q  K  Vimr

(15.37)

tr Here Vitr is the induced velocity of the tail rotor (see Eq. (15.50)), r is yaw rate, vf is the fraction of the vertical fin area exposed to full induced velocity from the tail rotor, ltr is the vertical distance between the c.g. and tail rotor hub, which is about the same distance to the center of pressure of the vertical fin, Vimr is the main rotor induced velocity, and K is the wake intensity factor, calculated in the tail rotor section. To accommodate for stall of the vertical fin (McConley 1998), the absolute value of the vertical fin side force is limited by

jYvf j  0:5 Svf



tr V1

2

C v2vf

(15.38)

300

V. Gavrilets

The vertical fin side force creates a yawing moment and a small rolling moment due to the offsets from the c.g.: Nvf D Yvf ltr Lvf D Yvf htr

15.4.5 Horizontal Stabilizer Forces and Moments The destabilizing effect of the main rotor flapping due to vertical speed is offset by the weathervaning provided by the horizontal tailplane. The horizontal tail produces lift and a stabilizing pitching moment around the center of gravity. An effective vertical speed at the horizontal tail location is determined, assuming that the stabilizer may be fully or partially submerged in the downwash of the main rotor: wht D wa C lht q  K  Vimr

(15.39)

The same wake intensity factor is used for the horizontal fin as for the vertical fin and the tail rotor. Next, the Z-force generated by the horizontal stabilizer is determined according to   Zht D 0:5 Sht CLht˛ jua j wht C jwht j wht (15.40) where Sht is the horizontal stabilizer area and CLht˛ D 3:0 is its lift curve slope. To accommodate for the stall of the horizontal stabilizer (McConley 1998), the absolute value of the horizontal stabilizer lift is limited by   jZht j  0:5 Sht u2a C w2ht

(15.41)

Finally, the pitching moment generated by the horizontal stabilizer is Mht D Zht  lht

(15.42)

15.4.6 Tail Rotor The tail rotor is subjected to a wide range of flow conditions, including those where the thrust-inflow iteration algorithm (given for the main rotor in Eqs. (15.4) and (15.5)) would fail (e.g., when the tail rotor operates in its own wake at a low in-plane airspeed). The thrust-inflow iteration equations for the tail rotor were linearized around the trim conditions corresponding to a no-sideslip flight at a range of forward speeds. The tail rotor thrust at such trim conditions is always nonzero to compensate for the main rotor torque, and zero sideslip implies that there is no airflow component normal to the rotor disk; therefore, Eqs. (15.4) and (15.5) are applicable:

15 Dynamic Model for a Miniature Aerobatic Helicopter Fig. 15.8 Linearized tail rotor side force due to side velocity

301

0.03 Y trv 0.04

sec1

0.05 0.06 0.07 0.08 0.09

0

5

10 15 Ue, m/sec

20

25

CTtr tr D

 @CTtr  jtr j ; trz D 0; ırtrim tr @z

(15.43)

CTtrır D

 @CTtr  jtr j ; trz D 0; ırtrim @ır

(15.44)

z

The partial derivatives in Eqs. (15.43) and (15.44) were computed numerically. Simple approximate analytical expressions for the tail rotor coefficients can be obtained by adapting those used for the main rotor coefficients (Padfield 1996, pp. 219, 229). They fall within 15 % of those computed via numerical or exact analytical linearization. The resulting nondimensional coefficients were used to calculate the corresponding dimensional stability derivatives: Yvtr D CTtr tr

ft tr Rtr Rtr2 m

(15.45)

Yıtrr D CTtrır

ft .tr Rtr /2 Rtr2 m

(15.46)

z

where ft is the fin blockage factor, as suggested in Padfield (1996, p. 142): ft D 1:0 

3 Svf 4 Rtr2

The tail rotor speed is given by t r D ntr mr , where ntr is the gear ratio given in Table 15.1. For reference, the computed dimensional derivatives are provided in Figs. 15.8 and 15.9. Finally, the side force generated by the tail rotor is given in Eq. (15.47): Ytr D mYıtrr ır C mYvtr trz tr Rtr (15.47)

302

V. Gavrilets

Fig. 15.9 Linearized tail rotor side force due to tail rotor pitch

4.6 Yδr

4.8 5

m/sec2

5.2 5.4 5.6 5.8 6 6.2 6.4

0

5

10 15 Ue, m/sec

20

25

In order to calculate Ytr , one needs to determine the normal (trz ) and the inplane (tr ) tail rotor inflow components. The main rotor wake affects the tail rotor thrust in a complex way; to model this influence accurately, an extensive modeling of the wake is required. It was decided to approximate just the increase in an apparent in-plane velocity seen by the tail rotor. For this, determine the main rotor wake intensity factor K . The geometry calculations are equivalent to those given in Leishman (2000), but computationally more efficient since an explicit evaluation of the trigonometric functions is avoided. Calculate the following variables (tangents of the angles determining the geometry): gi D

ltr  Rmr  Rtr htr

gf D

ltr  Rmr C Rtr htr

First, the tail rotor is out of the downwash if Vimr  wa , in which case there is an effective upwash. Next, at low enough forward speed with respect to air, the tail rotor is out of the wake as well. This can be represented by the condition: ua  gi Vimr  wa In both of these cases, K D 0. The tail rotor is fully in the wake if ua  gf Vimr  wa

(15.48)

15 Dynamic Model for a Miniature Aerobatic Helicopter

303

In the far wake, the downwash is twice the value at the rotor. It was assumed that K D 1:5 when the tail rotor is fully immersed. In the remaining case, when the tail rotor is partially immersed, assume a linear growth of the wake intensity factor with the forward speed: ua  gi Vimr wa K D 1:5 (15.49) gf  gi The derived expression is used to calculate the vertical component of airspeed at the tail rotor location, as shown in Eq. (15.37). Next, determine an advance ratio for the tail rotor: tr D

u2a C w2tr tr  Rtr

Velocity component normal to the tail rotor is given by vtr D va  ltr  r C htr  p and in nondimensional form ztr D

vtr tr  Rtr

The magnitude of the resulting tail rotor thrust is limited based on the assumed maximum thrust coefficient to model stall of the blades and other viscous losses: tr D ft CTtrmax .tr Rtr /2 Rtr2 Ymax tr jYtr j  Ymax

The yawing and small rolling moments due to offsets from the c.g. are computed as follows: Ntr D Ytr ltr Ltr D Ytr htr The tail rotor induced velocity, used in the calculation of the vertical fin side force (see Eq. (15.36)), needs to be computed. Using the same derivation as for the main rotor (Padfield 1996, pp. 115–123), the inflow ratio is approximated by Eq. (15.50):  tr0 D ztr  2

2CTtr  ır atr tr



2 1 C tr 3 2

 (15.50)

where CTtr is the computed tail rotor thrust coefficient, atr is the tail rotor blade lift 2ctr curve slope given in Table 15.1, and tr D R is the tail rotor solidity ratio. tr

304

V. Gavrilets

Finally, the tail rotor torque Qtr is computed similarly to Eqs. (15.11) and (15.12) using the tail rotor parameters in place of the main rotor parameters.

15.5

Actuator Models

The command ranges for the cyclics and collective blade pitch are symmetric around the center point. The tail rotor blade pitch is offset by the trim value given in Table 15.1 such that the tail rotor pitch is computed as ır D ırcmd C ırtrim

(15.51)

The following maximum commanded deflections were set, in radians: max ılat D 0:096 max D 0:096 ılon max D 0:183 ıcol

ırmax D 0:38 cmd max max Here ılat and ılon are actual maximum cyclic pitch angles of the main rotor blades, measured statically. The gearing between servos and pitch angles of control surfaces is close to being linear. Linear functions are used to relate servo pulse-width commands to control surface deflections. Lookup tables can be used for a different gearing. Hobby servos and pulse-width generation electronics used on the helicopter result in significant quantization effects. On average, 150 steps were used to encode rail-to-rail deflection of each control surface. This results, for example, in a tail rotor command quantization of 0.3ı . Linear transfer functions are used to model the servo dynamics. Futaba S9402 servos, used for collective and cyclic deflections of the main rotor blades, were subjected to small-amplitude frequency sweeps under 35 ozin mean load and small inertia, which was assume to be representative of the actual loads experienced by the servos during the flight. The following transfer function came up as a result:

Hservo .s/ D

!n2 s=Tz C 1 s=Tp C 1 s 2 C 2!n s C !n2

where Tz D 104 s, Tp D 33 s, !n D 36 rad/s, and  D 0:5. Note that 90ı phase lag occurs at roughly 30 rad/s, which also imposes a limitation on the control system bandwidth. A fast digital servo (Futaba S9450) was used for the tail rotor pitch. Since the torque required from the tail rotor servo is much lower than that required from the swashplate actuators, no-load small-signal bandwidth tests provide an

15 Dynamic Model for a Miniature Aerobatic Helicopter

305

adequate model of the servo. As a result of the tests, the servo transfer function was approximated by a second-order system with the undamped natural frequency of 7 Hz and the damping ratio of 0.6. Conclusion

A representative nonlinear dynamic model of a miniature aerobatic helicopter was developed with first-principles methods. In addition to the rigid body states, the model includes two states to represent lateral and longitudinal flapping angles of the main rotor, one state for the main rotor speed, and one state for the integral of the main rotor speed tracking error. Flight-test data were used for determining several key parameters. The model is valid up to advance ratios  < 0:15 for a variety of maneuvering flight conditions, including negative rotor loading and high angular rates. The model can be used for developing and evaluating control design methods for demanding tasks, including aerobatics. This simplified modeling framework is suitable for a class of miniature helicopters with hingeless rotors.

References P. Abbeel, A. Coates, A.Y. Ng, Autonomous auto-rotation of an rc helicopter, in Proceedings of ISER, Athens, Greece, 2008 P. Abbeel, A. Coates, A.Y. Ng, Autonomous helicopter aerobatics through apprenticeship learning. Int. J. Robot. Res. 29, 1608–1639 (2010) P. Abbeel, V. Ganapathi, A.Y. Ng, Learning vehicular dynamics with application to modeling helicopters, in Proceedings of NIPS 18, Vancouver, British Columbia, Canada, 2006 A.R.S. Bramwell, Bramwell’s Helicopter Dynamics (AIAA, Reston, 2001) R.T. Chen, A simplified rotor system mathematical model for piloted flight dynamics simulation. Technical Memorandum 78575, NASA, 1979 E. Feron, M. Brenner, J. Paduano, A. Turevskiy, Time-frequency analysis for the transfer function estimation and application to flutter clearance. AIAA J. Guid. Control Dyn. 21(3), 375–382 (1998) V. Gavrilets, Autonomous aerobatic maneuvering of miniature helicopters. Ph.D. dissertation, Massachusetts Institute of Technology, 2003 V. Gavrilets, M. Martinos, B. Mettler, E. Feron, Control logic for automated aerobatic flight of miniature helicopter, in Proceedings of the AIAA Guidance, Navigation, and Control Conference, Monterey, CA, Aug 2002 F.D. Harris, Articulated rotor blade flapping motion at low advance ratio. J. Am. Helicopter Soc. 17, 41–48 (1972) C. Harris (ed.), Shock and Vibration Handbook (McGraw-Hill, New York, 1996) A.M. Kuethe, C.Y. Chow, Foundations of Aerodynamics (Wiley, New York, 1986) M. LaCivita, T. Kanade, G. Papageorgiu, W. Messner, Design and flight testing of a highbandwidth h-infinity loop shaping controller for a robotic helicopter, in Proceedings of the AIAA Guidance, Navigation, and Control Conference, Monterey, CA, Aug 2002a M. LaCivita, W. Messner, T. Kanade, Modeling of small-scale helicopters with integrated firstprinciples and integrated system identification techniques, in Presented at 58th Forum of American Helicopter Society, Montreal, Canada, June 2002b J.G. Leishman, Principles of Helicopter Aerodynamics (Cambridge University Press, New York, 2000) L. Ljung, System Identification: Theory for the User (Prentice Hall, Upper Saddle River, 1999)

306

V. Gavrilets

M. McConley, Draper small autonomous aerial vehicle dynamic model. Technical report E41-98091, Draper Laboratory, Aug 1998 B. Mettler, Identification, Modeling and Characteristics of Miniature Rotorcraft (Kluwer, Boston, 2002) B. Mettler, M. Tischler, T. Kanade, W. Messner, Attitude control optimization for a small-scale unmanned helicopter, in AIAA Guidance, Navigation and Control Conference, Denver, CO, Aug 2000 B. Mettler, V. Gavrilets, E. Feron, T. Kanade, Dynamic compensation for high-bandwidth control of small-scale helicopter, in American Helicopter Society Specialist Meeting, San Francisco, CA, Jan 2002a B. Mettler, M.B. Tischler, T. Kanade, System identification modeling of a small-scale unmanned rotorcraft for control design. J. Am. Helicopter Soc. 47(1), 50–63 (2002b) Miniature Aircraft USA, X-Cell .60 graphite SE Helicopter Kit (Special Edition) Instruction Manual (Miniature Aircraft USA, Orlando, 1999) NASA Ames Research Center, Comprehensive Identification from Frequency Responses: An Interactive Facility for System Identification and Verification (NASA Ames Research Center, Moffet Field, 2000) G.D. Padfield, Helicopter Flight Dynamics: The Theory and Application of Flying Qualities and Simulation Modeling. AIAA Education Series (AIAA, Reston, 1996) J.M. Rolfe, K.J. Staples, Flight Simulation. (Cambridge University Press, Cambridge, 1986) K. Sprague, V. Gavrilets, D. Dugail, B. Mettler, E. Feron, Design and applications of an avionics system for a miniature acrobatic helicopter, in AIAA Digital Avionics Systems Conference, Daytona Beach, FL, 2001 T.D. Talbot, B.E. Tingling, W.A. Decker, R.T. Chen, A mathematical model of a single main rotor helicopter for piloted simulation. Technical Memorandum 84281, NASA, 1982 M. Tischler (ed.), Advances in Aircraft Flight Control (Taylor and Francis, Cornwall, 1996) W. von Grunhagen, G. Bouwer, H.-J. Pausder, F. Henchel, J. Kaletka, A high bandwidth control system for the helicopter in-flight simulator atthes – modelling, performance and applications, in Advances in Aircraft Flight Control, ed. by M. Tischler (Taylor and Francis, Cornwall, 1996)

Quadrotor Kinematics and Dynamics

16

Caitlin Powers, Daniel Mellinger, and Vijay Kumar

Contents 16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.1 Quadrotor Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.2 Coordinate System and Reference Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.3 Inertial Properties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.4 Motor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.5 Rigid Body Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2.6 Differential Flatness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.3 Robot Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.3.1 Linearized Controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.3.2 Nonlinear Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.4.1 Vicon Motion Capture System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.4.2 Software and Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

308 309 309 310 311 311 313 314 317 317 322 323 324 324 325 327 327

Abstract

This chapter presents an overview of the rigid body dynamics of a quadrotor as well as several controllers for the quadrotor. First, the Newton-Euler equations of motion that govern the quadrotor motion are described, and it is shown that the quadrotor model is differentially flat. Next, two controllers for the quadrotor C. Powers () • V. Kumar Department of Mechanical Engineering and Applied Mechanics, University of Pennsylvania, Philadelphia, PA, USA e-mail: [email protected]; [email protected] D. Mellinger KMel Robotics, Philadelphia, PA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 71, © Springer Science+Business Media Dordrecht 2015

307

308

C. Powers et al.

are presented. The first is a linear controller based on a linearized model of the dynamics. The second is a nonlinear controller derived from the original dynamic model. The architecture of the GRASP quadrotor testbed at the University of Pennsylvania is also presented. Finally, experimental results which illustrate the dynamics and control of small quadrotors are presented.

16.1

Introduction

Aerial robotics is a growing field with tremendous civil and military applications. Potential applications for micro unmanned aerial vehicles include search and rescue tasks, environmental monitoring, aerial transportation and manipulation, and surveillance. Quadrotor designs, rotorcrafts whose propulsive force is provided by four rotors, make flexible and adaptable platforms for aerial robotics. Rotorcraft designs are not limited to quadrotors – vehicle designs range from helicopters with tail rotors as seen in vehicles used for transport, coaxial helicopter designs such as the Skybotix CoaX (Skybotix Technologies 2012) to the Falcon 8, a vehicle with 8 rotors sold by Ascending Technologies, GmbH (Ascending Technologies 2012). However, quadrotors have recently emerged as the platform of choice for research in micro aerial vehicles (Kumar and Michael 2011). They are highly maneuverable and enable safe and low-cost experimentation in mapping, navigation, and control algorithms in three dimensions. They can hover in place and take off and land vertically, unlike their fixed-wing counterparts. They also have sufficient payload and flight endurance to support a number of indoor and outdoor applications. Their ability to move in 3-dimensional space brings new research challenges compared to the wheeled mobile robots that have driven mobile robotics research over the last decade. In recent years, small quadrotors have been demonstrated for mapping three-dimensional environments (Shen et al. 2011), transporting and manipulating payloads (Lindsey et al. 2011), assembling structures (Michael et al. 2011), and for autonomous exploration (Shen et al. 2012). The work in this chapter deals specifically with quadrotors that are on the order of 0.1–0.5 m in length and 0.1–0.5 kg in mass. As the length is scaled down, mass and inertia scale down much faster resulting in higher angular accelerations and superior maneuverability. As shown in Kushleyev et al. (2012), very small quadrotors (0.1 m in diameter) are remarkably agile. Another advantage is that the quadrotor design is easy to build and is quite robust. The blades have fixed pitch and the propellers rotate in one direction, allowing for simple motors and controllers. There are no hinges or flaps. As a result quadrotors are also easy to model and control. However, quadrotors are inefficient compared to fixed-wing UAVs. While fixedwing UAVs are able to exploit the lift generated by the structure of their wings, rotorcraft such as quadrotors must produce all of the energy they need to fly via their rotors, which is a much less efficient process. Because of limits on battery energy density and specific power, the battery represents close to 30 % of total mass, yielding flight times of less than 30 min. Another key challenge to creating autonomous quadrotors is the need to equip them with onboard sensors

16 Quadrotor Kinematics and Dynamics

309

and processors to enable 3-D state estimation. Sensors such as lidars are heavy and consume a lot of power, while cameras require significant processing power. The reader is referred to Kumar and Michael (2011) for a discussion of the challenges in developing autonomous quadrotors. This chapter presents an overview of the dynamics, control, and planning for quadrotors. The next section describes a dynamic model of a quadrotor. The dynamic model is used to develop controllers for the system, which are described in Sect. 16.3. Experimental results with controllers on real platforms are discussed in Sect. 16.4. Two popular quadrotor platforms are used extensively for modeling and experimental results. The first is the Hummingbird quadrotor sold by Ascending Technologies, GmbH (Ascending Technologies 2012). The second platform is the Nano Quadrotor, a micro quadrotor designed by KMel Robotics (KMel Robotics 2011; Kushleyev et al. 2012). Both platforms will be described in detail throughout this chapter.

16.2

Modeling

16.2.1 Quadrotor Platforms This section starts by describing two representative quadrotor platforms. Both platforms have onboard microcontrollers which allow low-level control and estimation to be done in flight by the robot. In addition, both platforms have onboard IMUs which provide feedback of accelerations and angular velocities. The Hummingbird quadrotor is shown in Fig. 16.1. The Hummingbird spans 55 cm from the tip of one rotor to the tip of the opposite rotor, has a height of 8 cm, and has a mass of about 500 g including a battery. The platform is durable enough to survive most crashes, while the blades are soft enough not to cause damage during such event. Furthermore, the 20 min battery life and 200 g payload capacity are also advantageous.

Fig. 16.1 A quadrotor platform with Vicon markers

310

C. Powers et al.

The Nano Quadrotor spans 15 cm, has a height of 5 cm, and has a mass of about 75 g including battery. The platform is very small and maneuverable and safe due to the small size of the rotors.

16.2.2 Coordinate System and Reference Frames A convenient description of the kinematics of a quadrotor can be provided using four reference frames. The inertial frame, A, is defined by the triad a1 , a2 , and a3 with a3 pointing upward. Two intermediate frames E and F are produced by first rotating through an angle about the a3 axis and then rotating about an angle  about the e1 axis. Finally, the body frame B of the quadrotor is defined by rotating about an angle  about the f2 axis. The origin of the body frame is attached to the center of mass of the quadrotor with b3 perpendicular to the plane of the rotors pointing vertically up (aligned with a3 ) during perfect hover. The basis vectors of the body frame are parallel to the principal axes of inertia of the quadrotor. The center of mass is denoted as C (Fig. 16.2).

Fig. 16.2 The body-fixed frame and the inertial reference frame. A pair of motors (1 and 3) spins counterclockwise, while the other pair spins clockwise (2 and 4). The pitches on the corresponding propellers are reversed so that the thrust is always pointing in the b3 direction for all propellers. However, while the reaction moments on the frame of the robot are also in the vertical direction, the signs are such that they oppose the direction of the angular velocity of the propeller

16 Quadrotor Kinematics and Dynamics

311

16.2.3 Inertial Properties Since bi are principal axes, the inertia matrix referenced to the center of mass along the bi reference triad, I , is a diagonal matrix. In practice, the three moments of inertia can be estimated by weighting individual components of the quadrotor and building a physically accurate model in SolidWorks or approximated using known inertia tensors of simple shapes (e.g., cylinders for the motors). The key parameters for the rigid body dynamics of the quadrotors used for simulation and experiments in this text are as follows: (a) Total mass of the quadrotor m (b) The distance from the center of mass to the axis of a motor: L (c) The components of the inertia dyadic using bi as the basis vectors: 2

3 Ixx 0 0 ŒIC bi D 4 0 Iyy 0 5 : 0 0 Izz The inertial properties of each of the quadrotor platforms are detailed in Table 16.1. Inspection of this table combined with reference to Kumar and Michael (2011) allows the reader to gain some insight into quadrotor scaling. Let L represent the characteristic length. The rotor radius R should scale linearly with L, the mass should as L3 , and the moments of inertia should scale as L5 . The force produced by the rotors, F , and the drag force D scale with the cross-sectional area and the square of the blade tip speed, v. Let the angular speed of the blades be denoted !. Clearly, ! D Lv . Thus F  ! 2 L4 and D  ! 2 L4 . Linear acceleration a D Fm scales as a  ! 2 L.

16.2.4 Motor Model Each rotor has an angular speed !i and produces a vertical force Fi according to Fi D kF !i2 : Table 16.1 Inertial properties of quadrotor platforms

Platform kg Ixx , 2 m kg Iyy , 2 m kg Izz , 2 m m, kg L, m

(16.1)

Hummingbird

Nano

3

2:32  10

4:28  105

2:32  103

4:28  105

4:00  103

8:36  105

0.5 0.175

0.075 0.0635

312

C. Powers et al.

The rotors also produce a moment according to Mi D kM !i2 :

(16.2)

The constants kM and kF can be determined by experimentation with fixed rotors or by matching the performance of a simulation to the performance of the real system. Data obtained from system identification experiments suggests that the rotor speed is related to the commanded speed by a first-order differential equation !Pi D km .!ides  !i /: This motor gain, km , is found from experimental data. The desired angular velocities, !ides , are limited to a minimum and maximum value determined through experimentation. The response of a motor from the Nano Quadrotor to a step input is shown in Fig. 16.3. However, as a first approximation, the motor controllers can be assumed to be perfect and the time constant km associated with the motor response to be arbitrarily small. In other words, assume that the actual motor velocities !i are equal to the commanded motor velocities, !ides . Specific values for the Hummingbird and Nano Quadrotor platforms are given in Table 16.2.

Fig. 16.3 Experimental step response for motor on Nano Quadrotor. The commanded speed is a step from 0 to 10,000 RPM at time 0. Dots represent measured motor speeds. The motor reaches the desired speed in approximately 0:25 s

16 Quadrotor Kinematics and Dynamics Table 16.2 Motor characteristics of quadrotor platforms

313

Platform Nm kM , rpm2 N kF , rpm2

Hummingbird 9

km , s1

Nano Quadrotor

1:5  10

4:737  1012

6:11  108

1:9985  109

20

200

16.2.5 Rigid Body Dynamics Kinematics The Z  X  Y Euler angle convention is used to model the rotation of the quadrotor in the world frame. As described in Sect. 16.2.2 to get from A to B, first rotate about a3 through the yaw angle, , to get the triad ei . A rotation about the e1 through the roll angle,  results in the triad fi (not shown in the figure). A third pitch rotation about f2 through  results in the body-fixed triad bi . This convention produces the rotation matrix from A to B, A RB : 2

c c  ss s A ŒRB D 4 cs C c ss cs

cs cc s

3 c s C css s s  c cs 5 : cc

(16.3)

The components of angular velocity of the robot in the body frame are given by p, q, and r: A !B D pb1 C qb2 C rb3 : These values are related to the derivatives of the roll, pitch, and yaw angles according to 32 3 2 3 2 c 0 cs p P 5 4 4q 5 D 4 0 1 (16.4) s P 5 ; P s 0 cc r Newton’s Equations of Motion Let r denote the position vector of C in A. The forces on the system are gravity, in the a3 direction, and the forces from each of the rotors, Fi , in the b3 direction. The equations governing the acceleration of the center of mass are 3 2 3 0 0 5: mRr D 4 0 5 CA RB 4 0 mg F1 C F2 C F3 C F4 2

The first input u1 will be defined to be u1 D †4iD1 Fi ; which is the total force produced by the rotors in the b3 direction.

(16.5)

314

C. Powers et al.

Euler’s Equations of Motion In addition to forces, each rotor produces a moment perpendicular to the plane of rotation of the blade, Mi . Rotors 1 and 3 rotate in the b3 direction, while 2 and 4 rotate in the Cb3 direction. Since the moment produced on the quadrotor is opposite to the direction of rotation of the blades, M1 and M3 act in the b3 direction while, M2 and M4 act in the b3 direction. The angular acceleration determined by the Euler equations is 2 3 2 2 3 3 2 3 L.F2  F4 / pP p p 4 5 4 5 I 4 qP 5 D 4  I  q5 : q L.F3  F1 / rP r r M1  M2 C M3  M4

(16.6)

This can be rewritten as 2 3 2 3 2 2 3 2 3 3 F1 pP p 0 L 0 L 6 7 p F2 7 4 5 4q 5 : I 4 qP 5 D 4L 0 L 0 5 6   I q 4F3 5 rP r     r F4

(16.7)

where  D kkMF is the relationship between lift and drag given by Eqs. (16.1) and (16.2). Accordingly, the second set of inputs is defined as the vector of moments u2 given by 2 3 3 F1 2 0 L 0 L 6 7 F2 7 u2 D 4L 0 L 0 5 6 4F3 5 :     F4 Since u1 and u2 are defined using a linear combination of the forces produced by the motors, they become the inputs to the system. Note that a rigid body in cartesian coordinates has six degrees of freedom, so the four inputs result in an underactuated system.

16.2.6 Differential Flatness The quadrotor has six degrees of freedom in position and orientation but only four actuators. This means that it is an underactuated system. In order to plan dynamically feasible trajectories and feedforward control inputs, it must be shown that the quadrotor system is differentially flat. A system is differentially flat if the states and inputs can be written as algebraic functions of the flat outputs  and their derivatives (van Nieuwstadt et al. 1994). This means that any trajectory .t/ in the space of flat outputs will be dynamically feasible for the quadrotor. In the case of the quadrotor, the four flat outputs can be shown to be  D Œx; y; z; . The quadrotor is differentially flat if it can be shown that there exists a smooth map .x; u/ D ˆ.; P ; R /; (16.8)

16 Quadrotor Kinematics and Dynamics

315

  where x is the state of the quadrotor, defined as x D x y z   xP yP zP p q r , and u are the control inputs as defined in Sect. 16.2.5. From the above definition of , it is trivial to show that this condition is satisfied for position and velocity, with the acceleration proving that the map is smooth: Œx; y; z D Œ1 ; 2 ; 3  Œx; P y; P zP D ŒP 1 ; P 2 ; P 3  Œx; R y; R zR D ŒR 1 ; R 2 ; R 3  :

(16.9)

Next, the condition must be satisfied for rotation and angular velocity as well. Let the rotation matrix A RB describe the rotation of the quadrotor, with the unit vectors in the body frame denoted by b1 ; b2 ; b3 . From the equations of motion, b3 D

1 r u1 .R j u11 .Rr

 mg/  mg/j

D

.Rr  mg/ j.Rr  mg/j

(16.10)

where the dependence on u1 can be eliminated since it is a scalar and b3 is a unit vector. It has been shown above that rR is a function of the flat outputs alone and the remaining terms are constant, so b3 is also a function of the flat outputs. A second unit vector in the body frame can be obtained using the following: e1 D cos a1 C sin a2 0

b3  e1 D b3  R b1 D b3  .cos b1  sin b3 / D cos b2

(16.11) (16.12)

where RT is a rotation of  about the b2 axis. Therefore the result of b3  e1 can be normalized to find b2 as a function of the flat outputs, provided that cos  is not zero. Finally, the cross product of b3 and b2 can be used to find b1 . To show that the angular velocity is also a function of the flat outputs, use the derivative of acceleration given by the Newton-Euler equations in (16.5): jD

1 1 uP 1 b3 C u1 A ! B  b3 : m m

(16.13)

Define j D « r to denote the jerk. Taking the dot product of the above equation with b3 shows that uP 1 D mj  b3 : (16.14) This can be substituted into (16.13) to find A

! B  b3 D

m .j  .j  b3 /b3 /: u1

(16.15)

Expressing A ! B in terms of components p; q, and r means that the left-hand side of the equations above can be equated to qb1  pb2 to find

316

C. Powers et al.

p D b2 

m m .j  .j  b3 /b3 /; q D b1  .j  .j  b3 /b3 /: u1 u1

(16.16)

For the third component of angular velocity, the derivatives of the Euler angles are related to the angular velocity using the following, 3 2 3 P p   7 6 A B ! D e1 b2 a3 4 P 5 D A RB 4 q 5 : r P 2

(16.17)

P which can P / This results in a system of three equations and three unknowns .r; ; be solved provided that e1 , b2 , and a3 are linearly independent. The mapping between the flat outputs and the derivatives of angular velocity is found by taking another derivative of Newton’s equation, giving mPj D uR 1 b3 C 2A ! B  uP 1 b3 C A ! B  A ! B  u1 b3 C A ˛ B  u1 b3 :

(16.18)

uR 1 can be found by projecting this equation along b3 . A ! B , u1 and uP 1 can be used to compute A ˛ B  b3 . This produces pP D .A ˛ B  b3 /  b2 ; qP D .A ˛ B  b3 /  b1 :

(16.19)

For the last component of angular acceleration, the derivative of the relationship between Euler angles and angular velocity (16.17) can be used: 02 3 2 3 2 31 2 3 pP p p R   6 B 6 7 6 7 C 7 6 A P 2 C e1 b2 a3 4 R 7 P 1 C B ! A  b RB @4 qP 5 C 4 q 5  4 q 5A D E ! A  e 5: R rP r r (16.20) If e1 , b2 , and a3 are linearly independent, then this equation is affine in the variables p; P qP and rP and results in a system of equations which can be solved for r. P Finally the control inputs need to be expressed as a function of the flat outputs. This can be done by simply rearranging the Newton-Euler equations to solve for u1 and u2 :   u1 D mjj R 1 R 2 R 3 C g jj

(16.21)

2 3 2 3 2 3 p p pP u2 D I 4 qP 5 C 4 q 5  I 4 q 5 : r r rP

(16.22)

16 Quadrotor Kinematics and Dynamics

317

Thus, the existence of the smooth map is shown in (16.9), (16.16), (16.17), and (16.19)–(16.22). However, as presented, the derivation is only valid provided that zB is never parallel to xC and yB is never parallel to zW .

16.3

Robot Controllers

16.3.1 Linearized Controller For control at small angles, the position and attitude control of the quadrotor can be decoupled. This allows the use of nested feedback loops which allows the attitude control to run at a much faster rate than the position control. The control problem is to determine the four inputs fu1 ; u2 g required to follow a desired trajectory zdes . As shown in Fig. 16.4, errors in the robot’s position will drive a position controller which directly determines u1 from (16.26). The model in (16.26) also leads to the desired orientation along the trajectory. The attitude controller for this orientation is derived from the model in (16.7). The transformation of the forces F1 to F4 into fu1 ; u2 g is straightforward and is described in Michael et al. (2010). The inner attitude control loop uses onboard accelerometers and gyroscopes to control the roll, pitch, and yaw and runs at approximately 1 kHz (Gurdan et al. 2007), while the outer position control loop uses estimates of position and velocity of the center of mass to control the trajectory in three dimensions at 100–200 Hz.

Fig. 16.4 The position and attitude control loops

318

C. Powers et al.

16.3.1.1 Linearization The controllers are derived by linearizing the equations of motion and motor model Eqs. (16.2)–(16.5) at an operating point that corresponds to the nominal hover state, x0 , with the inputs nominally required for hover, u0 . The hover configuration is given by r D r0 ,  D  D 0, D 0 , rP D 0, and P D P D P D 0. This approximation is valid where the roll and pitch angles are small (c  1, c  1, s  , and s  ). Linearizing the equations of motion about the hover state results in an equation of the form xP D

@ @ f .x; u/j.x0 ;u0 / .x  x0 / C f .x; u/j.x0 ;u0 / .u  u0 / D Ax C Bu: @x @u

At this state the lift from the propellers is given by Fi;0 D

mg ; 4

The nominal values for the inputs at hover are u1;0 D mg, u2;0 D 0. A and B are given by 2

033 6 036 AD6 4 033 033 2 g cos 0 a D 4 g sin 0 0 2 61 6m 6 6 B D60 6 6 40 0

I33 a33 036 036

3 036 033 7 7 I33 5 033

g sin g cos 0

045 0 0 043 1 0 Ixx 1 0 Iyy 0

0

0 0

(16.23)

3 0 05 0 3

0 7 7 7 7 : 0 7 7 7 0 5

(16.24)

(16.25)

1 Izz

Specifically, the linearized equations for the acceleration are rR1 D g. cos

0

C  sin

0/

rR2 D g. sin

0

  cos

0/

rR3 D u1 D u1  u1;0  D   0  D   0 :

(16.26)

16 Quadrotor Kinematics and Dynamics

319

Assuming that the rotor craft is symmetric so Ixx D Iyy , pP D

u2;x L D .F2  F4 / Ixx Ixx

u2;y L D .F3  F1 /; Iyy Iyy u2;z  rP D D .F1  F2 C F3  F4 /: Izz Izz

qP D

In other words, the equations of motion for the angular accelerations are decoupled. Each component of angular acceleration depends only on the appropriate component of u2 .

16.3.1.2 Attitude Control One attitude controller is a proportional plus derivative (PD) attitude controller which tracks a trajectory in SO.3/ specified in terms of a desired roll, pitch, and yaw angle. Since the development of the controller will be based on linearized equations of motion, the attitude must be close to the nominal hover state where the roll and pitch angles are small. Near the nominal hover state the proportional plus derivative control laws take the form: 2 3 kp; . des  / C kd; .p des  p/ u2 D 4 kp; . des  / C kd; .q des  q/ 5 : (16.27) des des  / C kd; .r  r/ kp; . 16.3.1.3 Position Control In the next subsection, the dynamic model will be used to derive a position control method that uses the roll and pitch angles as inputs to drive the position of the quadrotor. The position control algorithm will determine the desired roll and pitch angles,  des and  des , which can be used to compute the desired speeds from (16.27). The controller tracks a specified trajectory, rT .t/, in three dimensions. The yaw angle, , can be specified independently. It can either be constant, 0 , or time varying, T .t/. Since the model is linearized, differential flatness is not necessary for trajectory planning here. Let rT denote the closest point on the desired trajectory to the current position r, and let the desired velocity and acceleration obtained by differentiating the specified trajectory be given by rP T and rR T , respectively. The desired trajectory,   rT .t/ des z D ; T .t/ will be provided by a trajectory planner as an input to specify the trajectory of the position vector and the yaw angle that the vehicle should track. These are the flat outputs which were discussed in Sect. 16.2.6. Any continuous trajectory in the flat outputs specifies the trajectory for the system and the required inputs.

320

C. Powers et al.

Thus, appropriate trajectories can be planned in this space. In particular, for hovering, rT .t/ D r0 and T .t/ D 0 . The command accelerations, rRides , are calculated from a proportional plus derivative (PD) controller. Define the position error in terms of components using the standard reference triad ai by ep D .rT  r/: Similarly, define the velocity error by ev D rPT  rP : In order to guarantee that this error goes exponentially to zero, it is required that .Rri;T  rRides / C kd;i .Pri;T  rPi / C kp;i .ri;T  ri / D 0;

(16.28)

where rPi;T D rRi;T D 0 for hover. Equation (16.26) leads to the relationship between the desired accelerations and roll and pitch angles. Given that  D   0 D  and  D   0 D , rR1des D g. des cos

T

C  des sin

T/

(16.29a)

rR2des D g. des sin

T

  des cos

T/

(16.29b)

rR3des D

1 u1  g: m

(16.29c)

For hover, (16.29c) yields   u1 D mg C mrR3des D mg  m kd;3 rP3 C kp;3 .r3  r3;0 / :

(16.30)

Equations (16.29a) and (16.29b) can be used to compute the desired roll and pitch angles for the attitude controller:  des D

1 des .Rr sin g 1

T

 rR2des cos

T/

(16.31a)

 des D

1 des .Rr cos g 1

T

C rR2des sin

T /:

(16.31b)

The desired roll and pitch velocities are taken to be zero. p des D 0

(16.32a)

D 0:

(16.32b)

q

des

16 Quadrotor Kinematics and Dynamics T .t/;

Since the yaw, controller uses

321

is prescribed independently by the trajectory generator, the des

D

T .t/

r des D P T .t/:

(16.33a) (16.33b)

These equations provide the setpoints for the attitude controller in (16.27). Note that the attitude controller must run an order of magnitude faster than the position control loop in order for this to work. In practice as discussed in Michael et al. (2010), the position controller runs at 100 Hz, while the inner attitude control loop runs at 1 kHz. An alternate error metric can be used for more aggressive trajectories which only does not consider error in the tangent direction. Let the unit tangent vector of the trajectory (unit vector along rP T ) be Ot. The unit normal to the trajectory, Ot, is derived by differentiating the tangent vector with respect to time or arc length, and finally, O The position and velocity the unit binormal vector is the cross product bO D Ot  n. errors are defined according to the following equations: O bO O nO C ..rT  r/  b/ ep D ..rT  r/  n/ and ev D rP T  rP : Note that here the position error in the tangent direction is ignored, only considering position error in the plane that is normal to the curve at the closest point. Once again the commanded acceleration, rRides , is calculated from the PD feedback as shown in (16.28). In vector notation, this is .RrT  rR des / C kd ev C kp ep D 0:

(16.34)

Finally, (16.31)–(16.33) is used to compute the desired roll and pitch angles. Again the reader is referred to Michael et al. (2010) for more details including experimental results obtained from these controllers.

16.3.1.4 Linear Quadratic Regulators The gains used in a controller have a critical effect on its performance and stability. Since the model is linearized, the controller can be framed as a linear quadratic regulator. This allows the problem of choosing gains to be formatted as an optimization problem. A quadratic functional can be used for the cost: Z J D

0

1

.x  x0 /T Q.x  x0 / C .u  u0 /T R.u  u0 /dt

(16.35)

which is subject to the dynamic constraints that xP D Ax C Bu. Q and R are chosen to weight the position errors and control efforts depending on the task at hand.

322

C. Powers et al.

The analytic solution can be found using calculus of variations, which leads to a controller of the form .u  u0 / D K.x  x0/: (16.36) K can be solved for using the Riccati equation, which can be found in many texts on optimal control, such as Kirk (2004).

16.3.2 Nonlinear Control In order to control the quadrotor along trajectories that require large rotations, a different control law is needed. A nonlinear controller in the   can track trajectories differentially flat output space defined by .t/ D rdes .t/T ; des .t/T . The error metrics on position and velocity are defined as ep D r  rdes ; ev D rP  rP des :

(16.37)

The vector describing the total force to apply to the quadrotor is computed using Fdes D Kp ep  Kv ev  mga3 C mRrdes :

(16.38)

With Kp and Kv being positive definite matrices of gains. The first input is chosen to be the desired force vector projected onto the body frame z axis of the quadrotor as such: u1 D Fdes  b3 : (16.39) For the attitude control, the z axis of the quadrotor body frame should align with the desired force vector, as the quadrotor can only produce force along this axis. Using this fact and the specified differentially flat outputs, the desired rotation matrix Rdes can be defined. Next, the error metric for rotation is defined as eR D

1 T .R R  RT Rdes /_ 2 des

(16.40)

where _ represents the “unhat” operator which transforms a skew-symmetric matrix into a vector. The vector is arranged such that Ab D A_ b for any skew-symmetric matrix A and vector b. The derivative error metric for attitude control will be B : e! D A ! B  A !des

(16.41)

A

B !des can be found from the differential flatness derivation. So the attitude controller will simply be a PD controller using these two error metrics with some gains:

O T Rdes !des  RT Rdes !P des /: u2 D I.KR eR  K! e! / C A ! B  I A ! B  I.!R (16.42)

16 Quadrotor Kinematics and Dynamics

323

An advantage to using this controller design is that under a few assumptions, statements can be made about the stability of the controller. According to Lee et al. (2010), if actuator saturation is ignored and KR and K! are chosen to be positive, the attitude error dynamics are exponentially stable provided the initial configuration of the quadrotor satisfies the following: T .0/R.0// < 2 t r.I  Rdes

jje! .0/jj2
0, the term .! C / is increased, resulting in a larger frequency and fewer lines in the strobed animation. The bias simply shifts all points in the wingbeat profile fore or aft of the nominal neutral position of the wing spar, as can be seen with the right wing in the first subplot of the second row of Fig. 17.1. The remaining subplots show the effects of upstroke and downstroke angle-of-attack and stroke plane variation. The subsequent information presented allows for the independent actuation of wings. In other words, the right and left wings are taken to be driven by independent motion sources; however, the results are also applicable to designs where the left and right wings are coupled. Figure 17.2 graphically illustrates the effects of differences in asymmetric frequency, bias, and amplitude. The top two plots show the effect of asymmetric frequency for both positive and negative values of ı. The bottom figures shows the wing bias, , and the amplitude adjustment, A. The wingbeat cycle is split into two parts, the first part is the upstroke, while the second part is the downstroke. Considering Fig. 17.2, the upstroke is defined as moving from a wing position of 1 rad to a wing position of 1 rad, while the downstroke moves from a wing position of 1 rad to a wing position of 1 rad. A positive asymmetric frequency, ı, impedes or slows down the upstroke while the downstroke advances or speeds up, as shown in the upper left plot in Fig. 17.2. A negative asymmetric frequency has the opposite effect in that the upstroke is advanced, while the downstroke is impeded. The wing bias shifts the midstroke wingbeat angular position fore or aft by the bias amount, while the amplitude adjusts the limits of wing stroke. Both the bias and amplitude variation, as presented in Fig. 17.2, yield discontinuous wingbeat positions. The physical mechanism used to produce the wingbeat motion would not allow such discontinuities and would smooth the discontinuous functions but would result in an actual wingbeat motion that is different from that which was commanded. To avoid this issue, techniques exist that modify the

17 Dynamics and Control of Flapping Wing MAVs

333

Fig. 17.1 Strobed illustration of effect of variations in wing kinematics and definition of parameters

wingbeat forcing functions to yield continuous functions (Doman et al. 2010b; Oppenheimer et al. 2011a) even when the bias and/or amplitude change from one wingbeat cycle to another. The aerodynamic model employed to generate the results in this work is applicable to vehicle designs that employ passive angle-of-attack regulation in the form of limiters (Wood 2008; Doman and Regisford 2010). The subsequent results assume that the wing angle-of-attack changes instantaneously at the end of an upstroke or downstroke and remains on a limit, ˛u or ˛d , over the course of an upstroke or downstroke, respectively. Finally, the model allows for the variation in stroke plane inclination, , which allows the resultant aerodynamic force vector to be inclined with respect to the MAV fuselage.

334

D.B. Doman et al. Wing Position vs. Time:δ > 0

Wing Position vs. Time:δ < 0 1

4

0.5 0

0 −0.5

−0.5 −1

4

0.5 φ(t) (rad)

φ(t) (rad)

1

0

−1

1

0.5

0

Normalized Time Wing Position vs. Time: Bias 2

1

1

0.6

φ(t) (rad)

φ(t) (rad)

1

Normalized Time Wing Position vs. Time: Amplitude

2

0 −1

0 −1

0.4

0.2 −2

0.5

0

1

2

Normalized Time

3

−2

0.5 0

0.75 1

2

3

Normalized Time

Fig. 17.2 Examples of split cycle, bias, and amplitude wingbeat variations

17.3

Instantaneous Aerodynamic Forces and Moments

The seven parameters used to describe the motion of the wings can be used to approximate a wide range of wing motions; however, in practice, the actual motion is produced by drive elements and mechanisms and will therefore vary from the idealized model of wingbeat motion presented in Fig. 17.2. Nonlinear linkagebased mechanisms, load interactions with drive elements, and wing flexibility are just a few phenomena that can cause such departures from the idealized model of wing motion. The equations below provide a blade element-based estimate of the instantaneous forces and moments that act on a flapping wing vehicle. The equations are useful for simulations that are used to test the performance and robustness of model-based control laws designed using cycle-averaged models of such aircraft. They are also useful for conceptual design studies and computing quasi-equilibrium trim conditions for such vehicles. Note that the angular displacement of the wing in the stroke plane, .t/; the wing angle-of-attack, ˛.t/; and the stroke plane

17 Dynamics and Control of Flapping Wing MAVs

335

Table 17.1 Aerodynamic forces expressed in body frame Phase Body frame force vector 2 3 RW upstroke L.t / cos .t / C D.t / cos .t / sin .t / B 4 5 FRWU .t / D D.t / sin .t / L.t / sin .t / C D.t / cos .t / cos .t / 3 2 RW downstroke L.t / cos .t /  D.t / cos .t / sin .t / B 5 .t / D 4 D.t / sin .t / F RWD

LW upstroke

LW downstroke

L.t / sin .t /  D.t / cos .t / cos .t / 3 L.t / cos .t / C D.t / cos .t / sin .t / 5 FBLWU .t / D 4 D.t / sin .t / L.t / sin .t / C D.t / cos .t / cos .t / 2 3 L.t / cos .t /  D.t / cos .t / sin .t / B 5 FLWD .t / D 4 D.t / sin .t / L.t / sin .t /  D.t / cos .t / cos .t / 2

inclination, .t/, are written as general functions of time in the instantaneous model, in order to apply to a wide class of vehicles. Table 17.1 provides expressions for the instantaneous aerodynamic force vectors. B B FRWU and FRWD denote forces associated with the right wing on the upstroke B and downstroke written in the body-axes coordinate frame, respectively. FLWU and B FLWD denote the same for the left wing. The lift, L.t/, and drag, D.t/, are expressed in wing-carried frames and are given by P 2 L.t/ D kL .t/

(17.5)

P 2 D.t/ D kD .t/

(17.6)

where,

IA CL .˛/ 2

kD D IA CD .˛/ 2 kL D

(17.7) (17.8)

and IA is the wing planform area moment of inertia about the spar root hinge point that defines the stroke plane. The lift and drag coefficients are given by the empirical formulae (Sane and Dickinson 2001) CL .˛/ D 0:225 C 1:58 sin.2:13˛  7:2/

(17.9)

CD .˛/ D 1:92  1:55 cos.2:04˛  9:82/

(17.10)

where ˛ is expressed in degrees. The center-of-pressure of each planform must be obtained in order to compute moments on the fuselage. The center-of-pressure location for each wing can be computed as

336

D.B. Doman et al.

RR

2 1 ywp c.ywp /2 dywp xcp D 0R R 2 2 0 ywp c.ywp /dywp RR 3 ywp c.ywp /dywp ycp D R0R 2 0 ywp c.ywp /dywp

(17.11)

(17.12)

where the subscript wp indicates a coordinate axis within a planform frame and c.ywp / is the wing chord as a function of span. Let rBR and rBL denote the position vector from the vehicle center-of-gravity to the right and left wing root hinge points, i.e., T M  rBR D x w=2 z

(17.13)

T M  rBL D x w=2 z

(17.14)

where x; w; z are defined in Fig. 17.1. It is assumed that the vehicle centerof-gravity is fixed due to the wing mass being much lower than the mass of the fuselage and drive mechanisms. The centers of pressure associated with each wing and stroke, expressed in the body frame, are summarized in Table 17.2. The expressions for the aerodynamic moments associated with each wing and stroke are given by MBRWU D rcp BRWU  FBRWU MBRWD D rcp BRWD  FBRWD MBLWU D rcp BLWU  FBLWU MBLWD D rcp BLWD  FBLWD

(17.15)

Table 17.2 Centers of pressure for each wing expressed in the body frame CP location RW upstroke

RW downstroke LW upstroke

LW downstroke

Body frame expression 2 3 xcp sin ˛.t/ cos .t/ C sin .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C x 6 7 rcp BRWU .t/ D 4 xcp sin .t/ cos ˛.t/ C ycp cos .t/ C w2 5 xcp sin ˛.t/ sin .t/ C cos .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C z 2 3 xcp sin ˛.t/ cos .t/ C sin .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C x 6 7 rcp BRWD .t/ D 4 xcp sin .t/ cos ˛.t/ C ycp cos .t/ C w2 5 xcp sin ˛.t/ sin .t/ C cos .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C z 2 3 xcp sin ˛.t/ cos .t/ C sin .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C x 6 7 B w rcp LWU .t/ D 4 xcp sin .t/ cos ˛.t/  ycp cos .t/  2 5 xcp sin ˛.t/ sin .t/ C cos .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C z 2 3 xcp sin ˛.t/ cos .t/ C sin .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C x 6 7 rcp BLWD .t/ D 4 xcp sin .t/ cos ˛.t/  ycp cos .t/  w2 5 xcp sin ˛.t/ sin .t/ C cos .t/.xcp cos .t/ cos ˛.t/ C ycp sin .t// C z

17 Dynamics and Control of Flapping Wing MAVs

17.4

337

Quasi-equilibrium Hover and Vehicle Design

For insight into how to design a tailless flapping wing aircraft that can be trimmed, it is useful to consider the expression for a cycle-averaged generalized force: ! GD 2

Z

2 !

0

G..t//dt

(17.16)

The necessary conditions for a flapping wing aircraft to achieve a quasiequilibrium hover condition are that the cycle-averaged x-body force must be equal to the weight, W , of the aircraft and the cycle-averaged y- and z-body forces and body moment vector must be equal to zero, namely, M B B XN B D XN LW C XN RW DW

(17.17)

M B B YN B D YNLW C YNRW D0

(17.18)

M B B ZN B D ZN LW C ZN RW D0

(17.19)

M MN xB D MN xBLW C MN xBRW D 0

(17.20)

M MN yB D MN yBLW C MN yBRW D 0

(17.21)

M MN zB D MN zBLW C MN zBRW D 0

(17.22)

These conditions are necessary for a flapping wing aircraft to orbit a fixed point in space. These conditions are not sufficient for conventional steady equilibrium because of the periodic forces and moments produced by flapping wings. The periodicity of the forces causes limit-cycle oscillations to occur about a point, thus the use of the term quasi-equilibrium hover or trim condition. Furthermore, certain initial conditions must be met in order to avoid drifting away from the quasiequilibrium point (Oppenheimer et al. 2010a). By making use of the instantaneous equations for the forces and moments from the previous section and computing cycle-averages from Eq. 17.16, one can solve for values of the seven wing motion parameters and center-of-gravity locations that satisfy the necessary conditions. In general, this calculation will have to be performed numerically. A typical method (Doman and Regisford 2010) for solving the trim problem follows: 1. Determine the stroke plane inclination, , or upstroke and downstroke angle-ofattack combination, ˛U , ˛D , that causes ZN B D 0 for a given drive mechanism that produces wing spar motion .t/. 2. Adjust the symmetric frequency ! and/or amplitude A such that XN B D W . 3. Determine the center-of-gravity location, x and z, and/or wing bias, , such that MN yB D 0. 4. If the left and right wings are synchronized, then YN B D MN xB D MN zB D 0.

338

17.5

D.B. Doman et al.

Flapping Wing Aerodynamic Control Derivatives

Cycle-averaged control derivatives can be physically interpreted as the sensitivities of the cycle-averaged aerodynamic forces and moments to changes in the parameters that characterize the flapping motion of the wings. The method by which such control derivatives can be obtained is outlined below. Let G.t/ denote a generalized periodic aerodynamic force of period 2=! that is produced by a flapping wing. The generalized forces are written in body-axis coordinates where the origin of the coordinate frame is fixed to the fuselage center-of-gravity. A cycle-averaged generalized aerodynamic force is given by ! GN D 2

"Z

 !ı

0

Z G.t/dt C

2 !  !ı

# G.t/dt

(17.23)

The cycle-averaged control derivatives for a given generalized force are given by 

@GN=@! @GN=@ı @GN=@A @GN=@ @GN=@ @GN=@˛u @GN=@˛d

ˇ ˇ

hover

(17.24)

Table 17.3 was created using the procedure described above, and it presents all nonzero control derivatives associated with the seven wingbeat kinematic parameters for the specific case where D ı D  D 0 at hover. More general expressions for the control derivatives may be found in Oppenheimer et al. (2010b, 2011b). In Table 17.3, J1 .A/ is a Bessel function of the first kind. Similarly J0 .A/ and J2 .A/ are Bessel functions of the zeroth and second kind. Finally the derivatives of kL and kD with respect to angle of attack are denoted by dL D @kL =@˛ and dD D @kD =@˛, respectively.

17.6

Influence of Control Mechanisms on Wing Motion and Cycle-Averaged Generalized Forces

Wing drive mechanisms for flapping wing aircraft must produce periodic wingbeat motion. Control mechanisms will cause variations in one or more of the seven parameters described above. A general procedure for estimating the influence of changes in the position of control mechanisms on flapping wing motion is to first compute the wing spar position, .t/, over one complete wingbeat cycle at the quasi-equilibrium hover condition and then adjust the parameters of Eqs. 17.1 and 17.2 to fit the actual wing motion profile. By adjusting the state of the control mechanism, such that one or more of the seven parameters is altered, and recomputing the fit parameters, an approximation for the partial derivatives of the seven parameters with respect to changes in the control mechanism can be estimated. A simple fitting procedure for the periodic wing root position follows: 1. Determine the period, T , of the wingbeat and compute the symmetric frequency:

17 Dynamics and Control of Flapping Wing MAVs

339

Table 17.3 Nonzero cycle-averaged control derivatives evaluated at quasi-equilibrium hover Sensitivity Left wing Right wing B

@Fx =@! B @Fx =@A B @Fx =@˛u B @Fx =@˛d

A2 !kL A! 2 kL 1=4A2 ! 2 d L 1=4A2 ! 2 d L

A2 !kL A! 2 kL 1=4A2 ! 2 d L 1=4A2 ! 2 d L

@Fz =@ı B @Fz =@˛u B @Fz =@˛d B @Fz =@

A!kD J1 .A/ 1=2A! 2 J .A/d 1 D 1=2A! 2 J1 .A/dD 1=2A2 ! 2 kL

A!kD J1 .A/ 1=2A! 2 J .A/d 1 D 1=2A! 2 J1 .A/dD 1=2A2 ! 2 kL

1=2A!k

C J1 .A/w/   Aycp C J1 .A/w   1=4A! dD Aycp C J1 .A/w   A! 2 kL J1 .A/ycp C Aw=4    A!J1 .A/ kD xRB  xcp sin.˛/  kL xcp cos.˛/

1=2A!kD .Aycp C J1 .A/w/   1=4A! 2 dD Aycp C J1 .A/w   1=4A! 2 d D Aycp C J1 .A/w   A! 2 kL J1 .A/ycp C Aw=4    A!J1 .A/ kD xRB  xcp sin.˛/  kL xcp cos.˛/

A! 2 J1 .A/ycp kL

A! 2 J1 .A/ycp kL

B

B

@Mx =@ı B @Mx =@˛u B @Mx =@˛d B @Mx =@ B

@My =@ı B

@My =@ B

@My =@˛u

B

@My =@˛d

B

@My =@ B

@Mz =@! B @Mz =@A B

@Mz =@˛u B @Mz =@˛d

D .Aycp

1=4A! 2 d

D 2

A! J1 .A/  Cxcp .dL C kD / cos.˛/  C.dD  kL / sin.˛/  A! 2 J1 .A/ xRB dD  Cxcp .dL C kD / cos.˛/  C.dD  kL / sin.˛/

A! 2 J1 .A/ xRB dD  Cxcp .dL C kD / cos.˛/  C.dD  kL / sin.˛/  A! 2 J1 .A/ xRB dD  Cxcp .dL C kD / cos.˛/  C.dD  kL / sin.˛/

1=2A2 ! 2 x B k R L

1=2A2 ! 2 x B k R L

2

xRB dD

  2A!kL J1 .A/ycp C Aw=4  A! 2 kL .J1 .A/=A  C.J0 .A/J2 .A//=2/ycp C w=2 1=2A! 2 dL .J1 .A/ycp C Aw=4/ 1=2A! 2 dL .J1 .A/ycp C Aw=4/

!D

  2A!kL J1 .A/ycp C Aw=4  A! 2 kL .J1 .A/=A  C.J0 .A/J2 .A//=2/ycp C w=2 1=2A! 2 d .J .A/y C Aw=4/ L 1 cp 1=2A! 2 d .J .A/y C Aw=4/ L 1 cp

2 T

(17.25)

2. Determine the time at which wing stroke reversal occurs, TR , and compute the asymmetric frequency:  ıD! (17.26) TR 3. Compute the stroke amplitude: AD

1 Œmax  min  2

(17.27)

340

D.B. Doman et al.

4. Compute the wing stroke bias:  D max  A

(17.28)

5. Compute the best fit stroke plane angle: O D arg min

Z

z

2 !

. .t/  z/2 dt

0

(17.29)

Note that this computation is trivial if the control mechanism simply tilts the stroke plane at a constant angle. 6. Compute the best fit upstroke angle-of-attack Z ˛O u D arg min au

 !ı

0

.˛.t/  au /2 dt

(17.30)

7. Compute the best fit downstroke angle-of-attack Z ˛O d D arg min ad

2 !  !ı

.˛.t/  ad /2 dt

(17.31)

Approximations to the partial derivatives of the seven parameters with respect to the control inputs can be obtained by finite difference methods. These partial derivatives can be used in conjunction with the cycle-averaged aerodynamic control derivatives presented in Table 17.3 to compute estimates of a control effectiveness matrix for a wide range of control inputs that might be found on flapping wing aircraft. The multivariable chain rule can be used to compute the sensitivities of a cycleaveraged generalized force to changes in a general control input u. Application of the multivariable chain rule yields @GN @GN @! @GN @ı @GN @A @GN @ @GN @ @GN @˛u @GN @˛d D C C C C C C @u @! @u @ı @u @A @u @ @u @ @u @˛u @u @˛d @u (17.32) Equation 17.32 serves as a basis for generating the linear control effectiveness matrix, B, for a suite of control effectors that change the cycle-averaged forces and moments acting on a vehicle through changes in the wingbeat kinematics. The control effectiveness matrix is 3 2 B @FNx =@u1 @FNxB=@u2 : : : @FNxB=@um 6 @FNyB=@u @FNyB=@u : : : @FNyB=@u 7 1 2 m7 6 7 6 NB NB NB M 6 @Fz =@u1 @Fz =@u2 : : : @Fz =@um 7 B D 6@MN B (17.33) 7 6 x =@u1 @MN xB=@u2 : : : @MN xB=@um 7 7 6 NB 4@My =@u1 @MN yB=@u2 : : : @MN yB=@um 5 @MN zB=@u1 @MN zB=@u2 : : : @MN zB=@um

17 Dynamics and Control of Flapping Wing MAVs

341

 Given a desired cycle-averaged force and moment vector, FNxdes FNydes FNzdes MN xdes T MN ydes MN zdes , the objective is to find the control vector, Œu1 u2 : : : um T , such that 3 FNxdes 2 3 6 FN 7 u1 6 ydes 7 6 u2 7 6 N 7 6 7 6 Fzdes 7 7 D B6 : 7 6 N 6 Mxdes 7 4 :: 5 7 6 4 MN ydes 5 um MN zdes 2

(17.34)

Control allocation methods (Oppenheimer et al. 2010c) can be used to solve Eq. 17.34. These methods blend the available effectors to provide the cycle-averaged forces and moments commanded by a standard control law. The control designer must bear in mind that the dynamics of the fuselage must evolve on a much slower timescale than the wingbeats and that one can only influence the cycle-averaged forces and moments rather than the instantaneous periodic forces and moments. Thus, special provisions must be made to ensure that the controller and control allocator do not attempt to compensate for fuselage motion errors occurring within the timescale of the wingbeat. A simple method of addressing this problem is to implement a “cycle zero-order hold” that only allows new control input commands to be passed to the control effectors once per wingbeat cycle (Doman et al. 2010a). The advantage of this method is that it allows one to use any standard method to control a flapping wing aircraft that operates under the influence of periodic rather than steady forces and moments. The steady force and moment commands generated by the control law are interpreted as cycle-averaged force and moment commands, and the control allocation algorithm effectively finds the inverse mapping between those commands and a set of control effector positions that are required to produce them.

17.7

Flapping Wing Aerodynamic Stability Derivatives

Extracting stability derivatives requires that the effects of the fuselage velocity be added to the force and moment computations. Previous works have extended the quasi-steady model presented here to include such effects without great computational cost (Sigthorsson et al. 2012a; Faruque and Humbert 2010). There, the expression for the velocity field, used to compute the dynamic pressure over the surfaces of the wings, includes the rigid body velocities. The instantaneous forces and moments for a particular vehicle configuration can be computed by gridding the surfaces of the wings and applying numerical integration over the surface. The cycle-averaged forces and moments are obtained by numerical integration with respect to time over a full wingbeat cycle. Numerical estimation of partial derivatives with respect to fuselage velocity changes over the wingbeat cycle can then be computed to obtain cycle-averaged stability derivatives, which can be used to form the elements of a system matrix (Sigthorsson et al. 2012b).

342

D.B. Doman et al.

In order to account for the vehicle velocity in the aerodynamic force expressions, the computation of the dynamic pressure on the wing and the angle-of-attack is required. These terms are functions of the total velocity of the wing with respect to quiescent air. The total velocity of the wing at a point T M  rwp D xwp ywp 0

(17.35)

on the wing planform is composed of the velocity due to the wing flapping, vf ; fuselage translational velocity, vb ; and the fuselage rotational velocity,  iT h   M vB .r/ D vBx .r/ vBy .r/ vBz .r/ D vBf .r/ C vBb C B  rB

(17.36)

Furthermore, the small angle between the local freestream velocity due to flapping and fuselage motion at a point on the wing planform is found to be Sigthorsson et al. (2012a)  q wp wp wp 2 2 ˛v .r/ D atan vz .r/; .vx .r// C .vy .r//

(17.37)

M

The local angle-of-attack, defined as ˛.r/ D j˛v .r/j, is used to compute the lift and drag coefficients. By definition, the drag is along the local freestream velocity vector and the lift is perpendicular to the drag. Define a frame V with its origin at r, the drag is along the positive x-axis, the lift lies on the z-axis, and the y-axis lies in the wing planform plane. The infinitesimal aerodynamic force is expressed as 2

3 dD 5 d FV D 4 0 sgn.˛v .r//dL

(17.38)

where the infinitesimal lift and drag forces are, respectively,

kv.r/k2 CL .˛.r//dxwp dywp 2

dD D kv.r/k2 CD .˛.r//dxwp dywp 2 dL D

(17.39) (17.40)

and the infinitesimal aerodynamic moment is d MB D rB  d FB

(17.41)

where the moment arm associated with a point on the planform, rB , can be obtained directly from Table 17.2 by substituting the subscripts cp with wp. Many flapping wing aircraft make use of passive angle-of-attack regulation, where dynamic pressure acting on the wing causes it to rotate to a wing stop, or twist due

17 Dynamics and Control of Flapping Wing MAVs

343

to the load. When flapping in still air at hover, the wing rotation is well defined as it takes place at stroke reversal. As the fuselage velocity increases, the wing rotation points become more difficult to compute. To illustrate the difficulty, consider an extreme but impractical case where the fuselage forward velocity is equal to the maximum blade tip velocity of the wing, measured relative to the fuselage. In this case, the dynamic pressure at the blade tip would be zero on the upstroke; thus, the wing would experience little to no passive rotation at midstroke. It is therefore intractable to write a general expression for the total aerodynamic force and center-of-pressure location using a fixed wing rotation stop over each upstroke and downstroke. However, at small velocities the wing rotation can be approximated as constant over the upstroke, ˛u , and downstroke, ˛d . Using this approximation and, for brevity, dropping the explicit notation of which variables depend on r and t, the infinitesimal aerodynamic force for the right wing on the upstroke and downstroke, respectively, is expressed as follows: dFB RWU D 2

kvk 2

3 cos. / sin.˛u /  cos.˛u / cos./ sin. / sin. / sin./ cos.˛u / cos. / C cos./ sin.˛u / sin. / 6 7  4 cos.˛u / sin./ cos./  sin.˛u / sin./ 5  cos.˛u / cos. / cos./  sin.˛u / sin. / cos. / sin./ cos. / cos./ sin.˛u /  cos.˛u / sin. / q 2 1 3 0 wp wp wp wp vx vz = .vx /2 C .vy /2 6 C 7 B q 6 C B wp wp 2 wp 2 7 wp (17.42)  Bvwp CD .˛/ C 6vwp y vz = .vx / C .vy / 7 sgn.vz /CL .˛/C dxwp dywp 4 A 5 @ q wp 2 wp 2 .vx / C .vy / dFB RWD D 2

kvk 2

3 cos. / sin.˛d / C cos.˛d / cos./ sin. / sin. / sin./  cos.˛d / cos. / C cos./ sin.˛d / sin. / 6 7 4  cos.˛d / sin./ cos./  sin.˛d / sin./ 5 cos.˛d / cos. / cos./  sin.˛d / sin. / cos. / sin./ cos. / cos./ sin.˛d / C cos.˛d / sin. / q 2 1 3 0 wp wp wp wp vx vz = .vx /2 C .vy /2 6 C 7 B q 6 C 7 B wp wp wp (17.43) Bvwp CD .˛/ C6vwp v = .vx /2 C .vy /2 7sgn.vwp z /CL .˛/Cdxwp dywp 4 y qz A 5 @ wp 2 wp 2 .vx / C .vy /

The force for the left wing is the same for a symmetric stroke but with a sign change on the second component of the force vector. After gridding the wing planforms, the instantaneous aerodynamic forces and moments can be computed by numerically integrating the differential forces and moments over the planform area. In order to compute the cycle-averaged forces and moments, the total instantaneous forces and moments can be numerically integrated with respect to time over the period of the wingbeat. Cycle-averaged stability derivatives can be estimated through numerical difference computations applied to the vehicle linear and angular velocities.

344

17.8

D.B. Doman et al.

Simulation Methods

Rigid body equations of motion for aircraft can be used as a point of departure for simulation analysis, provided that the mass of the wings is small when compared to the mass of the fuselage (Doman et al. 2010a). The reader is referred to the multibody equations of motion for flapping wing aircraft for cases where this condition is violated (Bolender 2009). It is important to note that the forces and moments that drive these equations are periodic functions of time. The equations of motion for a rigid body aircraft are 2 3 pP   4 qP 5 D I1 MB .t/    I rP 2 3 2 3 2 3 uP 0 qw  rv   4 vP 5 D 4 ru  pw 5 C 1 FB .t/  RBI 4 0 5 m wP g pv  qu 2 3 2 3 u xP 4 yP 5 D RIB 4 v 5 w zP 2 3 2 32 3 q0 qP0 0 p q r 6 qP1 7 6 p 0 r q 7 6 q1 7 6 7D6 76 7 4 qP2 5 4 q r 0 p 5 4 q2 5 qP3

r q p 0

(17.44)

(17.45)

(17.46)

(17.47)

q3

where I is the mass moment of inertia matrix;  D Œp q rT is the body-axis angular rate vector; MB .t/ is the sum of the left and right wing instantaneous aerodynamic body-axis moment vectors consisting of the rolling, pitching, and yawing moments; u; v; w are the body-axis translational velocities; FB .t/ is sum of the left and right wing instantaneous aerodynamic body-axis force vectors; m is the vehicle mass; g is the acceleration due to gravity; x; y; z are the vehicle positions with respect to an inertial frame; and q D q0 C q1 iO C q2 jO C q3 kO is a quaternion used to perform the T 3-2-1 standard Euler transformation. Also, RIB D RBI are rotation matrices that transform from body to inertial axes and back. A simple blade element-based simulation can be constructed using the instantaneous forces and moments that can be computed from Tables 17.1 and 17.2 and Eq. 17.15. Such models account for periodic forces and moments due to fluctuations in dynamic pressure; however, they do not account for unsteady aerodynamic effects. Such simulations can be used to assess the efficacy of a cycle-averaged model-based control law when applied to a vehicle model that includes time periodic variations in the forces and moments due to the flapping wings without unsteady aerodynamic or structural flexibility effects. Directly modeling such effects can consume significant computational resources, which can produce a simulation that

17 Dynamics and Control of Flapping Wing MAVs

345

is impractical for the study of the behavior of control systems for flapping wing aircraft. It is recommended that such effects be captured in tabular form based upon the results of either high-fidelity computational methods or experiments, e.g., the low Reynolds number experiments of Sane and Dickinson (2001). Such results can be used to produce instantaneous lift and drag coefficients as a function of stroke angle, angle-of-attack, amplitude, frequency, and asymmetric frequency. This more complete representation may include translational and rotational aerodynamic contributions, unsteady aerodynamic effects, and fluid structure interactions. For each set of parameters considered in a table, the lift and drag coefficients should be estimated for a given stroke angle and angle-of-attack. Instantaneous estimates of the aerodynamic forces produced in this way should be representative of the steady-state periodic forces produced by the wings. In other words, transients due to changes in parameters other than stroke angle and angle-of-attack would not appear in such a model. Whether a simple blade element model or a higher-fidelity tablelookup is used, both simulations will provide the designer with a model that can be used to assess the performance of cycle-averaged model-based control laws on aircraft models that are driven by periodic forces and moments.

17.9

Conclusion

The methods presented in this chapter are intended for use with hover-capable flapping wing vehicles in the conceptual design phase. The methods allow engineers to iterate between vehicle design, trim, and controllability analysis until closure is achieved. Prototype model-based controllers can be designed by using the control derivative tables, the multivariable chain rule, and conventional control design methods. The resulting control laws provide a point of departure design that can be tested on the prototype aircraft or in higher-fidelity simulations. Tuning and optimization of the control law will likely be required to compensate for the effects of unmodeled unsteady aerodynamic effects, fluid-structure interactions, as well as interactions between wing loads, drive mechanism loads, and prime movers.

References M.L. Anderson, Design and control of flapping wing micro air vehicles. Ph.D. thesis, Air Force Institute of Technology, (2011) M.A. Bolender, Rigid multi-body equations of motion for flapping wing micro air vehicles using kane’s equations, in AIAA Guidance, Navigation, and Control Conference, Chicago, (2009). AIAA Paper 2010–6158 X. Deng, L. Schenato, W.C. Wu, S.S. Sastry, Flapping flight for biomimetic robotic insects: part I – system modeling. IEEE Trans. Robot. 22(4), 776–788 (2006a) X. Deng, L. Schenato, S.S. Sastry, Flapping flight for biomimetic robotic insects: part II – flight control design. IEEE Trans. Robot. 22(4), 789–803 (2006b) D.B. Doman, S. Regisford, Wing sizing, trim, and control consideration in the design of hovercapable flapping wing micro air vehicles, in AIAA Atmospheric Flight Mechanics Conference, Toronto, (2010). AIAA Paper 2010–7629

346

D.B. Doman et al.

D.B. Doman, M.W. Oppenheimer, D.O. Sigthorsson, Wingbeat shape modulation for flappingwing micro-air-vehicle control during hover. J. Guid. Control Dyn. 33(3), 724–739 (2010a) D.B. Doman, M.W. Oppenheimer, D.O. Sigthorsson, Dynamics and control of a biomimetic vehicle using biased wingbeat forcing functions: part i – aerodynamic model, in 48th AIAA Aerospace Sciences Meeting, Orlando, (2010b). AIAA Paper 2010–1023 D.B. Doman, C.P. Tang, S. Regisford, Modeling interactions between flexible flapping-wing spars, mechanisms, and drive motors. J. Guid. Control Dyn. 34(5), 1457–1473 (2011) I. Faruque, J.S. Humbert, Dipterian insect flight dynamics. Part 1 longitudinal motion about hover. J. Theor. Biol. 264(2), 538–552 (2010) M.V. Ol, K. Granlund, Abstraction of aerodynamics of flapping-wings: is it quasi-steady? in 50th AIAA Aerospace Sciences Meeting, Nashville, (2012). AIAA Paper 2012–0587 M.W. Oppenheimer, M.A. Bolender, D.O. Sigthorsson, D.B. Doman, Analysis of the translation motion of a flapping wing micro air vehicle, in AIAA Guidance, Navigation and Control Conference, Toronto, (2010a). AIAA Paper 2010–7708 M.W. Oppenheimer, D.B. Doman, D.O. Sigthorsson, Dynamics and control of a biomimetic vehicle using biased wingbeat forcing functions: part II – controller, in 48th AIAA Aerospace Sciences Meeting, Orlando, (2010b). AIAA Paper 2010–1024 M.W. Oppenheimer, D.B. Doman, M.A. Bolender, Control allocation, in Control System Application, ed. by W.S. Levine. Control Handbook, vol. II, 2nd edn. (CRC, Boca Raton, 2010c) M.W. Oppenheimer, D.B. Doman, D.O. Sigthorsson, Dynamics and control of a hovering biomimetic vehicle using biased wingbeat forcing functions. J. Guid. Control Dyn. 34(4), 647– 662 (2011a) M.W. Oppenheimer, D.O. Sigthorsson, D.B. Doman, Body torque generation for a flapping wing micro air vehicle by angle of attack change, in 49th AIAA Aerospace Sciences Meeting, Orlando, (2011b). AIAA Paper 2011–1281 S.P. Sane, M.H. Dickinson, The control of flight force by a flapping wing: lift and drag force production. J Exp. Biol. 204, 2607–2626 (2001) D.O. Sigthorsson, M. Oppenheimer, D. Doman, Flapping wing micro air vehicle aerodynamic modeling including flapping and rigid body velocity, in 50th AIAA Aerospace Sciences Meeting, Nashville, (2012a). AIAA-Paper-2012–0026 D.O. Sigthorsson,M. Oppenheimer, D. Doman, Insect sized flapping wing vehicles versus rotorcrafts, a comparative study, in 50th AIAA Aerospace Sciences Meeting, Nashville, TN.: (2012b). AIAA-Paper-2012–0028 B.K. Stanford, P.S. Beran, Analytical sensitivity analysis of an unsteady vortex method for flapping-wing optimization. J. Aircr. 47(3), 647–662 (2010) R.J. Wood, The first takeoff of a biologically inspired at-scale robotic insect. IEEE Trans. Robot. 24(2), 341–347 (2008)

Principles of Guidance, Navigation, and Control of UAVs

18

Gabriel Hugh Elkaim, Fidelis Adhika Pradipta Lie, and Demoz Gebre-Egziabher

Contents 18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2 Attitude and Position Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2.1 INS/GPS Integrated Navigation System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2.2 Implementation and Practical Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2.3 Flight Test Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3 Inner Loop Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3.1 PID Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3.2 Lateral Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3.3 Longitudinal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3.4 Trim Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.3.5 Simulation and Flight Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4 Guidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.1 General Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.2 Straight Line Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.3 LC 2 Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.4 Waypoint Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.5 Point Acquisition and RTB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.4.6 Simulation and Flight Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

348 351 352 357 359 359 361 363 364 367 367 369 370 371 372 374 376 377 378 379

G.H. Elkaim () Computer Engineering Department, University of California Santa Cruz, Santa Cruz, CA, USA e-mail: [email protected] F.A. Pradipta Lie • D. Gebre-Egziabher Aerospace Engineering and Mechanics, University of Minnesota, Minneapolis, MN, USA e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 56, © Springer Science+Business Media Dordrecht 2015

347

348

G.H. Elkaim et al.

Abstract

Two complete system architectures for a guidance, navigation, and control solution of small UAVs are presented. These systems (developed at the University of California Santa Cruz and the University of Minnesota) are easily reconfigurable and are intended to support test beds used in navigation, guidance, and control research. The systems described both integrate a low-cost inertial measurement unit, a GPS receiver, and a triad of magnetometers to generate a navigation solution (position, velocity, and attitude estimation) which, in turn, is used in the guidance and control algorithms. The navigation solution described is a 15-state extended Kalman filter which integrates the inertial sensor and GPS measurement to generate a high-bandwidth estimate of a UAV’s state. Guidance algorithms for generating a flight trajectory based on waypoint definitions are also described. A PID controller which uses the navigation filter estimate and guidance algorithm to track a flight trajectory is detailed. The full system architecture – the hardware, software, and algorithms – is included for completeness. Hardware in the loop simulation and flight test results documenting the performance of these two systems is given.

18.1

Introduction

In current usage, the term uninhabited aerial vehicles or UAVs refers to aircraft which fly without a human operator onboard. They span a wide range in size and complexity. The largest UAVs such as Predator or Global Hawk can weigh several thousand pounds and have wingspans on the order of 10 to 100 ft. At the other end of the size spectrum are UAVs whose maximum dimensions and mass are on the order of centimeters and grams, respectively. In general, these vehicles are being used or envisioned for use in operations where they serve primarily as a platform for a sensor payload. The sensor payload can be as simple as an electro-optical camera or as complex as synthetic aperture radar used for remote sensing. Regardless of their size and mission, all UAVs share the need for sensor or sensor systems which provide an estimate of the vehicle’s full state vector. The state vector normally consists of three position coordinates, three components of the velocity vector, and anywhere between three and nine parameters which describe the vehicle’s attitude. In addition to state sensing and estimation, UAVs (especially one that operates autonomously) need control and guidance systems which allow them to maneuver in a way consistent with their mission. In simple terms, the guidance system generates instruction on what state trajectory the UAV should follow in accomplishing its mission. The control system in turn operates the aircraft controls (aileron, elevators, thrust, etc.) to follow the trajectory generated by the guidance systems. A high-level depiction of a GNC system is shown in Fig. 18.1 and consists of both airborne and ground components. Off-the-shelf guidance, navigation, and control (GNC) solutions for small UAVs (i.e., UAVs which fall in the so-called “Class I” category of vehicles as defined in UAS CoE 2010) exist. Some of these off-the-shelf solutions consist of the

18 Principles of Guidance, Navigation, and Control of UAVs

349

Fig. 18.1 Unmanned aerial system (UAS) concept of operation

vehicle itself, avionics, and associated ground support systems. While most of these “turnkey” solutions allow some level of customization by the user, they can be very restrictive for use in research environments. For example, in research on the robust control problem, the effect of uncertainties and potential fault modes in the guidance and navigation algorithms must be known. Very few off-the-shelf solutions allow access to the internals of the GNC solution. Thus, a GNC solution that is inexpensive and easily reconfigurable is desirable for the research community. This chapter describes two such GNC solutions for small UAVs. This chapter discusses the implementation aspect of the GNC modules on University of California Santa Cruz’s (UCSC) and University of Minnesota’s (UMN) UAVs. The UAV laboratory at UMN is equipped with two fixed-wing aircraft models: Ultrastick 25e and Ultrastick 120. Shown in, Fig. 18.2 is the Ultrastick 120, also known as FASER (Free-flying for Subscale Experimental Research). FASER is equipped with conventional flight control surfaces and a pair of wingtip vanes on each side of the aircraft to measure aerodynamic angles (angle-of-attack and sideslip angle). Prior to joining the UAV Laboratory, FASER’s airframe was formerly used by NASA for their research program (Owens et al. 2006). Extensive wind tunnel testing during its time at NASAs has been used to develop a comprehensive aerodynamic model of the aircraft. At University of Minnesota, FASER is the workhorse for the multi-sensor navigation research such as GPS attitude and heading determination system, synthetic airdata estimator (Lie and Gebre-Egziabher 2012), and visionaided navigation (Chu et al. 2011). Thor, an Ultrastick 25e, is an approximately

350

G.H. Elkaim et al.

Fig. 18.2 Free-flying for Subscale Experimental Research (FASER) UAV (a modified Ultrastick 120)

Fig. 18.3 The UCSC SLUGS autopilot (a) and Multex Mentor UAV (b)

65 % scaled model of FASER with the same basic configuration. As an addition to the system identification work (Dorobantu et al. 2011), this UAV is used for faulttolerant control research at the University. The UAV Laboratory Web site (Murch 2012b) provides a detailed information on the airframe, avionics, and software architecture. The flight software is available as an open source software package available upon request. The UCSC autopilot, Santa Cruz Low-Cost UAV GNC Subsystem (SLUGS), has been developed over the past 5 years in order to implement a rapidly reconfigurable autopilot for UAV guidance, navigation, and control research. The SLUGS, pictured in Fig. 18.3a, consists of two fast dsPIC33 microcontrollers (DSC’s) and a suite of sensors; the SLUGS-based electric UAV is pictured in Fig. 18.3b. Details of the design, development, and deployment of the SLUGS can be found in Lizarraga (2009) and Lizarraga et al. (2011a,b, 2009a,b). Section 18.2 describes the UMN’s navigation module for attitude and position estimation. Sections 18.3 and 18.4 describe the inner loop control and guidance algorithm implemented on the UCSC SLUGS platform, respectively, including simulation and experimental results.

18 Principles of Guidance, Navigation, and Control of UAVs

18.2

351

Attitude and Position Estimation

In order to be able to guide and control the UAV, the state of the UAV must be available to the controller at high fidelity and high bandwidth. Accurate position is required to perform automatic control for precision applications (e.g., landing). The navigation module in the GNC system implements aircraft state estimation. The advent of powerful computers made available at relatively low cost has allowed sensor fusion for navigation; sensor fusion is a technique of optimally blending information from multiple different (flawed) sensors. Multi-sensor navigation framework aims to obtain the most information about the aircraft states by using minimum combination of sensors. This is key to UAV operations where size, weight, and power are all critical. This section will describe the navigation module implemented on University of Minnesota’s UAV. The complete state of the UAV comprises its position, velocity, attitude, airspeed, angle-of-attack, sideslip angle, and rotation (pitch, roll, and yaw) rates. Position, velocity, and attitude are also known as the navigation state (Gleason and GebreEgziabher 2009). The sensors used to measure these quantities are called navigation sensors: an inertial measurement unit (IMU) and global positioning system (GPS) receiver. Figure 18.4a shows the Analog Devices IMU ADIS16405 installed on the UAVs at University of Minnesota. It is a 6 degree-of-freedom temperature calibrated inertial measurement unit with 3-axis accelerometers, 3-axis gyros, and 3-axis magnetometer. With a flight computer that computes the navigation state only from IMU measurements, it is known as an inertial navigation system (INS). Figure 18.4b shows Crescent, an OEM GPS receiver from Hemisphere GPS. In addition to the differentially corrected position and velocity estimates, Crescent also outputs raw pseudorange and carrier phase measurement. These features enable a great deal of navigation research at the UMN UAV Laboratory. Crescent’s small form-factor and low power consumption also makes it a very suitable choice for UAV applications.

Fig. 18.4 Navigation sensor of UMN UAV: ADIS16405 (a) and Hemisphere Crescent OEM GPS receiver (b)

352

G.H. Elkaim et al.

Fig. 18.5 UMN’s Thor instrumented with five-hole pitot tube from Goodrich to measure airspeed, ˛, and ˇ

Airspeed, angle-of-attack .˛/, and sideslip angle .ˇ/ are known as the airdata quantities, which are traditionally measured using airdata sensors such as pitot tube and wind vane. Another alternative for measuring these angles is to use a multiple hole pitot tube such as the 5-hole pitot tube as shown in Fig. 18.5. Recent work (Lie and Gebre-Egziabher 2012) has shown that this airdata can also be estimated by using sensor fusion from the IMU and GPS. The navigation state is of particular interest. INS/GPS integration has been studied extensively in the past decades to estimate aircraft’s navigation state. This section will focus on the implementation aspects of an integrated navigation systems using inertial sensors and GPS (INS/GPS system). A brief description of INS/GPS is provided; a more in-depth discussion can be found in Gleason and Gebre-Egziabher (2009), Groves (2008), Farrell and Barth (1999), and Titterton and Weston (2004). System architecture and implementation challenges are discussed next. Following these descriptions, FASER’s flight test result is presented.

18.2.1 INS/GPS Integrated Navigation System An inertial navigation system (INS) uses the output of inertial sensors to estimate the vehicle’s position, velocity, and attitude. A complete six degree-of-freedom inertial sensor consists of 3-axis accelerometers and 3-axis gyros. The accelerometer measures the specific force acting on the platform and the gyros measures its rotation. Inertial sensors can be categorized according to its resulting navigational accuracy (Gleason and Gebre-Egziabher 2009; Gebre-Egziabher 2001). Automotive-grade micro electro-mechanical system (MEMS) inertial sensors are most suitable for low-cost UAV applications; however, when operating as a stand-alone navigator, these sensors produce positioning errors on the order of several hundreds of meter per minute. These large errors in the position, velocity, and attitude estimates are mainly due to sensor bias and noise that corrupt the measurements. They are also a

18 Principles of Guidance, Navigation, and Control of UAVs

353

function of the initial error on the state estimates. The position error grows linearly with the initial velocity error estimate, quadratically with uncorrected bias, and at a cubic rate with attitude error (Titterton and Weston 2004). Despite the unbounded growth of the error with time, one of the important advantages of inertial navigators is that they are self-contained; that is, they require no external signal to provide a navigation solution. In terms of the data rate, the navigation states can be computed at a rate limited only by the processing power of the flight computer. Contrast GPS to the advantages and disadvantages of an inertial navigator: GPS requires an external signal to operate, and although the navigation solution accuracy depends on the signal quality and the geometry of the satellite in view, this error is not a function of time (Misra and Enge 2001). Although there are specialized GPS receivers that can generate data up to 100 Hz, most low-cost GPS receivers provide output data at 1–10 Hz. In other words, although GPS navigation solution has longterm stability, the bandwidth of the solution is much lower than that of an inertial navigator. Integration of INS with GPS allows a navigation solution that has the high bandwidth of the inertial sensors and the drift-free long-term stability of the GPS solution. Attitude determination is an integral part of strapdown INS. This is because the specific force measured by the accelerometer needs to be transformed into the navigation frame in which the position and velocity are calculated. This transformation calls for the knowledge of the orientation of the platform (defined as the attitude). Attitude can be equivalently described by a set of three angles known as the Euler angle sequence or the four-parameter attitude quaternion. An Euler angle sequence consists of the yaw . /, pitch ./, and roll ./ angles that describe three successive rotations about the body z, y, and x axes, respectively. Although this representation carries physical interpretation, it is singular at  D ˙90ı . The attitude quaternion is a set of four numbers that can be related to the roll, pitch, and yaw angles using the following relationship:   D arctan

2q2 q3 C 2q0 q1 2q02 C 2q32  1



 D arcsin .2q1 q3 C 2q0 q2 /   2q2 q3 C 2q0 q1 D arctan : 2q02 C 2q32  1

(18.1)

  The quaternion q D q0 q1 q2 q3 represents the transformation from the local North-East-Down coordinate to the body-fixed frame. Although it has no singularity, directly visualizing the aircraft’s orientation can be challenging. Hence, for remote pilot display purposes and most control algorithms, the quaternion attitude needs to be transformed into its corresponding roll, pitch, and yaw angles. For applications where only benign maneuvers are expected, the Euler angle representation can instead be used.

354

G.H. Elkaim et al.

Attitude determination using gyro’s output calls for integrating the angular velocity to propagate the attitude forward in time. Since gyros measure inertial rotation, they must be compensated to account for both the earth’s rotation rate and the transport rate (Groves 2008) due to the Earth’s curvature. For most low-cost UAV applications, these terms are small (105 rad=sec) compared to the noise level in the sensors. Hence, they can be neglected for all practical purposes. When calculating attitude at the IMU sample rate, the following equations can be used: for an Euler angle attitude representation: 3 2 1 Œk C 1 4 Œk C 1 5 D 4 0 0 Œk C 1 2

sin.Œk/ tan.Œk/ cos.Œk/ sin.Œk/ sec.Œk/

3 32 B !x Œkt cos.Œk/ tan.Œk/ 5 4 !yB Œkt 5  sin.Œk/ cos.Œk/ sec.Œk/ !zB Œkt (18.2)

and for a quaternion attitude representation:   qŒk C 1 D qŒk ˝ 1 12 !x ŒkB t 12 !y ŒkB t 12 !z ŒkB t

(18.3)

where ˝ operator is the quaternion multiplication operator. Details on quaternion algebra can be found on Lefferts et al. (1982). In order to improve the navigation solution between subsequent GPS solutions, the filter makes frequent corrections to compensate for the inertial sensor errors. Although more sophisticated sensor error models exist, a simplified model presented in Gebre-Egziabher (2001) is used in this implementation. In this model, the sensor output (e.g., gyro) as a function of time can be written as !.t/ D !.t/ Q C bgs C bgd .t/ C wg

(18.4)

where !.t/ Q represents the true angular velocity, bgs is the turn-on bias, bgd .t/ is a time-varying bias, and wg is measurement noise that can be regarded as white noise. bgd .t/ is modeled as first-order Gauss-Markov process, written as 1 bPgd .t/ D  bgd .t/ C wb : 

(18.5)

This model is robust to parameters that are unobservable when the UAV is not accelerating (Gleason and Gebre-Egziabher 2009). Using this model, the estimated sensor bias is not simply the true bias corrupting the measurement; it also accounts for all unmodeled errors that corrupt the sensor measurement. The architecture of the INS/GPS filter using an extended Kalman filter (EKF) (Simon 2006) is shown on Fig. 18.6. Included in the state vector (Euler angle representation) are: xDŒ„ L ƒ‚ ƒ … h VNorth VEast VDown   b bay baz bgx bgy bgz  : (18.6) „ ƒ‚ … „ ƒ‚ … „ax ƒ‚ … „ ƒ‚ … Position

Groundspeed

Attitude

AccelBias

GyroBias

18 Principles of Guidance, Navigation, and Control of UAVs

355

Fig. 18.6 Block diagram of INS/GPS integration

The time update process, executed at the internal IMU sample rate, is: 

Œk C 1 Œk C 1

Œk C 1

T

D f .Œk; Œk; Œk/  ! B Œk  1  t (18.7)

VN Œk C 1 D VN Œk C CBN ŒkfB Œk  t

(18.8)

hŒk C 1 D hŒk  VDownŒk  t ƒŒk C 1 D ƒŒk C

VEast Œk  t RE C hŒk

LŒk C 1 D LŒk C

VNorth Œk  t RE C hŒk

(18.9)

The covariance is updated according to P./ Œk C 1 D ˆŒkP.C/ ŒkˆŒkT C QŒk:

(18.10)

The state transition matrix can be approximated as ˆŒk D I C FŒk  t. F is the Jacobian of the time-update equations and Q is the process noise covariance matrix. Details on the elements of F and Q can be found in Gleason and Gebre-Egziabher (2009). When GPS position and velocity becomes available at time-step k, the measurement update process is outlined as follows. The measurement, y, is defined as follows: ˇ  (18.11) yjGPS D pN pE pD VNorth VEast VDown ˇGPS where pN , pE , and pD are the reported North-East-Down position with respect to a specific location, e.g., the initial position.

356

G.H. Elkaim et al.

Define the state error as: ˇ  ıy D yjGPS  pN pE pD VNorth VEast VDown ˇINSŒk :

(18.12)

The covariance is updated using 

 I33 033 039 033 I33 039  1 KŒk D P./ ŒkHT R C HP./ ŒkHT HD

(18.13)

P.C/ Œk D .I  KŒkH/ P.C/ Œk .I  KŒkH/T C KŒkRŒkKŒkT : The state is updated according to: ıx D KŒkıy

(18.14)

Position: h.C/ Œk D h./ Œk  ıx.3/ ƒ.C/ Œk D ƒ./ Œk C

ıx.2/ RE C h.C/ Œk

L.C/ Œk D L./ Œk C

ıx.1/ : RE C h.C/ Œk

(18.15)

Velocity: .C/

./

VNorth Œk D VNorth Œk C ıx.4/ .C/

./

VEast Œk D VEast Œk C ıx.5/ .C/

(18.16)

./

VDown Œk D VDown Œk C ıx.6/: When using Euler angle representation, the attitude is updated using the transformation matrix CBN . CBN .C/ D .I  Œıx.7 W 9//  CBN ./

(18.17)

where Œ indicates the skew-symmetric matrix operator. Euler angles can be derived from the updated transformation matrix, CBN .C/. Accelerometer and gyros bias are updated as follows: ./ b.C/ a Œk D ba Œk C ıx.10 W 12/:

(18.18)

./ b.C/ g Œk D bg Œk C ıx.13 W 15/:

(18.19)

18 Principles of Guidance, Navigation, and Control of UAVs

357

Fig. 18.7 Flight software architecture

18.2.2 Implementation and Practical Challenges The avionics suite in the UMN’s UAV is implemented on a 32-bit PowerPC Phytec MPC5200B-tiny SoM. It utilizes a real-time operating system (RTOS) and flight software written in C. This computer handles data acquisition; performs guidance, navigation, and control tasks; stores relevant data; and sends information to the ground control station via the telemetry modem. The flight software uses a multithreaded architecture in which all of the flight critical tasks execute in the highest priority thread at 50 Hz, while additional tasks (i.e., those not required to control the aircraft, such as a fault detection filter) are executed in separate, lower-priority threads (Murch 2012a). As shown in Fig. 18.7, there are ten modules executed in the highest-priority thread (thread 0). Navigation tasks are executed immediately after data acquisition (DAQ). It starts as valid GPS measurements are available. The position and velocity are initialized at the reported position and velocity. Since this initialization occurs on the ground, attitude is initialized to a known attitude of the aircraft on the ground. One of the key aspects to the correct fusion of inertial sensors and GPS information is aligning data acquisition from non-timing sensors, such as an IMU, with sensors that have inherent timing capability such as GPS. Data acquisition to an IMU is achieved by polling the IMU at the desired sample rate. GPS receivers, however, stream out their navigation solutions (position and velocity) at a preset rate (e.g., 1 Hz). This rate is usually driven by a receiver clock that has been steered to the GPS time. Despite being precise, the flight computer clock might be running at a different rate from that of the GPS clock. When the IMU data stream is not aligned with the GPS output, incorrect measurements are fed to the EKF with resulting errors in the vehicle state estimates. The magnitude of the error depends on the acceleration experienced by the aircraft and the clock rate difference.

358

G.H. Elkaim et al.

Thread 0 DAQ

Thread 1

NAV

GL

SF

CL

SI

SF

DL

A

TM

Research Algorithm #1

Thread 2

Research Algorithm #2

0.0

5.0

10.0

15.0

20.0

Time, msec

Fig. 18.8 Flight software timing diagram

There are many ways to synchronize IMU and GPS outputs. One way to do this is by performing data acquisition at a deterministic sample rate. By checking the availability of GPS solution at the highest rate available in the flight computer, IMU and GPS data are always synchronized within the resolution of one sampling interval. In this architecture, this is achieved by using the alarm functions provided by the RTOS kernel which are used to trigger module execution. The timing diagram that shows the execution time of each module is shown in Fig. 18.8. Tuning the extended Kalman filter is known to improve filter convergence time and short-term accuracy (Gleason and Gebre-Egziabher 2009). However, in order to tune the filter, a truth reference is needed so that a comparison can be made to the resulting filter estimate. Although tuning the filter is an art, experience can help smooth the process. Having a truth reference system might be prohibitive for most small scale UAVs; it is recommended that the data set used for tuning is from a known trajectory. This helps to develop a feeling for how the errors grow and how effective is the bias compensations. Since the states’ observability depends highly on the acceleration experienced by the vehicle, it is important that a trajectory with sufficient acceleration is used: acceleration and deceleration on a straight line path and frequent turns should be included. It might also be worth testing the filter performance both during powered and gliding operation since structural vibration might adversely affect the filter performance. Adding damping to the system or simply moving the IMU placement on the UAV can often alleviate vibration problems. Lastly, when the IMU is not collocated with the GPS antenna, the distance between the IMU and the GPS antenna results in different velocity and position measured by both sensors. This is known as the lever arm problem, and the Kalman filter must account for it to make correct state estimates. There are generally two ways to account for this error. The first way is to transfer the GPS measurement to the location of the IMU by using the information of the aircraft attitude and the known lever arm vector in the body frame. The dependency on estimated attitude to transfer this measurement might lead to inaccurate GPS position and velocity estimates. The second way is by inflating the GPS measurement covariance. This method fails when the inflation factor is so large that it renders the GPS position

18 Principles of Guidance, Navigation, and Control of UAVs

359

and velocity useless. For most UAV applications, however, the distance between the GPS antenna and the IMU is not large, thus only a very small inflation required to compensate for this error. In this work, the latter method is employed.

18.2.3 Flight Test Result Figure 18.9 shows the trajectory flown during one of FASER’s flight test. The navigation data are extracted from the flight computer’s data recorder and are plotted on Fig. 18.10. During the flight, there were several GPS outages, and the navigation state estimates during these period of free inertials are plotted in green. When bias corrections are continuously fed to the IMU, this inflight calibration improves the performance of the inertial navigation and slows down error growth during these GPS outages. In Fig. 18.10a, b, the position and velocity estimates from INS/GPS integration system are differenced with the velocity from the GPS measurement. Due to lack of a reference (truth) system, it is not possible to compare the attitude and bias estimates. Hence, the attitude and bias estimates from the INS/GPS filter are plotted on Fig. 18.10c, d. Qualitative evaluation of the filter performance indicates good performance. This can be seen from the smoothness of the INS/GPS solution and good convergence of the inertial sensor bias estimates. As such, the INS/GPS integration provides a high-bandwidth solution with long-term stability suitable for UAV control.

18.3

Inner Loop Control

In order to fly an aircraft, a low-level control system must stabilize the airframe using available sensor inputs and actuators. A higher-level outer loop control will implement path following (see Sect. 18.4) while the inner loop keeps the aircraft flying. There are myriad ways to implement an inner loop control on a UAV. In this chapter the inner loop controller is based on the SLUGS autopilot (details at Lizarraga (2009), ASL, and SLU), which uses relatively simple PID controllers in the inner loop. The PID-based inner loop control that has flown on the SLUGS platform is presented, and actual flight data of the implementation is included. The UCSC SLUGS autopilot is divided into two hardware sections: a control processor and a sensor processor. The sensor processor is tasked with taking the raw sensor measurements and fusing them into a high-quality position and attitude estimate (similar to Sect. 18.2). The control processor is tasked with both the inner loop (stabilization) and the outer loop (guidance). The SLUGS implementation is illustrated in Fig. 18.11. While there are many different ways to implement a low-level control system, the one presented is based on decoupling the longitudinal and lateral flight dynamics into two separate control systems (Stewart 2001). The lateral control channel

360

G.H. Elkaim et al. Trajectory in North−East Plane GPS INS/GPS INS only

700 600 500

North (m)

400 300 200 100 0 −100 −200 −300 −400

−200

0

200 East (m)

400

600

800

Altitude 550

500

Altitude (m)

450

400

350

300

250

200

0

100

200

300

400

500

600

700

Time (sec)

Fig. 18.9 Flight test trajectory

controls rudder and ailerons, while the longitudinal control handles throttle and elevator. While this is reasonable for traditional aircraft-like UAV’s flying gentle maneuvers, it is not appropriate for aircraft with high degrees of lateral-longitudinal cross coupling, nor for aircraft performing aggressive aerobatic maneuvers.

18 Principles of Guidance, Navigation, and Control of UAVs

200

300

400

500

600

700

0 −20

0

100

200

300

400

500

600

700

10 0 −10

0

100

200

300

400

500

600

700

0

100

200

300

400

500

600

700

0

100

200

300

400

500

600

700

0

100

200

300

400

500

600

700

10 0 −10 5 0 −5

Time (sec)

d

Euler Angles 0 −100

0

100

200

300

400

500

600

700

0.2 0

0

200

400

600

800

1 2

a y (m/s )

θ (deg)

50 0 −50

0

100

200

300

400

500

600

700 2

a z (m/s )

200

ψ (deg)

Accel Bias

0.4

2

a x (m/s )

φ (deg)

100

p bias (deg/s)

c

Time (sec)

0 −200

0

100

200

300

400

Time (sec)

500

600

700

0 −1

0

200

400

600

800

0.2 0

−0.2

0

200 400 600 Time (sec)

800

q bias (deg/s)

100

0 −10

r bias (deg/s)

0

20

VE Error (m/s)

0 −50

Velocity Difference

10

VD Error (m/s)

East Error (m) Altitude Error (m)

Position Difference

50

VN Error (m/s)

b

North Error (m)

a

361

Gyro Bias

2 0 −2

0

200

400

600

800

0

200

400

600

800

0.5 0 −0.5 0 −2 −4

0

200 400 600 Time (sec)

800

Fig. 18.10 INS/GPS solution: position error (a), velocity error (b), attitude estimate (c), and sensor bias estimate (d)

Furthermore, each of the two main (lateral and longitudinal) controllers is divided into successive loop closure using proportional-integral-derivative (PID) controllers. PID controllers are simple to tune experimentally, and while they can theoretically perfectly stabilize a second-order plant, in practice they work quite well with higher-order plants. Some attention must be paid to integrator windup, and the use of a numerical derivatives in the PID loop, or the resulting controller will have poor performance. Lastly, these controllers are implemented digitally, in order to accommodate the actual autopilot hardware.

18.3.1 PID Control As stated above, the proportional-integral-derivative (PID) control can perfectly stabilize a second-order plant, given the right gains. However, in practice, it can usually perform well in a variety of settings without the need for a precise model of the underlying plant (a large advantage when system identification is either difficult or imprecise). The classical PID controller consists of an input derived from the measurement of the system and the desired reference for it to track and an actuator signal on the output. The difference between the measurement and the reference,

362

G.H. Elkaim et al. Sub-mode Selector

Inner Loop

ψc

dr Lateral Channel

hc Uc

da UAV

de Longitudinal Channel

dt

Controller Gains UAV Response Navigation

Waypoints Navigation Config

Ground Control Station Operator Input

Fig. 18.11 SLUGS block diagram of inner loop and guidance

referred to as the error, is fed into the PID block, and an output is generated from the three gains: Kp , the proportional gain is multiplied directly by the error; Kd , the derivative gain is multiplied by the time rate of change of the error; and lastly, Ki , which multiplies the integral of the error. At its most basic form, the PID control simply consists of three gains, with a first-order difference for the derivative and a running sum for the integral. Several embellishments are used to make the PID controller behave better at the corner cases. Firstly, the derivative is changed from the time rate of change of the error to the time rate of change of the measurement; this is done so that the PID controller does not throw the output due to a reference change. If a direct derivative signal is available (for instance, pitch rate gyro), then that is used in preference to the firstorder difference. Also, the first-order difference can be changed to longer time steps if noise sensitivity is an issue. Secondly, saturation limits are included to clip the output if it would saturate the actuator. The integral is changed from a running sum to trapezoidal integration, and anti-windup logic is included. Integral state windup is caused by the integrator continuing to integrate even after the control is saturated. This creates a memory effect within the controller that causes overshoot and degrades the controller performance when coming out of saturation. In order to prevent this, an anti-windup scheme is implemented which checks if the actuator would saturate on the current time step and does not perform the integration if this is the case.

18 Principles of Guidance, Navigation, and Control of UAVs

Commanded Turn-rate

Compute Roll

+

Commanded Roll

−

Roll Measured

Airspeed

Lateral Acceleration

363

PID Controller

PID Controller

–1

Aileron Command

Rudder Command

Fig. 18.12 SLUGS lateral inner loop control

The integrator state can be reset under software control, such that when the control is engaged, a bumpless transfer is achieved. Furthermore, the integral state can be manipulated when changing the control gains, such that the output remains stable as the new gains are latched into the controller. Again, all of this is done to maintain smooth control for flight.

18.3.2 Lateral Control The lateral controller (Fig. 18.12) uses the rudder and ailerons to keep the aircraft flying in a coordinated turn and following a commanded turn rate (including a zero turn rate for straight flight). The lateral dynamics of an aircraft include the roll-rate damping mode, a spiral mode, and the dutch roll mode (yaw-roll coupling). In this formulation of the inner loop control, yaw rate, P c , is commanded and is converted to a commanded roll angle, c , using the formulation c D arctan

P c Um g

! (18.20)

where Um is the measured airspeed and g is gravity. Equation 18.20 assumes that the aircraft is flying in a coordinated turn, that is, where the turn rate is constant and the body-fixed lateral acceleration, ay , is zero. The commanded roll angle, c , is used as the reference to the PID control, with the plant output being the actual roll angle, , that comes from the attitude estimation algorithm. The commanded bank angle is limited with a saturation block to keep the roll angle from becomming too extreme. In the case of the SLUGS small UAV, this is limited to ˙40ı . The output generated by the PID block is to the ailerons, which are used to drive the roll error (commanded  actual) to zero. In this case, the derivative of the roll

364

G.H. Elkaim et al.

error is taken directly from the body-fixed gyros (with the bias again taken care of by the attitude estimation algorithm). That is, while the full derivative of roll rate P D p C .q sin  C r cos / tan 

(18.21)

where Œp; q; r are the body-fixed roll, pitch, and yaw rates and ;  are the roll and pitch euler angles, respectively. Note that with a small pitch angle and yaw rate, this can be approximated as P ' p: (18.22) Lastly, when the roll angle command, c , is the opposite polarity from the previous command, the integral state is reset to improve roll tracking performance. The above describes the turn rate command to aileron control system, which will drive the roll error to zero. However, Eq. 18.20 depends on coordinated flight to be effective. In order to ensure coordinated flight, a second PID loop is closed around to rudder. The commanded input to the turn coordinator PID block is the negative of the lateral acceleration (this accounts for the fact that the rudder is behind the center of mass of the aircraft and has a negative sign in the transfer function), and the output is the rudder actuator command. The PID loop will actuate the rudder in order to drive the body-fixed lateral acceleration, ay , to zero. Here, again, the fact that the output of the aircraft affected by the rudder has a direct measurement of its derivative from the body-fixed gyros can be used for the D term. That is, P D

1 .q sin  C r cos / cos 

(18.23)

which for small angles of pitch and roll, can be reliably approximated as P ' r:

(18.24)

Note that the turn coordinator, with the derivative feedback from the body-fixed yaw rate gyro, r, has the effect of also acting as a yaw damper for the aircraft. More traditional autopilots often use the rudder solely as a yaw damper, without the turn coordination function, and rely on the directional stability of the aircraft to keep the turns coordinated. In that case, the rudder feedback command would be based on a high-passed version of the yaw rate gyro. Experience with small UAVs indicates that the former approach works better for model scale aircraft, giving both coordinated turns and decent yaw damping.

18.3.3 Longitudinal Control The longitudinal low-level control loops (Fig. 18.13) operate similarly to the lateral ones, with the other two flight controls being throttle and elevator. There are two

18 Principles of Guidance, Navigation, and Control of UAVs

Commanded Airspeed

Dynamic Pressure Altitude

Compute Airspeed

365

PID Controller

−

Throttle Command

Airspeed

Compute r

Altitude Command

Altitude

−

PI Controller

Pitch Command

−

PID Controller

+

Elevator Command

Pitch Measured Roll Feedforward Roll Measured

1 1

k

cos(f)

Fig. 18.13 SLUGS longitudinal inner loop control

modes of operation for a longitudinal low-level controller: climb/descent and level flight. In climb/descent mode, the throttle is set to a fixed value (high for climb, low for descent), and the airspeed and climb rate are controlled through the aircraft pitch via the elevator. In level flight mode, airspeed is controlled directly by the throttle and altitude by aircraft pitch via the elevator. Since airspeed is being kept constant, the altitude responds rapidly to even small changes in aircraft pitch. As a design choice, only the level flight mode is implemented, given that the UAV will spend most of its mission in level flight at constant altitude following waypoints. The airspeed hold control loop takes in a commanded airspeed, Uc , and compares this to the measured airspeed. Note that the airspeed control uses airspeed, and not ground speed (as available from GPS). Compensation for wind is handled within the higher-level guidance controller rather than in the low-level stabilizing controllers. The aircraft does not, however, have an onboard measurement of airspeed. Rather, it has an onboard measurement of dynamic pressure, q, and altitude, from which the airspeed can be extracted. That is, s Um D

2q 

(18.25)

where q is the dynamic pressure and  is the atmospheric density. However,  is a function of altitude, such that

366

G.H. Elkaim et al.

  D 0 1 

hm 44331

4:255876 (18.26)

where 0 is the sea-level air density (1.025 Kg/m3 ) and hm is the measured altitude in meters. This approximation is good for most altitudes reachable by UAVs and is sufficient for control. The output of the PID block is the throttle actuator. In the case of the airspeed, there is no direct derivative to be used in the throttle loop (though the body-fixed longitudinal acceleration could be used). As such, some care must be used in both the derivatives and in tuning the airspeed hold loop in general. Too aggressive gains on the airspeed hold loop causes the throttle to surge and cut in flight, with unpleasant results. This is due both to the noise on measured airspeed and also due to the lag in response to the throttle input. Here tuning the gains for an acceptable error and relying on the integral term to close the error works well. The altitude control loop consists of two chained PID controllers. The first is a straight PI controller (the derivative gain is set to 0) that takes in commanded altitude, hc , and measured altitude, hm , as its error and outputs a pitch command. The pitch command is saturated at ˙15ı with the integral windup stopped if the saturation is reached. The limited pitch command is used as a reference input for the second PID loop and is compared to the aircraft pitch, , from the attitude estimation to generate the error. The output of the second PID loop drives the elevator, which works to match the aircraft pitch attitude to the commanded one. Here, the derivative term comes straight from the body-fixed gyros, with the full equation P D q cos   r sin 

(18.27)

which for small bank angles can be reduced to P ' q:

(18.28)

The direct derivative term allows the pitch to elevator PID control to be aggressive, yet still retain good damping characteristics. This PID controller will drive the aircraft pitch to the commanded pitch until the desired altitude is reached. While the aircraft is flying straight and level, this control works very well. However, while the aircraft is in a turn, the aircraft will descend. This is because the lift from the wings must match the aircraft weight, but in a turn part of the lift is directed inwards to cause the turn itself. More precisely, in order not to lose altitude in a turn, the lift is related to the weight and the bank angle as LD

W cos 

(18.29)

where L is lift, W is aircraft weight, and  is roll angle. This is a simplified model assuming lift only from the wings and not accounting for sideslip angles of the fuselage for knife-edge type flight. Using this model, the change in lift is:

18 Principles of Guidance, Navigation, and Control of UAVs

 L D W

1 1 cos 

367

 (18.30)

which indicates how much the lift must be increased to maintain altitude in a coordinated circular turn. Pilots are trained to increase elevator (back stick) when entering a turn to maintain altitude. In order to ensure that the UAV holds altitude during turns, a feedforward gain proportional to the increase in lift is included:  ıeff D Kff

1 1 cos 

 (18.31)

where ıeff is the additional elevator command due to the feedforward term. Note that in the implementation, the roll angle, , is low pass filtered to reduce unwanted pitch oscillations resulting from attitude estimation noise and is also limited to ˙60ı . At angles beyond this amount, the UAV will simply enter into an accelerated stall trying to hold altitude, and the low-level control system will no longer be able to stabilize the aircraft.

18.3.4 Trim Conditions Before leaving the subject of the low-level inner loop controllers, it is necessary to discuss aircraft trim conditions. In the case of the SLUGS small-scale UAV, trim is established by the safety pilot before the autopilot is engaged. At the moment the autopilot is switched on, the inner control loops assume the aircraft is currently trimmed for straight and level flight. The inner PID control loops above each output a change in actuator output summed with the original trim condition. Note again that for the SLUGS autopilot, this trim is established by a human safety pilot. The trim could just as easily be established using a conventional 6DOF model of the airframe and computing trim for various flight conditions. That is, for every combination of airspeed, climb rate, and altitude, there is a throttle and elevator setting that will keep the UAV flying in steady state at those conditions. The ailerons and rudder surfaces are assumed to be set at zero for straight flight.

18.3.5 Simulation and Flight Test Results In order to validate the inner loop control, data from the hardware-in-the-loop (HIL) simulation is presented showing the results of the inner loop PID controllers. Figure 18.14a shows the lateral controller. Note that at every change in commanded roll angle, the lateral acceleration receives a large disturbance, which must be washed out. In panel (b) is the longitudinal control, with both pitch and airspeed control. Again, note that the spikes in airspeed are due to large climb or descent inputs to the controller.

368

G.H. Elkaim et al.

a

Lateral Channel 40

Measured Commanded

φ (deg)

20 0 –20 –40 0

50

100

150

200

150

200

Lateral Acceleration (m/s)

Time(s)

1 0.5 0 –0.5 –1 0

50

100 Time(s)

Airspeed (m/s)

b

Longitudinal Channel Measured Commanded

22 21 20 0

50

100

150

200

150

200

150

200

Time(s)

θ (deg)

20 10 0 –10 –20 0

50

100 Time(s)

Altitude (m)

520 515 510 505 500

0

50

100 Time(s)

Fig. 18.14 Simulation of the inner loop control using the HIL simulator: (a) the lateral channel, (b) the longitudinal channel

18 Principles of Guidance, Navigation, and Control of UAVs

369

Inner Loop Measurements. AP

φ (deg)

10

Commanded Measured

0 –10 −20 −30 260

270

280

290

300

260

270

280

290

300

310

320

330

340

350

360

310 320 Time (s)

330

340

350

360

θ (deg)

10 5 0

0.5 0 −0.5 −1 −1.5

U (m/s)

A y (m/s2)

−5

260

280

300 320 Time (s)

340

360

14.8 14.6 14.4 14.2 14 260

280

300 320 Time (s)

340

360

Fig. 18.15 SLUGS inner loop flight test data

Flight test data from the inner loop control is presented in Fig. 18.15, which shows excellent performance on pitch and roll commands, a reasonable attenuation of lateral acceleration (with a limit to less than 1 g for most of the flight time) and a fairly good airspeed hold. This is in the presence of wind, gusts, and other disturbances. Note that the bias on the lateral acceleration is most likely from the accelerometer being mounted with a slight tilt on the airframe.

18.4

Guidance

UAV guidance, navigation, and control (GNC) ultimately signifies the ability to follow a desired trajectory through the sky. With attitude estimation established (Sect. 18.2), and inner loop stabilizing the aircraft (Sect. 18.3), what is left is to guide the UAV along the desired trajectory rejecting disturbances such as wind. Again, while there are several ways to implement this guidance control, this section discusses the guidance algorithms implemented on the SLUGS autopilot (Lizarraga 2009). In order to deliver a mission to the UAV from the ground station, a simple set of GPS-based waypoints, along with an altitude and speed for each leg of the mission, is transmitted to the UAV via telemetry link. Furthermore, the interest

370

G.H. Elkaim et al.

is in flying the aircraft on the trajectory, and thus the legs are blended into each other using circular arcs rather than forcing the UAV to overfly each waypoint. Note that this is a design choice, and both variants have a utility depending on the specific mission. The SLUGS guidance algorithm is based on an extension of a simple line-ofsight guidance law originally developed for ground robotics. Much of the art of waypoint guidance consists of determining which leg of the trajectory the UAV is on and when to switch to the next leg of the trajectory. The basic approach is to define a Serret-Frenet frame which points from the current waypoint to the next (TE , along the current leg), with one axis down and E unit the third defined by the cross product between the along track, TE , and down, B, E vectors. The projection of the UAV position onto the T direction is used to determine when to switch to the next leg. In practice this is better than using a distance from the next waypoint to switch (especially in very windy conditions).

18.4.1 General Tracking The outer loop guidance law follows closely the work developed in Amidi and Thorpe (1991), Park et al. (2004), and Park et al. (2007). The original derivation is presented here for completeness. The guidance algorithm steers the velocity vector toward the line of sight; this is one form of pursuit guidance. Making the commanded acceleration proportional to sin  is only one of number of possible guidance laws. The basic guidance algorithm is to determine an aim point and steer the velocity vector toward it. Referring to Fig. 18.16, Vg is the UAV’s ground speed, and C is a circular arc of radius R that originates at the UAV and intercepts the desired path. L1 is a constant look-ahead distance from the UAV to the path in the desired direction of travel. From elementary trigonometry: jL1 j D R sin : 2

(18.32)

Vg C

h

L1 Desired Path

acmd

R

Fig. 18.16 Navigation control law geometry (Reproduction of Fig. 1 in Park et al. 2004)

R h

h

18 Principles of Guidance, Navigation, and Control of UAVs

371

Additionally, from elementary kinematics, it is known that the centripetal acceleration, ac , required for a point mass to follow the circular arc C is given by ac D

jVg j2 : R

(18.33)

Thus the UAV must command a lateral acceleration of ac . Solving Eq. 18.32 for R and substituting it into Eq. 18.33 produces the following control law for commanded acceleration: jVg j2 sin : (18.34) acmd D 2 jL1 j The only requirements for the implementation of this control law are to select the lookahead distance jL1 j and determine sin , the sine of the angle from the velocity vector to L1 .  is sometimes called the line of sight angle. Choice of jL1 j is analogous to feedback gain, with a larger L1 corresponding to smaller gains; sin  is found from the vector cross product of Vg and L1 . sin  D

jVg  L1 j : jVg jjL1 j

(18.35)

For the UAV to actually track the desired trajectory, the lateral acceleration command must be converted to an appropriate bank angle command using the steady-state turn equation: cmd D tan1

acmd : g

(18.36)

18.4.2 Straight Line Tracking The guidance algorithm is most easily described using a straight line (though it is certainly not limited to such). In Fig. 18.17, the UAV is following a straight line segment from waypoint P0 to P1 . A Serret-Frenet coordinate frame (Etkin 2005) is attached to the initial waypoint, P0 , such that the calculations are always in local path coordinates.

Vg Puav N1

Fig. 18.17 Geometry description of the angle  computation

h L1 eN

γ max F

P0

{F } T1

E

Ddt

P1

372

G.H. Elkaim et al.

T D ŒxT ; yT ; zT  D N D

.P1  P0 / ; jP1  P0 j

1 ŒyT xT 0> ; jŒyT xT 0j

(18.37)

B D T  N: Given that eN is the the error in the N direction. Then jeN j D N > .Puav  P0 /;

(18.38)

E is the closest point on the path to the UAV. The intersection of L1 with the desired path is shown by the point F . The down-track distance or EF separation, denoted Ddt , can be computed as Ddt D

p jL1 j2  jeN j2 ;

(18.39)

with the vectors E and F given by E D Puav  jeN jN;

(18.40)

F D E C Ddt T:

(18.41)

From these relations, the vector L1 can then be computed as L1 D F  Puav D Ddt T  jeN jN;

(18.42)

which is then used in Eq. 18.35 to determine  and hence the commanded acceleration. For sufficiently small tracking errors, the L1 guidance looks like a PD controller on lateral error (Park et al. 2007).

18.4.3 LC Control 2 The SLUGS autopilot extends the L1 pursuit guidance to account for some shortcomings of the control law (and for clarity refers to this as LC 2 control). Firstly, it was noticed during flight test experiments that the L1 guidance exhibited large overshoots when turning downwind (hence an increasing Vg ). Analysis showed that in order to solve this, the new L2 vector should be a function of ground speed. Thus, jL2 j D T  jVg j, where T  is a constant and the commanded acceleration becomes acmd D 2

jVg j sin : T

(18.43)

18 Principles of Guidance, Navigation, and Control of UAVs

373 Vg

Fig. 18.18 LC 2 geometry h

∗

MDP |L2| Vg Puav

h |L2| eN

P0

γmax

Pa

P1

Ddtmin

Furthermore, when the lateral error, jeN j, is larger than jL2 j, the aim point cannot be found. To enforce that an aim point is always found, define a maximum intercept angle, max ; the down-track distance to the aim point defined by max is Ddt D

jeN j ; tan max

(18.44)

as shown in the lower aircraft in Fig. 18.18. When jeN j is greater than jL2 j, such as during initial intercept, Ddt may become too large. For example, the aim point may go beyond the next waypoint. To prevent this a bound is placed on the down path distance of the aim point. This bound is set to be a constant, MDP , times jL2 j: Ddtmin D min.Ddt ; jL2 j  MDP /:

(18.45)

This is shown for the upper aircraft in Fig. 18.18. The down-track distance of the aim point is (  Ddt min

D

Ddtmin

p max.Ddtmin ; jL2 j2  jeN j2 /

jeN j > jL2 j jeN j  jL2 j

(18.46)

If the aim point gets beyond P1 , the UAV will continue along this line without ever changing direction. First compute Dwp1 , the along-track distance from the UAV to P1 :

374

G.H. Elkaim et al.

Dwp1 D T > .P1  Puav /:

(18.47)

This will be a negative number if the UAV is beyond P1 . The distance from P1 back along T to the aim point is  Da D Dwp1  Ddt : min

(18.48)

The final aim point Pa is then computed: Pa D T1  max.0; Da /  Pwp1 :

(18.49)

The max operation ensures that the aim point does not extend beyond P1 , i.e., if Da ever becomes negative due to UAV position or failure of the waypoint switching logic, the acceleration command is then computed with Eq. 18.43, where the sin  is obtained from the cross product of Vg and Pa , Eq. 18.35. Note that when the aircraft is initially pointed in the opposite direction from T ,  is greater than 90ı . In this case, a maximum lateral acceleration is imposed to return the vehicle to the correct flight path, where: amax D g tan max

(18.50)

and max is the maximum bank angle permitted (typically set to 30–60ı for a small UAV). In order to use the LC 2 controller in this condition, a maximum allowable  is required:

amax =2  (18.51) ; TLEAD max D min 2 Ucomm where Ucomm is the commanded airspeed from the inner loop controller, and TLEAD is the lead time required for the aircraft to roll out of a steep bank (and determined experimentally for the airframe).

18.4.4 Waypoint Switching As previously stated, the guidance strategy is to move from one segment to the next of the flight mission by connecting the two segments with a circular arc. There are myriad ways to switch between segments; this is simply the one implemented on the SLUGS. The LC 2 implementation uses a policy of early waypoint switching to prioritize path following instead of waypoint precision. Let C be a circle of radius R given by RD

.Uc C jwi nd j/2 : amax

(18.52)

Here Uc is the commanded airspeed, jwi nd j is the windspeed, and amax is the maximum acceleration for the UAV. In the presence of wind the actual radius of

18 Principles of Guidance, Navigation, and Control of UAVs

375

P2

O R Psw2

M R N1

δ

Psw1

Γ P1

P0

T1

Puav

TLEAD · |Vguav|

p

T2

T1 N2

Fig. 18.19 Waypoint switching geometry

curvature of the vehicle over the ground changes as the track angle changes. To avoid this problem in defining switching points, use the maximum possible ground speed in the numerator, thus creating a circle whose radius is larger than any curve over the ground. The circle C is centered at a point O such that it is tangent to each of two waypoint path legs, Psw1 and Psw2 , as shown in Fig. 18.19. The first tangent point, Psw1 , determines the location where the UAV will switch to start tracking the next waypoint segment. Referring to the geometry in Fig. 18.19,  D arccos.T1> T2 /  

2 p D M cos ı ıD

R sin ı R pD : tan ı

M D

(18.53) (18.54) (18.55) (18.56) (18.57)

In flight tests, it was discovered that initiating the turn just at the switch point given by Eq. 18.57 was not adequate because of the lag time of the roll dynamics of the UAV. Therefore, the concept of a lead time, TLEAD , was introduced to initiate the turn. When this lead time is multiplied by the ground speed, it gives the extra distance from the waypoint to the switch point. The new switch point distance is

376

G.H. Elkaim et al.

p D TLEAD  jVg j C

R : tan ı

(18.58)

Finally, the vector position of the switch point Psw1 is given by Psw1 D P1  pT1

(18.59)

Note that during the transition from missions leg, the L2 vector intercepts the circular arc, not the straight line segments. Also, depending on the exact geometry of the waypoints, and the aircraft speed, it may be that just after switching legs, the aircraft is already beyond the next segment switching point. In this case, the logic immediately switches again to the next leg.

18.4.5 Point Acquisition and RTB One advantage of the LC 2 controller is that it is quite robust. For instance, without any change in the logic, the same controller (with its commanded lateral acceleration/bank angle) can just as easily drive the UAV to a point as well as follow an arbitrary curve. For the case of an acquisition point, Pa , the angle  is computed from the current UAV position and Pa : sin  D

jVg  .Pa  Puav /j : jVg jj.Pa  Puav /j

(18.60)

Again with the limits of max and downrange distance imposed to limit bank angle and lateral acceleration, implemented in Eqs. 18.33 and 18.36. This controller will essentially “point at the point” until it overflies the acquisition point, Pa , at which point the switch logic will cause it to circle the point. Of interest is that nowhere in the formulation does the point Pa have to be fixed; the same LC 2 controller can track a moving target using its same logic assuming you have a position estimate of the target. Simulations have shown that for target speeds moving at or below the ground speed of the UAV, the UAV always acquires the target. The initial mission point, PI , before the first waypoint is determined using concept borrowed from instrument flying in which all aircraft must first fly to a well-defined point before proceeding to land. The initial point is determined by projecting a fixed point in front of the first leg of the mission a constant distance in front of the initial waypoint, Fig. 18.20: Vg Puav

h

P0

PI

Fig. 18.20 Initial point geometry

Ip∗· |L2|

P1

18 Principles of Guidance, Navigation, and Control of UAVs

377

PI D Pa D P0  T1  .IP  jL2 j/

(18.61)

where IP is simply a constant tuned to the specific aircraft. Lastly, the LC 2 controller implements a return to base (RTB) functionality by simply recording the original base position, Pb , and using this as the acquisition point if either the mission waypoints are completed or if communication to the UAV fails.

18.4.6 Simulation and Flight Test Results The SLUGS autopilot has both a full simulation (based on MATLAB’s Simulink) and a hardware-in-the-loop (HIL) simulation with the algorithms running on the SLUGS embedded hardware, while the aircraft is simulated through a full 6DOF model running on a separate computer. Figure 18.21 demonstrates the full LC 2 running in simulation, showing the initial point, transitions, and a RTB at the end of the flight. The SLUGS-based UAV, a small electric RC aircraft (Multex Mentor), was flown at UCSC, using both the inner loop control and outer loop guidance described above. Figure 18.22a shows the flight test results (Lizarraga 2009) in the presence of strong winds and real flight disturbances. Panel (b) shows the L2 vector in real time as the aircraft transitions through the waypoints.

Track History

a 600

b

3

300

500

2

2 4

400

3

200

300

5

4

100

200

X(m)

X(m)

Track History

400

100 1

0

1

0

6

−100

−100

−200

−200

−300

−300 −400 −300 −200 −100

0

100 200 300 400 500 Y(m)

−300

−200

−100

0

100

200

300

Y(m)

Fig. 18.21 Return to base implementation in two different waypoint paths. Both have an initial point (green square) and return to the base (red circle). Note the logic skips the waypoint 2 in (a) and waypoint 5 in (b)

378

G.H. Elkaim et al.

a

XY Plote Baseline 50

WP Path UAV Trajectory

0

Y(m)

Fig. 18.22 Flight data for the LC 2 controller on the Mentor platform, showing path tracking performance in (a) and the evolution of the L2 vector at the waypoint switches in (b)

–50 –100 –150 –200 –150 –100 –50

b

0

50 X(m)

100

150

200

250

XY Plot Baseline 50

Y (m)

0

–50

–100 WP Path UAV Position Navigation L2

–150

50

18.5

100

150 200 X (m)

250

300

Conclusion

In this chapter the complete system architecture for two UAV GNC solutions is described. The systems described a low-cost and easily reconfigurable solution which is intended to support research efforts associated with guidance, navigation, and control of small UAVs. The purpose of the description and discussion provided was to show how to implement (hardware and software) a low-cost GNC solution. It was not intended, however, to be the definitive GNC solution for small UAVs. Such solutions do not exist, and this fact was the impetus for writing this chapter. Size, weight, and power constraints associated with small UAVs preclude a “one size fits all” GNC solution. Thus, while the guidance, navigation, and control solutions presented in the chapter were suited for the two UAVs described and the mission

18 Principles of Guidance, Navigation, and Control of UAVs

379

they were intended for (flight control and guidance research), it does not mean they are ideal for application based on other sensors, vehicles, or mission. However, they are flexible enough that they can be adapted to support other missions.

References O. Amidi, C. Thorpe, Integrated mobile robot control, in Proceedings of the SPIE, Boston, vol. 1388, 1991, p. 504 C.-C. Chu, F.A.P. Lie, L. Lemay, D. Gebre-Egziabher, Performance comparison of tight and loose INS-Camera integration, in Proceedings of the 24th International Technical Meeting of the Satellite Division of the Institute of Navigation (ION GNSS 2011), Portland, 2011, p. 3516 A. Dorobantu, A. Murch, B. Mettler, G. Balas, Frequency domain system identification for a small, low-cost, fixed-wing UAV, in AIAA Guidance, Navigation, and Control Conference, Portland, 2011 B. Etkin, Dynamics of Atmospheric Flight (Dover, Mineola, 2005) J. Farrell, M. Barth, The Global Positioning System and Inertial Navigation (McGraw-Hill, New York, 1999) D. Gebre-Egziabher, Design and performance analysis of low-cost aided dead reckoning navigator. PhD thesis, Department of Aeronautics and Astronautics, Stanford University, Stanford S. Gleason, D. Gebre-Egziabher, GNSS Applications and Methods (Artech House, Boston, 2009) P. Groves, Principles of GNSS, Inertial, and Integrated Navigation Systems (Artech House, Boston, 2008) E.J. Lefferts, F.L. Markley, M.D. Shuster, Kalman filtering for spacecraft attitude estimation. J. Guid. Control Dyn. 5(5), 417–429 (1982) F.A.P. Lie, D. Gebre-Egziabher, A synthetic airdata system, in AIAA Guidance, Navigation, and Control Conference, Minneapolis, 2012 M. Lizarraga, Design, implementation and flight verification of a versatile and rapidly reconfigurable UAV GNC research platform. PhD thesis, Department of Computer Engineering, University of California Santa Cruz, Santa Cruz, 2009 M. Lizarraga, V. Dobrokhodov, G.H. Elkaim, R. Curry, I. Kaminer, Simulink based hardwarein-the-loop simulator for rapid prototyping of uav control algorithms, in AIAA Infotech Conference, Seattle, 2009a M. Lizarraga, G.H. Elkaim, G. Horn, R. Curry, V. Dobrokhodov, I. Kaminer, Low cost rapidly reconfigurable uav autopilot for research and development of guidance, navigation and control algorithms, in ASME/IEEE MESA09, San Diego, 2009b. International Conference on Mechatronic and Embedded Systems and Applications M. Lizarraga, R. Curry, G. Elkaim, Reprogrammable uav autopilot system (part 1) – system hardware and software. Circuit Cellar (249), 24–35 (2011a) M. Lizarraga, R. Curry, G. Elkaim, Reprogrammable uav autopilot system (part 2) – testing and results. Circuit Cellar (250), 36–43 (2011b) P. Misra, P. Enge, Global Positioning System, Signals, Measurements, and Performance (GangaJamuna Press, Lincoln, 2001) A. Murch, UMN UAV Flight Code Documentation (2012a), http://http://www.uav.aem.umn.edu/ uav/doxygen/html/index.html A. Murch, University of Minnesota UAV Laboratory (2012b), http://www.uav.aem.umn.edu D. Owens, D. Cox, E. Morelli, Development of a low-cost sub-scale aircraft for flight research: the FASER project, in 25th AIAA Aerodynamic Measurement Technology and Ground Testing Conference, San Francisco, 2006 S. Park, J. Deyst, J.P. How, A new nonlinear guidance logic for trajectory tracking, in AIAA Guidance, Navigation and Control Conference and Exhibit, Portland, 2004

380

G.H. Elkaim et al.

S. Park, J. Deyst, J.P. How, Performance and Lyapunov stability of a nonlinear path-following guidance method. J. Guid. Control Dyn. 30(6), 1718 (2007) D. Simon, Optimal state estimation: Kalman, H1 , and nonlinear approaches (Wiley, Hoboken, 2006) SLUGS website, http://slugsuav.soe.ucsc.edu J. Stewart, Calculus: Early Transcendentals, 4th edn. (Brooks/Cole, Belmont, 2001) D. Titterton, J. Weston, Strapdown Inertial Navigation Technology (Institution of Engineering and Technology, Stevenage, 2004) UAS CoE, Eyes of the army: US army roadmap for unmanned aircraft systems 2010–2035 (2010), http://www.rucker.army.mil/usaace/uas/ University of California Santa Cruz Autonomous Systems Lab, http://asl.soe.ucsc.edu

Section IV Sensors and Sensing Strategies Ben Upcroft and Salah Sukkarieh

Sensors and Sensing Strategies: Introduction

19

Kimon P. Valavanis and George J. Vachtsevanos

Sensors and Sensing Strategies enable an unmanned aircraft to “sense,” “see,” “hear,” and “understand” the world around it so that it may function intelligently in an unknown and cluttered environment and in the absence of an onboard pilot. In essence, sensors and sensing strategies are crucial since they provide the technologies that will result in “unmanned aircraft operating as if there were a human pilot onboard.” Sensors for Missions by Mejias, Lai, and Bruggemann sets the tone for sensors used on UAVs that are assigned complex missions. The sensor suite onboard a UAV is tightly coupled with payload capabilities, as payload dictates UAV usability and market value. However, advances in miniaturization of electronics are enabling replacement of multiprocessing, power-hungry general-purpose processors with more integrated and compact electronics that contribute to more onboard sensors. Several common payload sensors are described along with their usefulness to solve real-world problems. Inertial Sensor-Based Simultaneous Localisation and Mapping for UAVs by Bryson and Sukkarieh provides an overview of algorithms for inertial sensor-based simultaneous localization and mapping (SLAM) within the context of UAVs, using the extended Kalman filter (EKF) and the extended information filter (EIF) due to their ease of understanding, applicability to online implementation, and prevalence in airborne localization applications outside of SLAM.

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 135, © Springer Science+Business Media Dordrecht 2015

383

384

K.P. Valavanis and G.J. Vachtsevanos

UAV Localisation Using Inertial Sensors and Satellite Positioning Systems by Bryson and Sukkarieh provides an overview of UAV localization with a focus on aided inertial localization, that is, algorithms for fusing data from, for example, satellite positioning systems, barometric sensors, and magnetometers with inertial sensors to provide real-time position and orientation. An example implementation of aided inertial localization on a UAV is presented as a tutorial to understand key concepts in airborne localization and as a basic guide toward more complicated implementations. Data Fusion and Tracking with Multiple UAVs by Ridley, Upcroft, and Sukkarieh describes decentralized data fusion (DDF) algorithms for a team of multiple autonomous platforms. It is shown how through the DDF algorithms each platform can maintain a consistent global solution from which decisions may be made. The overall system design is detailed, providing insight into the overall complexity of implementing a robust DDF system for use in information-gathering tasks in outdoor UAV applications. Collectively, after reading and understanding the first four sections of the handbook, the reader, novice, or expert will be ready to continue with the actual control of UAVs and all other more advanced technical aspects.

Sensors for Missions

20

Luis Mejias, John Lai, and Troy Bruggemann

Contents 20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.2 Navigation Sensors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.2.1 Electro-Optical (EO) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.2.2 Radio-Wave Sensors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.3.1 Collision Avoidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.3.2 Remote Sensing: Power Line Inspection and Vegetation Management . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

386 386 387 389 390 390 394 398

Abstract

An onboard payload may be seen in most instances as the “Raison d’Etre” for a UAV. It will define its capabilities, usability and hence market value. Large and medium UAV payloads exhibit significant differences in size and computing capability when compared with small UAVs. The latter has stringent size, weight, and power requirements, typically referred as SWaP, while the former still exhibit endless appetite for compute capability. The tendency for this type of UAVs (Global Hawk, Hunter, Fire Scout, etc.) is to increase payload density and hence processing capability. An example of this approach is the Northrop Grumman MQ-8 Fire Scout helicopter, which has a modular payload architecture that incorporates off-the-shelf components. Regardless of the UAV size and capabilities, advances in miniaturization of electronics are enabling the replacement of multiprocessing, power-hungry general-purpose processors with more integrated and compact electronics (e.g., FPGAs).

L. Mejias () • J. Lai • T. Bruggemann Australian Research Centre for Aerospace Automation, Queensland University of Technology, Brisbane, QLD, Australia e-mail: [email protected]; [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 6, © Springer Science+Business Media Dordrecht 2015

385

386

L. Mejias et al.

The payload plays a significant role in the quality of ISR (intelligent, surveillance, and reconnaissance) data, and also in how quickly that information can be delivered to the end user. At a high level, payloads are important enablers of greater mission autonomy, which is the ultimate aim in every UAV. This section describes common payload sensors and introduces two cases in which onboard payloads were used to solve real-world problems. A collision avoidance payload based on electro optical (EO) sensors is first introduced, followed by a remote sensing application for power line inspection and vegetation management.

20.1

Introduction

There are two main categories of payloads onboard a UAV: those that allow the vehicle to navigate and those that allow the vehicle to perform its main task. The distinction between them is sometimes very subtle. In some cases, a payload aimed to perform a task can be used for navigation purposes (e.g., cameras), and navigation payloads may in some instances be integrated with other sensors to perform a specific task (e.g., GPS/INS with LIDAR). This section will describe common sensors that are used in many typical payloads onboard UAVs.

20.2

Navigation Sensors

At the core of most UAV guidance and navigation systems, one can find Global Navigation Satellite Systems (GNSS) (including the Global Positioning System – GPS) and Inertial Navigation Systems (INS). Their complementary nature has been recognized, and as a result, GPS and INS sensors are the preferred sensor couple for the majority of autopilot systems. The integration of GPS and INS is without doubt the area in which researchers have spent considerable efforts proposing approaches such as uncoupled integration, loosely coupled integration, tightly coupled integration, and deeply coupled integration (Grewal et al. 2007). GPS and INS are not the only two sensors used for navigation. They can be complemented with altimeters (laser-based, barometric, etc.) to enhance the estimation of the vehicle state. Additionally, infrared attitude sensors are typically found in micro UAVs. Recently in Cao et al. (2010), a survey of UAV autopilot alternatives and typical sensor combinations was presented. At the heart of the integration scheme lies a estimator (usually a form of Kalman filter) which estimates position, velocity, attitude, GPS errors and inertial sensor errors. Due to their complementary nature, GPS and INS are often the preferred core sensor suite. However, researchers have also investigated the integration of other sensor combinations, such as GPS with computer vision (Dusha et al. 2011; Wein et al. 2011; Dusha and Mejias 2012) and INS with computer vision (Merz et al. 2006). Additionally, factors such as the trade-off between cost and accuracy

20 Sensors for Missions

387

of INS sensors and the susceptibility of GPS to spoofing and jamming have also contributed to increased interest in alternative sensor combinations. While alternative integration schemes using other sensors is an attractive option, the low cost and future navigation availability and integrity that Space-Based Augmentation Systems (SBAS) such as WAAS, EGNOS, GAGAN, and MSAS will provide cannot be ignored. Submeter accuracy for civilian users will also be possible with the commissioning of Galileo, Compass, and the modernization of GLONASS. This in turn will encourage interoperatibility and the possibility of a triple-frequency civilian-band GPS over the next decades.

20.2.1 Electro-Optical (EO) Nowadays, it is difficult to realize a UAV without an EO sensor. They have become standard fit-out onboard many aerial vehicles. The challenge is now in the processing and interpretation of the information acquired by EO sensors. Perception through EO sensors can be seen as one of the most important tasks in a UAV. Whether it is performed for navigation or for surveillance (as an end application), it defines the necessary peripherals to process the EO data. This section will introduce some of the most common EO sensors typically found in UAVs.

20.2.1.1 Visible Spectrum A visible-spectrum camera operates about 390 nm (3.9 m)–750 nm (7.5 m) wavelength. These cameras can be found in two main categories: digital still cameras and machine vision cameras (including surveillance and webcam). Digital still cameras offer very high resolution, but cannot provide a continuous stream of images. The amount of images they can provide usually depends on the amount of internal memory. This type of camera sees application in remote sensing and aerial photography. Machine vision cameras have relatively lower resolution but can provide a continuous stream of images up to a few hundred frames per second. The speed is related with the resolution and output format used (digital or analog). They are suitable for processes or tasks that require very fast perception of environment. Common output protocols and interfaces for machine vision cameras include IEEE 1394, Camera Link, SD/HD Analog, USB, GigE Vision, and in coming years Thunderbolt cameras. They can provide color or gray level (or both) images. The data representation or color space usually varies from one manufacturer to another. Typical color spaces or models used in most machine vision cameras are RGB, YUV, YPbPr, and YCbCr, etc. Regardless of its end use and camera type, camera geometric models are necessary. They provide parameters that are needed to correct lens distortions, perform the mapping or representation of 3D objects onto the 2D surface called image, and overall allow manipulation of the data acquired by these devices. These parameters are normally estimated through a calibration process that is regularly

388

L. Mejias et al.

performed. For more details on camera models and calibration theory, refer to Forsyth and Ponce (2002) and Szeliski (2011).

20.2.1.2 Infrared An infrared (IR) camera is a device that detects and converts light in the same way as common visible-spectrum cameras, but is sensitive to light at longer wavelengths. They form an image using infrared radiation in the spectrum at wavelengths from 5,000 nm (5 m) to 14,000 nm (14 m). Infrared cameras are used to convey a measure of thermal radiation of bodies. The intensity of each pixel can be converted for use in temperature measurement, where the brightest parts of the image (colored white) represent the warmest temperatures. Intermediate temperatures are shown as reds and yellows, and the coolest parts are shown in blue. Often, IR images are accompanied by a scale next to a false color image to relate colors to temperatures. The resolution of IR cameras (up to a maximum of 640 480 pixels) is considerably lower than the resolution of optical cameras. Furthermore, the price of IR cameras is considerably higher than their visible-spectrum counterparts. Infrared cameras can be categorized in two main groups: Cooled Infrared Detectors Cooled IR detectors are normally inside a vacuum-sealed container that is cryogenically cooled. Cooling is needed for the efficient operation of the semiconductor used. Typical operating temperatures (in Kelvin degrees (K)) range from 4 to 293 K. Most modern IR detectors operate in 60–100 K range. The operating temperature is related with the type of technology used and performance expected. Cooled infrared cameras provide superior image quality compared to uncooled ones. They provide greater sensitivity that allows the use of higher F-number lenses, making high-performance long focal length lenses both smaller, and cheaper for cooled detectors. The drawbacks of cooled infrared cameras are that they are expensive both to produce and to run. Cooling is power hungry and time consuming. A camera may need several minutes to cool down before it can begin working. Uncooled Infrared Detectors This type of camera use sensors that are at or close to room temperature. For this reason, they do not require bulky, expensive cryogenic coolers, and therefore are smaller and cheaper than cooled ones, but with the drawback that their resolution and image quality tend to be lower than cooled detectors. This is due to a difference in their fabrication processes, limited by currently available technology. Given the ability of IR cameras to reveal the hidden world not perceived by the human eye or visible-spectrum camera, they see application in tasks that are visibly challenging such as night vision or bad weather scenarios, inspection tasks, surveillance (bushfire monitoring), search, and rescue.

20.2.1.3 Hyperspectral Imaging Sensors in this category acquire image data simultaneously in multiple adjacent spectral bands. This gives a wealth of data, but its processing and interpretation

20 Sensors for Missions

389

require good knowledge of what specific properties are to be measured and how they relate to the actual measurement made from the sensor. For example, a single cell position in an image will have a set of brightness (or intensity) levels for each wavelength (spectral band). Different materials under examination by a hyperspectral sensor will often exhibit different intensity versus wavelength relationships. Hence, with some prior knowledge, hyperspectral image data can be useful for identifying the type or composition of materials. Cost and complexity are the two major drawbacks of this technology. Storage is another limiting factor given the multidimensional nature of hyperspectral datasets. The availability of graphics processing units (GPU) as powerful parallel processing hardware could bring this technology a step closer to widespread adoption. However, further research efforts are needed to create analytical techniques and algorithms to unleash the full potential of hyperspectral imaging.

20.2.2 Radio-Wave Sensors 20.2.2.1 Airborne Radio Detection and Ranging (Radar) Radar is a radio system used to determine range, altitude, direction, or speed of objects. The system transmits controlled radio pulses which are reflected back by objects. The distance to objects is estimated by measuring the signal return time. The received power declines as the fourth power of the range, assuming transmitting and receiving antennas are in the same location, hence the need for high transmitting power in most cases. Speed can be estimated by tracking the change in distance with time or by exploiting the Doppler effect (Stimson 1998). Airborne radar has been in operation since WWII, and can be considered as an integral part of systems such as Ground Proximity Warning Systems (GPWS) or TCAS. In the automotive industry, radar is starting to appear in the form of collision warning systems (www.ford.com). In a UAV context, the main drawback of radar is the high power consumption. The size, weight, and power (SWaP) challenges that are faced by UAVs are well known. However, new systems such as Synthetic-Aperture Radar (SAR) (Soumekh 1999) are beginning to make radar technology a feasible option onboard UAVs (Hanlon 2008).

20.2.2.2 Light Detection and Ranging (LIDAR) LIDAR operates in a similar manner to radar, in that it can estimate distance by measuring the time of return of a signal reflected from an object. LIDARs have been widely used in atmospheric and meteorology research (Couch et al. 1991; Kiemle et al. 1997) and remote sensing (Dubayah and Drake 2000; Lefsky et al. 2002). Given its ability to provide very high definition (under 2.5 cm (Terranean 2011)), LIDAR has become a common sensor for mapping and infrastructure inspection (Yuee et al. 2009).

390

L. Mejias et al.

However, LIDAR shares many of the SWaP obstacles that are faced by radar technology. Hence, LIDARs are not often found in micro/small size UAVs, but have been flown in light general aviation aircraft (Ara 2011). Recent advances in technology have been made, an example of which is the Riegl LMS-Q160 LIDAR in 0, Rww > 0, all plant dynamics are constant in time, ŒA; Cy  is detectable, and ŒA; Bw  is stabilizable. A system is said to be stabilizable if all unstable modes are controllable and is detectable if all unstable modes are observable. In this case, the forward propagation of the covariance Qe .t/ settles down to a constant Qss independent of Qe .0/, as t ! 1 where 1 AQss C Qss AT C Bw Rww BwT  Qss CyT Rvv Cy Qss D 0 1 . Note that the stabilizable/detectable assumptions give a and Lss D Qss CyT Rvv unique Qss 0. Also, if Qss exists, the steady-state filter

PO x.t/ D .A  Lss Cy /x.t/ O C Lss y.t/

(27.89)

is asymptotically stable given these assumption, and thus, Lss can be used as the estimator gains Le for the system.

27.3.4.3 Dynamic Output Feedback Control System Given a closed-loop estimator with gain Le , PO x.t/ D .A  Le C /x.t/ O C Bu.t/ C Le y.t/:

(27.90)

Then if Le is chosen as outlined in the previous sections, it is known that x.t/ O ! x.t/ as t ! 1, but in general, x.t/ O ¤ x.t/ 8t. Even so, the approach taken in the output feedback control case is to implement u.t/ D K x.t/, O leading to the closed-loop system dynamics 

 " A x.t/ P D PO x.t/ Le C

BK A  BK  Le C

#

x.t/ x.t/ O

 (27.91)

or xP cl .t/ D Acl xcl .t/:

(27.92)

The stability of this closed-loop system can be analyzed using the similarity   I 0 that preserves the location of the eigenvalues. transformation matrix T D I I In fact it is easy to show that the dynamics matrix of the transformed system is given by " # A  BK BK 1 N Acl , TAcl T D : 0 A  Le C Thus, the closed-loop pole locations are given by the roots of the polynomial det.sI  ANcl / , det.sI  .A  BK//  det.sI  .A  Le C // D 0

566

J.P. How et al.

which consist of the union of the regulator poles and estimator poles. The significance of this result is that it implies that one can design the estimator and regulator separately and combine them at the end, and the closed-loop system will be stable. This is called the separation principle, and when it holds, it greatly simplifies the control design process. The resulting dynamic output feedback compensator is the combination of the regulator and closed-loop estimator written using the new state xc .t/  x.t/ O as xP c .t/ D Ac xc .t/ C Bc y.t/

(27.93)

u.t/ D Cc xc .t/;

(27.94)

where the compensator dynamics are Ac , A  BK  Le C , Bc , Le , and Cc , K. Note that, as expected, the compensator maps sensor measurements to actuator commands. The DOFB solution provided in this section is called the (infinite horizon) linear quadratic Gaussian (LQG), and it is the controller that optimizes the expected timeaveraged value of the cost in Eq. (27.76) for the stochastic system in Eqs. (27.86) and (27.87) (Bryson and Ho 1969; Stengel 1994; Franklin et al. 1994; Skogestad and Postlethwaite 2005). While LQG has proven to be very effective design algorithm for many MIMO systems, it does not have the same robustness properties guaranteed for the LQR design (Doyle 1978). Furthermore, LQG designs can exhibit high sensitivity to modeling errors. Thus, it is prudent to be cautious and check the peak magnitude of the sensitivity S.s/, which provides a measure of the closeness of the loop transfer function to the critical point at s D 1. High sensitivities (above 20–50) are indicative of designs that could be susceptible to modeling errors, so further testing and simulation using perturbed dynamics should be considered.

27.3.4.4 Robust Control Techniques The alternative approach to control design mentioned in Sect. 27.3.1 works with the S.s/ and T .s/ functions to shape them directly (McFarlane and Glover 1992). For example, consider the following design problem in Fig. 27.14 where G.s/ D

Ws(s)

r

e

Gc(s)

−

Fig. 27.14 Direct design problem

200 : .0:05s C 1/2 .10s C 1/

z1

Wu(s)

u

G(s)

z2

y˜

27 Linear Flight Control Techniques for Unmanned Aerial Vehicles

567

Note that there is one input (r) and two performance outputs – one that penalizes the sensitivity S.s/ of the system and the other that penalizes the control effort used with the weight functions Ws .s/ and Wu .s/ to be defined later. With z1 D Ws .s/.r  Gu/

(27.95)

z2 D Wu .s/u

(27.96)

e D r  Gu;

(27.97)

then closing the loop with u D Gc .s/e yields    Ws G Ws C Gc .I C GGc /1 0 Wu     Ws  Ws GGc S Ws S D D Wu Gc S Wu Gc S 

PCL .s/ D

so that



z1 z2



 D

 Ws S r; Wu Gc S

(27.98) (27.99)

(27.100)

and thus, the closed-loop transfer function from the reference input to the performance variable z1 is the weighted sensitivity function. This setup gives us direct access to the sensitivity function in the analysis. It is now possible to embed tests such as whether jS.s/j satisfies a desired bound such as jS.j!/j
0, then

d V1 .x; e/ D W .x/  k1 e 2 < 0; dt

8.x; e/ ¤ 0;

thus proving stability of system Eq. 28.15. Finally, assuming that g1 .x; z/ ¤ 0, 8.x; z/ 2 RnC1 , a stabilizing feedback for the original system in Eq. 28.14 can be recovered as 1 u.x; z/ D g1 .x; z/ D

1 g1 .x; z/

 

@u0 .x/ .f0 .x/ C g0 .x/z/  f1 .x; z/ C v @x



@u0 .x/ .f0 .x/ C g0 .x/z/ @x

 @V0 .x/ g0 .x/  k1 .z  u0 .x// :  f1 .x; z/  @x

(28.16)

Since the control law in Eq. 28.16 stabilizes system Eq. 28.14, with Lyapunov function V1 , a similar argument can be used to recursively build a control law for a system in strict-feedback form of arbitrary order, xP D f0 .x/ C g0 .x/ z1 ; zP1 D f1 .x; z1 / C g1 .x; z1 / z2 ; :: :

(28.17)

zPm D fm .x; z1 ; : : : ; zm / C g1 .x; zm / u: The procedure is summarized as follows: Start from the “inner” system, for which a stabilizing control law and a Lyapunov function are known. Define an error between the known control law and the actual input to the inner system. Augment the Lyapunov function with the square of the error. Design a new control law that

588

G.V. Chowdhary et al.

ensures the stability of the augmented system. Repeat using the augmented system just defined as the new inner system. This recursion, “stepping back” from the inner system all the way to the control input, gives the name to the method. A detailed description of control synthesis for small unmanned helicopters using backstepping techniques is provided in Raptis and Valavanis (2011).

28.3.3 Model Predictive Control (MPC) MPC has been successfully used for many years in industrial applications with relatively slow dynamics (e.g., chemical reactions (Maciejowski 2002; Kerrigan 2000; Mayne et al. 2000)); it is only in the past decade that the computational power has been available to enable online optimization for fast system dynamics typical in aerospace applications (early relevant demonstrations are in Manikonda et al. (1999), Dunbar and Murray (2002), Dunbar and Murray (2004), Dunbar (2007), Richards and How (2004), and Richards (2005)). MPC attempts to solve the following problem: Design an admissible piecewise continuous control input u.t/ that guarantees the system behaves like a reference model without violating the given state and input constraints. As such, a key benefit MPC is the ability to optimize the control input in the presence of state and input constraints. Furthermore, because MPC explicitly considers the operating constraints, it can operate closer to hard constraint boundaries than traditional control schemes. MPC can be formulated using several types of cost function; however, due to the availability of robust quadratic solvers, the following formulation is popular: Find the optimal control u.t/ such that for given positive (semi)-definite matrices Q, R, and S , the following quadratic cost is minimized: Z J1 .e; u/ D

1 t0

e T .t/Qe.t/ C uT .t/Ru.t/dt; x 2 „; u 2 …:

(28.18)

In the presence of constraints, a closed form solution for an infinite horizon optimization problem cannot be found in the general case (Camacho and Bordons 1999; Rawlings 2000; Nicolao et al. 2000; Mayne 2000). Hence, the approach in MPC is to numerically solve a receding horizon optimization (RHO) problem online over the interval Œt; t C N  to find the new control input at time t C 1; this process is then repeated over every discrete update (see, Fig. 28.3). The idea is that if the horizon is sufficiently large, the solution to the receding horizon optimization problem can guarantee stability. Let h be a mapping between the states x of the nonlinear dynamical system in Eq. 28.2 and the output z such that z D h.x/. If the function h.x/ is observable, this problem can be recast into a problem of minimizing a discrete output-based cost. Observability for linear systems was discussed in chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book. Local observability for nonlinear systems can be established in an analogous manner by considering the rank of the Jacobian matrix of n  1 Lie derivatives of h.x/ along the trajectories of Eq. 28.2 (see, e.g., Kou et al. 1973; Nijmeijer and van der Schaft 1990). Typically, a quadratic cost function is preferred to leverage

28 Nonlinear Flight Control Techniques for Unmanned Aerial Vehicles Fig. 28.3 Schematic of a model predictive controller which minimizes a quadratic cost over a finite horizon

u

x

UAV Dynamics: f

MPC

589

Candidate control input

Receding Horizon Optimizer

Model Predictor

Predicted states

External Commands

existing results in quadratic programming (Camacho and Bordons 1999; Rawlings 2000; Demircioglu and Karasu 2000). Let zrmk denote the sampled output of the reference model, and define the output error eNk D zk  zrmk . Then the problem can be reformulated to finding the optimal sequence of inputs uk such that the following quadratic cost function is minimized for given positive (semi)-definite N subject to the reformulated output constraint y 2 „, N where matrices QN and R, N D fy W y D z .x.t/; u.t//; x 2 „g and the input constraint u 2 …: „ JT .e; N u/ D

tX CN

N k C Vf .et CN /; y 2 „; N u 2 …; eNkT QN eNk C uTk Ru

(28.19)

kDt

where the term Vf .et CN / denotes a terminal penalty cost. Several computationally efficient nonlinear MPC algorithms have been proposed and their stability properties established (Zheng and Allgower 1998; Zheng 2000). To numerically solve the RHO problem in Eq. 28.19, a prediction model xOP D O f .x/ O is required to predict how the system states behave in the future, where xO is the estimated state and fO is the prediction model. The prediction model is a central part of MPC, and in many cases (especially when the system dynamics are nonlinear or unstable), an inaccurate prediction model can result in instability (Rawlings 2000). In many MPC implementations, one of the most costly effort in control design is to develop a reliable prediction model (Camacho and Bordons 1999; Rawlings 2000; Lee 2000). Furthermore, approximations made in modeling, changes in the system dynamics due to wear and tear, reconfiguration of the system, or uncertainties introduced due to external effects can affect the accuracy of the predicted system response. Robustness of MPC methods to estimation errors (Findeisen et al. 2003; Michalska and Mayne 1995; Lee et al. 2002), plant variability (Chisci et al. 2001; Lee and Kouvaritakis 2000; Cuzzola et al. 2002; Richards 2005), and disturbances (Lee and Kouvaritakis 1999; Bemporad 1998; Kerrigan and Maciejowski 2001; Kerrigan and Mayne 2002; Richards and How 2007) remain active research areas. The resulting optimization problems are typically solved using linear matrix inequalities, linear programming, or quadratic programming. The key challenge here is to provide sufficient robustness guarantees while keeping the problem computationally tractable. The fact remains that without an accurate prediction model, the performance and stability guarantees remain very conservative (Rawlings 2000; Nicolao et al. 2000).

590

G.V. Chowdhary et al.

To this effect, some authors have recently explored adaptive MPC methods that estimate the modeling uncertainty (Camacho and Bordons 1999; Fukushima et al. 2007; Adetola et al. 2009; Karra et al. 2008). Another active area of research in MPC is that of analytically guaranteeing stability. It should be noted that stability and robustness guarantees for MPC when the system dynamics are linear are at least partially in place (Camacho and Bordons 1999). With an appropriate choice of Vf .et CN / and with terminal inequality state and input constraints (and in some cases without), stability guarantees for nonlinear MPC problems have been established (Nicolao et al. 2000; Zhang et al. 2010; Marruedo et al. 2002). However, due to the open loop nature of the optimization strategy, and dependence on the prediction model, guaranteeing stability and performance for a wide class of nonlinear systems still remains an active area of research. A challenge in implementing MPC methods is to ensure that the optimization problem can be solved in real time. While several computationally efficient nonlinear MPC strategies have been devised, ensuring that a feasible solution is obtained in face of processing and memory constraints remains an active challenge for MPC application on UAVs, where computational resources are often constrained (Sutton and Bitmead 2000).

28.3.4 Model Inversion-Based Control 28.3.4.1 Dynamic Model Inversion Using Differentially Flat Representation of Aircraft Dynamics A complete derivation of the nonlinear six-degree-of-freedom UAV dynamics was presented in chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book. For trajectory tracking control using feedback linearization, simpler representations of UAV dynamics are often useful. One such representation is the differentially flat representation (see, e.g., Hauser and Hindman 1997). The dynamical system in Eq. 28.1 with output y D h.x; u/ is said to be differentially flat with flat output z if there exists a function g such that the state and input trajectories can be represented as a function of the flat output and a finite number of its derivatives:   d ny .x; u/ D g y; y; P y; R :::; n : (28.20) dt In the following, assume that a smooth reference trajectory pd is given for a UAV in the inertial frame (see chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book) and that the reference velocity pPd and reference acceleration pRd can be calculated using the reference trajectory. Consider the case of a conventional fixed wing UAV that must be commanded a forward speed to maintain lift; hence, pPd ¤ 0. A right-handed orthonormal frame of reference wind axes can now be defined by requiring that the desired velocity vector is aligned with the xW axis of the wind axis (see chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book for details), and the lift and drag forces are in the xW –zW plane of the wind axis, the zW axis, such that there are no side forces. The acceleration can be written as

28 Nonlinear Flight Control Techniques for Unmanned Aerial Vehicles

pRd D g C fI =m;

591

(28.21)

where fI is the sum of propulsive and aerodynamic forces on the aircraft in the inertial frame. Let RI W denote the rotation matrix that transports vectors from the defined wind reference frame to the inertial frame. Then pRd D g C RI W aW ;

(28.22)

where aW is the acceleration in the wind frame. Let ! D Œ!1 ; !2 ; !3 T denote the angular velocity in the wind frame, then the above equation can be differentiated to obtain d 3 pd D RI W .!  aW / C RI W aP W ; (28.23) dt 3 by using the relationship RP I W aW D RI W !a O W D RI W .!  aW / (see Sect. 27.2.2 of chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book). Let at denote the tangential acceleration along the xW direction, an denote the normal acceleration along the zW direction, and V denote the forward speed. Furthermore, let e1 D Œ1; 0; 0T . Then in coordinated flight pPd D VRI W e1 , hence pRd D VP RI W e1 C VRI W .!  e1 /:

(28.24)

Combining Eqs. 28.22 and 28.24, the following relationship can be formed: 2 pRd D RI W

3 VP 4 V !3 5 : V !2

(28.25)

Equation 28.25 allows the desired !2 and !3 to be calculated from the desired acceleration pRd . Furthermore, from Eq. 28.23, 2

3 2 3 2 aP t !2 an 1 0 4 !1 5 D 4 !3 at =n 5 C 4 0 1=an aP n !2 at 0 0

3 0 d 3 pd : 0 5 RITW dt 3 1

(28.26)

The above equation defines a differentially flat system of aircraft dynamics with flat output pd and inputs ŒaP t ; !1 ; aP n , if V D kpPd k ¤ 0 (nonzero forward speed) and an ¤ 0 (nonzero normal acceleration). The desired tangential and normal accelerations required to track the path pd can be controlled through the thrust T .ıT / which is a function of the throttle input ıT , the lift L.˛/, and the drag D.˛/ which are functions of the angle of attack ˛ by noting that in the wind axes: at D T .ıT / cos ˛  D.˛/; an D T .ıT / sin ˛  L.˛/:

(28.27)

592

G.V. Chowdhary et al.

To counter any external disturbances that cause trajectory deviation, a feedback term can be added. Let p be the actual position of the aircraft and u be the control input, 3 and consider a system in which ddtp3d D u: 2 3 2 0 1 p d 4 5 4 pP D 0 0 dt 0 0 pR

32 3 2 3 0 p 0 1 5 4 pP 5 C 4 0 5 u: 1 pR 0

(28.28)

Defining e D p  pd , the above equation can be written in terms of the error: 2 3 2 0 e d 4 5 4 eP D 0 dt 0 eR

32 3 2 3  0  e 1 0 d 3 pd 5 4 5 5 4 u  : eP C 0 0 1 dt 3 1 eR 0 0

(28.29)

 T 3 Therefore, letting u D ddtp3  K e eP eR , where K is the stabilizing gain, and computing at ; an ; !1 from .p; p; P p; R u/ guarantee asymptotic closed-loop stability of the system (see Hauser and Hindman 1997 for further detail of the particular approach presented). This approach is essentially that of feedback linearization and dynamic model inversion, which is explored in the general setting in the next section.

28.3.4.2 Approximate Dynamic Model Inversion The idea in approximate dynamic inversion-based controllers is to use a (approximate) dynamic model of the UAV to assign control inputs based on desired angular rates and accelerations. Let x.t/ D Œx1T .t/; x2T .t/T 2 Rn be the known state vector, with x1 .t/ 2 Rn2 and x2 .t/ 2 Rn2 : let u.t/ 2 Rn2 denote the control input: and consider the following multiple-input nonlinear uncertain dynamical system: xP 1 .t/ D x2 .t/;

(28.30)

xP 2 .t/ D f .x.t/; u.t//; where the function f is assumed to be known and globally Lipschitz continuous, and control input u is assumed to be bounded and piecewise continuous. These conditions are required to ensure the existence and uniqueness of the solution to Eq. 28.30. Furthermore, a condition on controllability of f with respect to u must also be assumed. Note also the requirement on as many control inputs as the number of states directly affected by the input (x2 ). For UAV control problem, this assumption can usually be meet through the successive loop closure approach (see chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book). For example, for fixed wing control aileron, elevator, rudder, and throttle control directly affect roll, pitch, yaw rate, and velocity. This assumption can also meet for rotorcraft UAV velocity control with the attitudes acting as virtual inputs

28 Nonlinear Flight Control Techniques for Unmanned Aerial Vehicles

593

for velocity dynamics and the three velocities acting as virtual inputs for the position dynamics (Johnson and Kannan 2005). In dynamic model inversion-based control, the goal is to find the desired acceleration, referred to as the pseudo-control input .t/ 2 Rn2 , which can be used to find the control input u such that the system states track the output of a reference model. Let z D .x; u/, if the exact system model f .z/ in Eq. 28.30 is invertible; for a given .t/, u.t/ can be found by inverting the system dynamics. However, since the exact system model is usually not invertible, let  be the output of an approximate inversion model fO such that  D fO.x; u/ is continuous and invertible with respect to u, that is, the operator fO1 W RnCn2 ! Rl exists and assigns for every unique element of RnCn2 and a unique element of Rl . An approximate inversion model that satisfies this requirement is required to guarantee that given a desired pseudocontrol input  2 Rn2 a control command u can be found by dynamic inversion as follows: u D fO1 .x; /: (28.31) The model in Sect. 28.3.4.1 is an example of a differentially flat approximate inversion model. For the general system in Eq. 28.30, the use of an approximate inversion model results in a model error of the form xP 2 D  C .x; u/;

(28.32)

where  is the modeling error. The modeling error captures the difference between the system dynamics and the approximate inversion model: .z/ D f .z/  fO.z/:

(28.33)

Note that if the control assignment function were known and invertible with respect to u, then an inversion model can be chosen such that the modeling error is only a function of the state x. Often, UAV dynamics can be represented by models that are affine in control. In this case the existence of the approximate inversion model can be related directly to the invertibility of the control effectiveness matrix B (e.g., see Johnson 2000). For example, let G 2 Rn2 n2 , and let B 2 Rn2 l denote the control assignment matrix and consider the following system: xP 1 .t/ D x2 .t/;

(28.34)

xP 2 .t/ D Gx2 .t/ C B.‚.x/ C u.t//; where ‚.x/ is a nonlinear function. If B T B is invertible, and the pair .G; B/ is controllable, one approximate inversion model is .t/ D Bu.t/, which results in a unique u for a unique : u.t/ D .B T B/1 B T .t/. Adding and subtracting  D Bu yields Eq. 28.32, with .x/ D Gx2 C B.‚.x/ C u/  Bu D Gx2 C B‚.x/.

594

G.V. Chowdhary et al.

A reference model is used to characterize the desired response of the system: xP 1rm D x2rm ;

(28.35)

xP 2rm D frm .xrm ; r/; where frm .xrm .t/; r.t// denote the reference model dynamics which are assumed to be continuously differentiable in xrm for all xrm 2 Dx  Rn . The command r.t/ is assumed to be bounded and piecewise continuous; furthermore, frm is assumed to be such that xrm is bounded for a bounded reference input. The pseudo-control input  is designed by combining a linear feedback part pd D ŒK1 ; K2 e with K1 2 Rn2 n2 and K2 2 Rn2 n2 , a linear feedforward part rm D xP 2rm , and an approximate feedback linearizing part ai .z/:  D rm C pd  ai :

(28.36)

Defining the tracking error e as e.t/ D xrm .t/  x.t/ and using Eq. 28.32, the tracking error dynamics can be written as  x2 : (28.37) eP D xP rm  C 

0 I1 Letting A D , B D Œ0; I2 T , where 0 2 Rn2 n2 , I1 2 Rn2 n2 , and K1 K2 I2 2 Rn2 n2 are the zero and identity matrices, and using Eq. 28.36 gives the following tracking error dynamics that are linear in e: eP D Ae C BŒai .z/  .z/:

(28.38)

The baseline linear full-state feedback controller pd should be chosen such that A is a Hurwitz matrix. Furthermore, letting ai D .z/ using Eq. 28.33 ensures that the above tracking error dynamics is exponentially stable, and the states of the UAV track the reference model. This framework is depicted in (Fig. 28.4).

28.3.5 Model Reference Adaptive Control The model reference adaptive (MRAC) control architecture has been widely studied for UAV control in presence of nonlinearities and modeling uncertainties (see, e.g., Johnson and Kannan 2005; Lavertsky and Wise 2005; Nguyen et al. 2006b; Patel et al. 2009). MRAC attempts to ensure that the controlled states track the output of an appropriately chosen reference model (see, e.g., Narendra and Annaswamy 1989; Ioannou and Sun 1996; Astr¨om and Wittenmark 1995; Tao 2003). Most MRAC methods achieve this by using a parameterized model of the uncertainty, often referred to as the adaptive element and its parameters referred to as adaptive weights. Aircraft dynamics can often be separated into a linear part whose mathematical

28 Nonlinear Flight Control Techniques for Unmanned Aerial Vehicles

595

xrm

r(t)

Reference Model

νrm

ν

(Approximate) Inversion model fˆ−1

νai

νpd

u

UAV dynamics f

x

− e

Feedback linearizer PD compensator

Fig. 28.4 Approximate dynamic model inversion framework using feedback linearization

model is fairly well known and an uncertain part that may contain unmodeled linear or nonlinear effects. This representation is also helpful in representing nonlinear external disturbances affecting the system dynamics. Therefore, one typical technique for implementing adaptive controllers is to augment a baseline linear controller, designed and verified using techniques discussed in chapter Linear Flight Control Techniques for Unmanned Aerial Vehicles in this book, with an adaptive controller that deals with nonlinearities and modeling uncertainties. Let x.t/ 2 0 ^ y T rf . / > 0 Proj .; y/ D ˆ ˆ y; if not ˆ ˆ :   .    /T 1 Proj .; y/  y  0; 8  2 ˝0 ;  2 ˝1 ; y 2 Rn P D Proj .; y/ Œ .0/ 2 ˝0 ) Œ  .t / 2 ˝1 ; 8t  0 ˝1 D f W

Projection operator

n f . /  0g D  W

Also of key importance is the projection operator modification (Ioannou and Fidan 2006), shown in Table 30.3. As defined, this modification acts on two column vectors .; y/. For matrices, the projection operator is applied column-wise. By design, projection-based MRAC laws will force the tracking error to become small while keeping the adaptive parameters within their prespecified bounds. Without robustness modifications, the adaptive law dynamics are defined by integrating a nonlinear function, represented by the regressor  vector ˆ .x/, multiplied by a linear combination of the state tracking errors e T P B . This product is further multiplied by a constant matrix ‚ (the integral gain), and finally it is integrated to O .t/ (see Fig. 30.6). yield the adaptive parameters ‚ As seen from the block-diagram, there is a chain of nonlinear integrators in a feedback loop, whose output constitute the adaptive parameters. In all practical applications, feedback integrators must be “managed” in the sense that their

30 Robust and Adaptive Control Methods for Aerial Vehicles

ycmd

697

xref

Reference Model

− + System

x

u

Φ(x) ˆ Θ

1 s

e

PB

ΓΘ

Fig. 30.6 MRAC system viewed as a nonlinear integral feedback controller

output signals (i.e., the adaptive parameters) need to be constrained. This prevents integrators against “winding up” due to nonlinear saturation functions in the control channels, where the system achievable control limits are defined and enforced. Control techniques that prevent the integrator windup problems are called the “antiwindup” methods, and the projection operator is one of them. So in practice, an MRAC architecture would consist of the smoothed dead-zone modification coupled with the projection operator. These are the two must-have modifications for enabling MRAC systems to efficiently operate in unknown environment.

30.5.3 Observer-Based MRAC Design with Transient Guarantees Even though an adaptive controller enables command tracking asymptotically in time (as t ! 1/, it provides no uniformly guaranteed bounds on how large the transients might become prior to acquiring the command. In order to yield fast tracking and thus shorten transients, one needs to increase the rates of adaptation and, thus, to speed up MRAC laws. However, experience shows that if these rates grow large, unwanted transient oscillations will appear during the initial few seconds (the transient time) of operation. The balance between achieving fast tracking and avoiding undesired transients constitutes the MRAC design trade-off phenomenon. In essence, the rates of adaptation must be chosen large enough for fast tracking but not too large so that unwanted transients are precluded. To understand the intricacies in MRAC design, reconsider the scalar design (30.42). What complicates the MRAC tuning process is the direct dependence of the transient dynamics (30.46) on (a) the external command and (b) the initial conditions for the system and the adaptive controller. These dependencies may too lead to undesirable transients. Consider the error dynamics (30.46). Using Lyapunov arguments, one can prove that the time-varying signal ' .t/ D b .kx .t/ x .t/ C kr .t/ r .t//

(30.73)

698

E. Lavretsky

is uniformly bounded in time and that the tracking error e .t/ globally asymptotically tends to zero, as shown in (30.48). Still, the time constant of the transient dynamics (30.46) e D ja1ref j is exactly the same as in the reference model (30.42). Although having the same time constant in both systems is theoretically correct, any control practitioner would want to have the transient dynamics (30.46) evolve faster than the desired reference model. In other words, the transients must die out quickly, relative to the dynamics of the reference model trajectories. This design requirement is identical to the one that takes place during the construction of asymptotic state observers, originally developed by Luenberger in his PhD thesis at Stanford (1963). Per Luenberger, the reference model in (30.42) represents an open-loop observer. So, just like in choosing the closed-loop observer dynamics, one can add an error feedback term to the reference model and arrive at the observer-like reference model: xP ref D aref xref C bref r C ke .x  xref /

(30.74)

Error Feedback Term

where ke > 0 is the reference model feedback gain. The newly introduced error feedback term in (30.74) is equivalent to the output innovation feedback in a state observer. It is easy to see that in this case, the corresponding error dynamics become faster than the open-loop reference model from (30.42): eP D .aref  ke / e C b .kx x C kr r/

(30.75)

Once again, Lyapunov-based arguments can be easily repeated to prove (a) global asymptotic stability of the modified error dynamics (30.74) and (b) uniform boundedness of all signals in the related closed-loop system. Readers familiar with the MRAC stability proof concept should recognize that using the same Lyapunov function candidate (30.47), one needs to compute its time derivative along the trajectories of (30.75), substitute the adaptive law from (30.42), and then show that the resulting time derivative is globally nonpositive. This will prove uniform boundedness of the tracking error e and of the parameter estimation errors (30.45). Furthermore, since in the observer-like reference model (30.74), aref < 0 and the error feedback term is bounded, then the model state xref is bounded as well. The rest of the proof follows standard (in MRAC) stability arguments, finally arriving at (30.48). Revised block-diagram with the observer-like reference model (30.74) is shown in Fig. 30.7. As the reference model error feedback gain ke is increased, the system transient dynamics become less oscillatory. In order to gain further insights into the transient behavior, choose k0 > 0, a small positive parameter ", and redefine the reference model feedback gain: k0 (30.76) ke D "

30 Robust and Adaptive Control Methods for Aerial Vehicles

699

ke External Command

Reference Model

Ref. Model System Output – + Response

Adaptive Law Control Command

Controller

System Response

Process

Fig. 30.7 MRAC block-diagram with observer-like reference model

Then the modified error dynamics (30.75) become " eP D ." aref  k0 / e C " Œb .kx x C kr r/ „ ƒ‚ …

(30.77)

'.t /

Since all signals in the closed-loop system are uniformly bounded, it is not difficult to show that there exists a strictly positive finite constant 0 < 'max < 1 such that for any " > 0, the upper bound j' .t/j  'max holds uniformly in time and ". Furthermore, starting from an initial condition e .0/ D e0 , the solution of (30.77) can be written explicitly: e .t/ D e

  k aref  "0 t

Zt e .0/ C



e

aref 

k0 "



.t  /

' ./ d 

(30.78)

0

and one can compute an upper bound for this signal: t

je .t/j  e k0 " je0 j C

'max " k0

(30.79)

This relation is valid  for anyfixed " > 0 uniformly in time. So, the system state x .t/ " of the reference model state xref .t/ exponentially fast converges within ˙ 'kmax 0 t

and at the rate no slower than e k0 " . This term gives an upper-bound quantification for the decay rate of the MRAC transient dynamics due to initial conditions mismatch x .0/ ¤ xref .0/. Otherwise, the system transients would remain within

700

E. Lavretsky

  "-dependent bounds ˙ 'kmax " . Consequently, the system transients are reduced by 0 decreasing ", which according to (30.76) corresponds to increasing the reference model feedback gain ke . Being able to influence and shape the MRAC transient dynamics constitutes the essential benefit of the Luenberger-like reference model modification (30.74)–(30.76). Relation (30.79) can also be written as x .t/ D xref .t/ C C e ke t C o .1/

(30.80)

where C > 0 is a constant independent of ke and o .1/ is the “small-O” signal (a function of time that decays to zero asymptotically, as t ! 1/. The second term in (30.80) defines the transient dynamics due to initial conditions. Consequently, with a large enough feedback gain ke , MRAC transient dynamics can be quantified and forced to decay as fast as needed. Note that since ke is inversely proportional to ", then the obvious trade-off in the modified MRAC design would be to avoid high gain effects in the reference model. As for standard MRAC, it is also possible to generalize the observer-based MRAC to a broad class of nonlinear MIMO uncertain dynamical systems in the form 1 0 d .x p /         ‚ …„ ƒC  B Imm ePy I ey I 0mm Cp 0mm T C B D C ƒ @u C ‚d ˆd xp A C ycmd 0np m Ap 0np m xP p xp Bp „ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ … x A Bref xP  B y D 0mm Cp x „ ƒ‚ … C

(30.81)

These dynamics incorporate an np -dimensional open-loop system with m control inputs u- and m- regulated outputs y. This is the original plant, whose state is xp 2 Rnp . The plant is augmented by the m-dimensional integrated output tracking error dynamics ePy I D Cp xp ycmd , where Cp 2 Rmnp is a known constant matrix. The order of the complete system (30.81) is n D np C m. In addition, x 2 Rn is the system state vector, u 2 Rm is the control input, y 2Rpis the regulated  output, ycmd 2 Rm is the commanded signal for y to follow, d xp D ‚Td ˆd xp 2 Rm N m is a nonlinear state-dependent matched parametric uncertainty, is the   ‚d N2 R matrix of unknown constant “true” parameters, and ˆd xp 2 R is the known N -dimensional regressor vector, whose components are locally Lipschitz continuous in x, that is, there exists a finite positive known constant 0 < Lˆd < 1 such that for any .x1 ; x2 / 2 Rnp from a bounded neighborhood of the origin, the following inequality holds kˆd .x1 /  ˆd .x2 /k  Lˆd kx1  x2 k

(30.82)

Also in (30.81), A 2 Rnn , B 2 Rnm , Bref 2 Rnm , and C 2 Rmn are constant known matrices, while ƒ 2 Rmm is a constant diagonal unknown matrix with strictly positive diagonal elements.

30 Robust and Adaptive Control Methods for Aerial Vehicles

701

Consideration of the process dynamics (30.81) is largely motivated by aerospace   applications, where xp models the 6-DoF motion of an airborne platform and d xp represents uncertainties in the vehicle aerodynamic moments: By definition, the moment uncertainties appear together with the system control inputs, thus enforcing the matching conditions needed to justify mere existence of a control solution. Moreover, control actuator uncertainties, control effectiveness reduction, and other control failures are modeled by an unknown constant matrix ƒ. Finally, inclusion of the integrated output tracking error ePy I D Cp xp  ycmd into the open-loop system leads to the extended system formulation (30.81). This inclusion is optional, yet it allows the designer to explicitly account for baseline controllers with integral feedback, and it also allows to avoid feedforward terms in a control solution. Other dynamics, such as structural notch filters, sensors, and actuators, can also be added in the formulation of the extended open-loop system. In order to control a dynamical system such as (30.81), one needs the nominal system (no   uncertainties) to be controllable. It is well known that controllability of Ap ; Bp , coupled with the rank condition,  rank

Ap Bp Cp 0pm

 D np C m D n

(30.83)

ensures controllability of the extended pair .A; B/. Disregarding the system uncertainties, it is straightforward to form the ideal reference model dynamics:

where

xP ref ideal D Aref xref ideal C Bref ycmd

(30.84)

  1 T Aref D A  B Rref B Pref „ ƒ‚ …

(30.85)

T Klqr

is Hurwitz, Klqr is the baseline LQR feedback gain, Pref is the unique symmetric positive definite solution of the ARE, 1 T Pref A C AT Pref  Pref B Rref B Pref C Qref D 0

(30.86)

and .Qref ; Rref / are some appropriately chosen symmetric positive definite matrices. Using the LQR design is the preferred way to formulate reference model dynamics and to embed basic performance of an LQR PI controller into system specifications. Due to inclusion of the integrated tracking error in (30.81), the DC gain of the reference model (30.84) is unity. Consequently, if ƒ D Imm and T d .x/ D 0m1 , then the baseline LQR PI linear state feedback control ulqr D Klqr x enforces global exponential stability of the reference model (30.84), and it makes the regulated output y .t/ track any bounded command ycmd .t/, with bounded errors. For a step-input command, the LQR PI controller provides global exponential tracking with zero steady-state errors. Also, it is easy to see that such a choice of the reference model enforces the model matching conditions, whereby given a

702

E. Lavretsky

Hurwitz matrix Aref and an unknown constant positive definite diagonal matrix ƒ, there exists a constant possibly unknown gain matrix Kx such that Aref D A  B ƒ KxT

(30.87)

It is important to understand that in this case, existence of Kx is guaranteed for any controllable pair .A; B/ and for any nonsingular matrix ƒ. In particular, relations (30.85) and (30.87) imply Kx D Klqr ƒ1 (30.88) Using (30.87), it is convenient to rewrite the system dynamics (30.81) in the form  

xP D Aref x C B ƒ .u C KxT x C ‚Td ˆd xp / C Bref ycmd „ ƒ‚ …    T T x  Kx ‚d „ ƒ‚ … ˆd xp „ ƒ‚ … ‚T

(30.89)

ˆ.x/

and then get

  xP D Aref x C B ƒ u C ‚T ˆ .x/ C Bref ycmd

(30.90)

The control goal of interest is bounded tracking of ycmd in the presence of the system parametric uncertainties fƒ; ‚g. Specifically, one needs to find a control input u such that the regulated output y D C x 2 Rm tracks any bounded time-varying command ycmd .t/ 2 Rm with bounded errors, while the rest of the signals in the corresponding closed-loop system remain bounded. In addition, it is desirable to have smooth and quantifiable transient characteristics in the closed-loop dynamics. Using Lyapunov-based arguments (Khalil 1996), coupled with asymptotic analysis (Kevorkian and Cole 1996), one can derive MRAC systems with quantifiable transient performance. Similar to (30.74) and for the system dynamics (30.90), consider a Luenbergerlike reference model in the form xP ref D Aref xref C Lv .x  xref / CBref ycmd

(30.91)

Error Feedback Term

where xO 2 Rn is the reference model state and Lv 2 Rnn is the error feedback gain, parameterized by a positive scalar v > 0 (to be defined). The system control input u is selected as O T ˆ .x/ u D ‚ (30.92) Substituting (30.92) into the system dynamics (30.90) gives   O  ‚T ˆ .x/ C Bref ycmd xP D Aref x  B ƒ ‚ „ ƒ‚ … ‚

where ‚ 2 RN m denotes the matrix of parameter estimation errors.

(30.93)

30 Robust and Adaptive Control Methods for Aerial Vehicles

703

  O will be selected such that the system state x In what follows, the pair Lv ; ‚ globally asymptotically tracks xref – the state of the observer-like reference model (30.91), and so y ! yref . Also, one can show that xref tracks xref ideal , which in turn t !1 implies that yref ! yref ideal . Furthermore, since the output of the ideal reference t !1 model (30.84) follows its command yref ideal ! ycmd , with bounded errors, and y ! yref ! yref ideal , then the system-regulated output y will also track ycmd with t !1 t !1 bounded errors. This argument constitutes the proposed design strategy. O so that x globally asymptotically tracks Begin by choosing adaptive laws for ‚ xref , in the presence of the system uncertainties. Let, e D x  xref

(30.94)

denote the state tracking error. Subtracting (30.91) from (30.93), gives the system transient dynamics: eP D .Aref  Lv / e  B ƒ ‚T ˆ .x/

(30.95)

Choose the error feedback gain Lv as Lv D Pv Rv1

(30.96)

where Pv D PvT > 0 is the unique solution of the following ARE, Pv ATref C Aref Pv  Pv Rv1 Pv C Qv D 0

(30.97)

with the ARE weight matrices .Qv ; Rv / selected as  Qv D Q0 C

vC1 v

 Inn ;

Rv D

v Inn vC1

(30.98)

using a constant parameter v > 0. This constant will eventually become the design “tuning knob,” where small values of v yield better MRAC transients. However, the corresponding feedback gain Lv will increase at the rate of 1v . In fact, as v tends to zero, the error feedback gain tends to infinity:    1 1 Pv D O Lv D 1 C v v 

(30.99)

while the solution Pv of the ARE (30.97) tends to a constant positive definite symmetric matrix P0 . It is easy to verify that the ARE (30.97) possesses the unique symmetric positive definite solution Pv . Furthermore, because of (30.97), the observer closed-loop matrix

704

E. Lavretsky

Av D Aref  Lv D Aref 

Pv Rv1

  1 D Aref  Pv 1 C v

(30.100)

satisfies 0

1T

1

0

B C C B Pv @Aref  Pv Rv1 A C @Aref  Pv Rv1 A Pv C Pv Rv1 Pv C Qv D 0 (30.101) „ ƒ‚ … „ ƒ‚ … „

ƒ‚ Av

or equivalently

Lv

…

„

ƒ‚ Av

Lv

…

Pv ATv C Av Pv D Pv Rv1 Pv  Qv < 0

(30.102)

and therefore, Av is Hurwitz for any v > 0. Since Pv is the unique symmetric positive definite solution of the ARE (30.97), then the matrix inverse PQv D Pv1 exists for any v  0 and the following relation takes place: ATv PQv C PQv Av D Rv1  PQv Qv PQv < 0 (30.103) O so that the tracking error e The design task is to choose adaptive laws for ‚ globally asymptotically tends to the origin. Toward that end, consider the following Lyapunov function candidate:   V .e; ‚/ D e T PQv e C trace ƒ ‚T ‚1 ‚

(30.104)

where ‚ D ‚T > 0 is the adaptation rate. The time derivative of V , along the trajectories of the error dynamics (30.95), can be computed as   PO VP .e; ‚/ D e T PQv eP C eP T PQv eP C 2 trace ƒ ‚T ‚1 ‚    T D e T PQv Av e  B ƒ ‚T ˆ .x/ C Av e  B ƒ‚T ˆ .x/ PQv eP OP C2trace ƒ‚T ‚1 ‚   D e T PQv Av C ATv PQv e 2e T PQv B ƒ ‚T ˆ .x/ PO C2 trace ƒ ‚T ‚1 ‚ (30.105) Because of (30.102) and using the properties of the matrix trace operator,      PO  ˆ .x/ e T PQ B VP .e; ‚/ D e T Rv1 C PQv Qv PQv e C 2 trace ƒ ‚T ‚1 ‚ v (30.106)

If the adaptive laws are chosen as PO D ˆ .x/ e T PQ B ‚ ‚ v

(30.107)

30 Robust and Adaptive Control Methods for Aerial Vehicles

then

  VP .e; ‚/ D e T Rv1 C PQv Qv PQv e  0

705

(30.108)

and, hence, V .e; ‚/ is the Lyapunov function for the error dynamics (30.95). For this reason, the tracking error signal e as well as the parameter error matrix ‚ are uniformly bounded in time, that is, .e; ‚/ 2 L1 . Since Aref in (30.91) is Hurwitz by design and .e; ycmd / 2 L1 , then .xref ; xP ref / 2 L1 and consequently x 2 L1 . Since the unknown parameters ‚ are constant and ‚  2 L  1 then   O O ‚ 2 L1 . The regressor vector ˆ xp is Lipschitz continuous and x; ‚ 2 L1 . Therefore, from definition (30.92) it follows that u 2 L1 and consequently xP 2 L1 . Also, since xP ref 2 L1 , then eP 2 L1 . Using (30.108) yields   VR .e; ‚/ D 2 e T Rv1 C PQv Qv PQv eP 2 L1

(30.109)

The function V from (30.104) is lower bounded and has a nonincreasing time derivative as in (30.108). Thus, V tends to a limit, as t ! 1. Also the function’s second time derivative is uniformly bounded. Therefore, VP is a uniformly continuous function of time. Using Barbalat’s lemma (Khalil 1996) implies that VP .t/ tends to zero, as t ! 1. Finally, and due to (30.108), lim ke .t/k D 0

t !1

(30.110)

which proves global asymptotic stability of the tracking error, attained by the adaptive controller (30.92), the adaptive laws (30.107), and the observer-like reference model (30.91). In order to show that xref asymptotically tracks xref ideal , it is sufficient to subtract (30.84) from (30.91) and write the dynamics of the reference model error eref D xref  xref ideal : ePref D Aref eref C Lv e .t/ „ƒ‚… o.1/

(30.111)

Then, Zt eref .t/ D exp .Aref t/ eref .0/ C 0

exp .Aref .t  // Lv e ./ d  D o .1/ ! 0 t !1 „ƒ‚… o.1/ (30.112)

So x ! xref ! xref ideal , and hence, t !1

t !1

.y D C x/ ! .yref D C xref / ! .yref ideal D C xref ideal / ! ycmd .t/ (30.113) t !1

t !1

In other words, the system-regulated output y asymptotically tracks its ideal reference command yref ideal , and y also tracks its original command ycmd with bounded errors.

706

E. Lavretsky

Table 30.4 Observer-based MRAC design summary   Open-loop plant xP D Aref x C B ƒ u C ‚T ˆ .x/ C Bref ycmd Observer-like reference model xP ref D Aref xref C Lv .x  xref / C Bref ycmd State tracking error e D x  xref Riccati equation for adaptive laws Pv ATref C ArefPv  Pv Rv1 Pv C Qv D 0 Qv D Q0 C

ARE weight matrices

vC1 v

Inn ;

Rv D

v vC1

Inn

Pv Rv1 O T .x/

Lv D u D ‚ ˆ PO D ˆ .x/ e T P 1 B ‚ ‚ v

Observer gain Total control input MRAC laws

The design summary is given in Table 30.4. In order to analyze the transient dynamics (30.95), substitute (30.96) into (30.95) and write the transient error dynamics as   eP D Aref  Pv Rv1 e  „ ƒ‚ … Hurwitz Matrix

B ƒ ‚ .t/T ˆ .x .t// ƒ‚ … „

(30.114)

'.t / D Uniformly Bounded Function of Time

Using (30.98) gives     1 eP D Aref  1 C Pv e  ' .t/ v

(30.115)

In (Lavretsky 2011), it is shown that the asymptotic relation Pv D P0 C O .v/ ;

as v ! 0

(30.116)

holds with a constant positive definite symmetric matrix P0 . Then,     1 eP D Aref  1 C .P0 C O .v// e  ' .t/ v

(30.117)

v eP D .v Aref  .v C 1/ .P0 C O .v/// e  v ' .t/

(30.118)

or equivalently

Rewrite (30.118) as v eP D .v 0 Aref  .v C 1/ .P0 C O .v/// e  v ' .t/ 1 C B D @P0 C .v Aref  v .P0 C O .v//  O .v//A e C v ' .t/ „ ƒ‚ … O.v/ D .P0 C O .v// e C v ' .t/

(30.119)

30 Robust and Adaptive Control Methods for Aerial Vehicles

707

or, equivalently,

1 (30.120) .P0 C O .v// e C ' .t/ v It is not difficult to show (by direct integration), that solutions of (30.120) satisfy the following asymptotics eP D

  t e .t/ D O e  v C O .v/ ;

.v ! 0/

(30.121)

uniformly in time, with a positive constant  and for all sufficiently small v > 0. So, the transient dynamics   exponentially decay to a neighborhood of the origin,  vt no slower than O e . Moreover, the “diameter” of the convergence set can be made arbitrarily small, by choosing v to be sufficiently small. This argument formally proves and quantifies transient dynamics improvements in MIMO MRAC systems with observer-like reference models. There is also an alternative way to analyze the transient dynamics in (30.118). This system can be viewed as singularly perturbed, where v plays the role of a small parameter. To understand the intricacies of the system behavior, one can employ the singular perturbation arguments (Khalil 1996; Kevorkian and Cole 1996). Setting v D 0 gives the isolated root e D 0 for the corresponding reduced system, which describes asymptotic behavior as t ! 1, that is, for a sufficiently small v > 0, the error trajectories converge to a small neighborhood of the manifold e  0 and will evolve near this manifold thereafter. Next, the boundary-layer system is formed to quantify and characterize the transient dynamics. These dynamics are derived by “stretching” the time t D (30.122) v rewriting (30.118) in the “fast” time scale , and then setting v D 0. The resulting boundary-layer dynamics de D P0 e (30.123) d are globally exponentially stable, since P0 is symmetric and positive definite (Kevorkian and Cole 1996). In this case, one can claim (Khalil 1996) that for a sufficiently small v > 0, while starting from an initial time t0  0, the singular perturbation system (30.118) has a unique solution e .t; v/, defined on an infinite interval Œt0 ; 1/, and the asymptotic relation e .t; v/ D eN

  t C O .v/ v

(30.124)

  holds uniformly on Œt0 ; 1/, where eN vt is the solution of the boundary-layer system (30.123) and t0 > 0 is the initial time instant. Since, eN

  t D exp .P0 .t  t0 // eN .0/ v

(30.125)

708

E. Lavretsky

then substituting (30.125) into (30.124) results in    t  t0 .x .t0 /  xref .t0 // C O .v/ e .t; v/ D exp P0 v

(30.126)

This asymptotic relation is conservative. In fact, it has been already proven that the tracking error e .t; v/ asymptotically converges to the origin, starting from any initial condition. Consequently, h i ' .t/ D B ƒ ‚ .t/T ˆ .x .t// D o .1/ ; ƒ‚ … „ o.1/

.t ! 1/

(30.127)

and so, (30.126) can be rewritten as    t  t0 x .t; v/ D exp P0 .x.t0 /  xref .t0 // C xref .t/ C O.v/o.1/ v Global Asymptotic Stability Transient Dynamics

(30.128) where P0 is a constant symmetric positive definite matrix, o .1/ is a function of time with lim o .1/ D 0, and O .v/ decays to zero no slower than v. Design details and t !1 stability proofs can be found in (Lavretsky 2011). The asymptotic expansion (30.128) quantifies the MRAC transient dynamics. Indeed, for a sufficiently small v > 0, the transients, described by the first term in (30.128), decay exponentially fast, while the second term defines asymptotic behavior of the tracking error, as t ! 1. This constitutes the main benefit of using the error feedback in an observer-based reference model. Essentially, with a sufficiently small parameter v > 0, one ensures quantifiable transient characteristics, and the latter are given by the first term in (30.128). The observer-based MRAC design method represents a numerically efficient technique of reducing unwanted transient oscillations in state feedback/feedforward MRAC systems. The plant dynamics (30.81) and the corresponding control problem formulations can be modified to include nonparametric uncertainties, such as matched uncertainty approximation errors and bounded possibly nonmatched process noise. In that case, one can use known robustification techniques (i.e., -modification, e-modification, and the projection operator) to prove bounded tracking performance and then establish transient characteristics. Also, the state feedback MRAC design, with an observer-like reference model, can be extended to adaptive output feedback controllers (Lavretsky 2012).

30 Robust and Adaptive Control Methods for Aerial Vehicles

709

Specifies desired closed-loop dynamics Reference Model

Desired Response Tracking Error

Actual Response

Command

Robust Baseline Autopilot Nonlinear Adaptive Augmentation Adaptive/Learning Process Adaptive Flight Control System = Robust Baseline Autopilot + Nonlinear Adaptive Augmentation

Fig. 30.8 (Robust + adaptive) flight control system

30.6

Conclusion

Robust and adaptive methods can be seamlessly combined to construct resilient controllers, applicable to a wide range of systems, including aerial platforms. A notional block-diagram is shown in Fig. 30.8. This system is designed to track and execute external commands, provided by a pilot, a guidance logic, or an autonomous mission planner. The architecture embeds a robust baseline controller (LQR PI feedback). The reference model represents the baseline closed-loop dynamics that would be achieved under the baseline controller and without uncertainties. The adaptive control acts as an augmentation to the baseline. Its purpose is to recover the desired baseline performance while operating in the presence of “unknown unknowns” in the system dynamics and operational environment. As depicted in the figure, this control architecture was designed and flown on various vehicles at the Boeing Company. Some are in production today, and yet others were designed to test and verify extreme capabilities and resilience of aerial platforms equipped with robust and adaptive flight controllers.

References R.L. Butchart, B. Shackcloth, Synthesis of model reference adaptive control systems by Lyapunov’s second method, in Proceedings of 1965 IFAC Symposium on Adaptive Control, Teddington, UK, 1965 B. Etkin, Dynamics of flight. Stability and control, 2nd edn. (Wiley, New York, 1982)

710

E. Lavretsky

G.F. Franklin, J.D. Powel, A. Emami-Naeni, Feedback Control of Dynamic Systems (AddisonWesley, Reading, 1986) P. Ioannou, P. Fidan, Adaptive Control Tutorial. Advances in Design and Control (SIAM, Philadelphia, 2006) J. Kevorkian, J.D. Cole, Multiple Scale and Singular Perturbation Methods. Applied Mathematical Sciences, vol 114 (Springer, New York, 1996) H. Khalil, Nonlinear Systems, 3rd edn. (Prentice Hall, Upper Saddle River, 1996) E. Lavretsky, Reference dynamics modification in adaptive controllers for improved transient performance, in Proceedings of AIAA Guidance, Navigation and Control Conference, Portland, OR, 2011 E. Lavretsky, Adaptive output feedback design using asymptotic properties of LQG/LTR Controllers. IEEE Trans. Autom. Control 57(6), 1587–1591 (2012) D. McRuer, I. Ashkenas, D. Graham, Aircraft Dynamics and Automatic Control (Princeton University Press, Princeton, 1990) K.S. Narendra, A.M. Annaswamy, Stable Adaptive Control (Dover, New York, 2005) P.D. Parks, Lyapunov redesign of model reference adaptive systems. IEEE Trans. Autom. Control 11, 362–367 (1966) J.-J.E. Slotine, W. Li, Applied Nonlinear Control (Prentice Hall, Englewood Cliffs, 1995) B.L. Stevens, F.L. Lewis, Aircraft Control and Simulation (Wiley, New York, 1992) H.P. Whitaker, J. Yamron, A. Kezer, Design of Model-Reference Control Systems for Aircraft, Rep. R-164, Instrumentation Laboratory, MIT, Cambridge, MA (1958)

Section VII UAV Communication Issues Eric W. Frew

UAV Communication Issues: Introduction

31

Kimon P. Valavanis and George J. Vachtsevanos

The absence of a pilot onboard an unmanned aircraft necessitates the development and application of robust, effective, and secure communication technologies that will enable unmanned aircraft to unmanned aircraft, unmanned aircraft to ground or mother ship communications, and unmanned aircraft to air traffic controller communications. Unmanned aircraft control and communication issues are tightly coupled and they are treated as such, since the “Achilles heel” of UAVs is controlling them when they are out of range. Problem of UAV Communications by Heppe summarizes the current consensus view (as of the end of 2012) regarding critical data flows for UAS, control link performance requirements, potential frequency bands that can satisfy the relevant technical as well as regulatory constraints, challenges that must be overcome to ensure reliable operation, and possible data link design principles that could lead to a safe and workable implementation. Both line-of-sight and beyond-line-of-sight data links are addressed. It is shown that the challenges to be faced and overcome are significant, but a safe and workable implementation appears to be achievable through data link diversity and other error correction and error recovery techniques. Cognitive Networking for UAV Swarms by Brown, McHenry, and Jaroonvanichkul examines the enabling role of cognitive radio technologies for UAS to access more spectrum needed for flight operations. After requirements are set, different architecture choices available to the UAS cognitive radio designer

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 138, © Springer Science+Business Media Dordrecht 2015

713

714

K.P. Valavanis and G.J. Vachtsevanos

are presented, including the communication architecture, the spectrum awareness techniques for assessing what spectrum is available, and the spectrum access techniques for deciding which available spectrum to use. Information in this chapter is relevant for the development of future UAS rules and standards. Layered Approach to Networked Command and Control of Complex UAS by Elston, Stachura, Dixon, Agrow, and Frew discusses different networking hardware, protocols, and sensors that when combined, they create a diverse and complex UAS through a layered design approach with modular supporting software. It is shown that critical software components, such as service discovery, simplify the inclusion of a diverse set of subsystems and sensors. Maintaining the modularity of these software components ensures that the system can be expanded while requiring minimal software changes. A detailed description of a system is presented, which enabled flight operations of a multi-vehicle unmanned aircraft system for performing targeted, in situ sampling of supercell thunderstorms during the 2010 VORTEX2 field campaign. Cognitive Radio Architectures for Unmanned Aircraft Systems by Wietfeld and Daniel discusses challenges and solution approaches with respect to self-organized and robust multi-UAV systems. The focus is on designing novel service-oriented system architecture to achieve a flexible deployment of UAS. The design aspects of mobility control algorithms addressing the competing requirements of communication reliability and spatial distribution of UAV swarms are demonstrated using aerial sensor networks and ad hoc aerial relay networks. Solution approaches rely on agent-based control of UAV swarms operating autonomously in dynamically changing environments. Different communication-aware algorithms for microscopic as well as macroscopic mobility control are presented, such as Cluster Breathing, Smart Cube, Potential Fields, and Role-Based Connectivity Management. The chapter also discusses how the network and application task-specific performance requirements are met with the proposed mobility control algorithms even in the cases of temporarily unavailable communication links. The self-healing capabilities therefore allow for reliable networking and control of aerial robot swarms in diverse use cases, such as emergency response, environmental monitoring, and ad hoc network provisioning. Collectively, this section addresses current and future trends and challenges in single and networked UAV communication issues, also highlighting UAV integration issues and communication and networking requirements as dictated by air traffic organizations.

Problem of UAV Communications

32

Stephen B. Heppe

Contents 32.1 Introduction and Scope of Communications Challenge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Detailed CNPC Performance Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.3 Frequency Bands for CNPC Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.4 Technical Challenges for CNPC Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.5 A Strawman Architecture for CNPC Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

716 722 729 732 742 747 748

Abstract

Unmanned aircraft have been phenomenally successful in military operations, but have yet to achieve widespread civilian use. This is chiefly due to a concern that unmanned aircraft would pose a danger to other aircraft in the air (including manned aircraft) as well as humans and property on the ground. In order to address this concern and ensure an adequate level of safety, the communication link between the unmanned aircraft and the ground-based pilot – the control link – must carry certain safety-critical data and must be extremely robust and reliable. This chapter summarizes the current consensus view (2012) regarding critical data flows for unmanned aircraft systems, control link performance requirements, potential frequency bands that can satisfy the relevant technical as well as regulatory constraints, challenges that must be overcome to ensure reliable operation, and possible data link design principles that could lead to a safe and workable implementation. Both line-of-sight and beyond-line-ofsight (i.e., satellite based) data links are addressed. It will be seen that the

S.B. Heppe Telenergy, Inc., Hood River, OR, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 30, © Springer Science+Business Media Dordrecht 2015

715

716

S.B. Heppe

challenges are significant, but a safe and workable implementation appears to be achievable through data link diversity and other error correction and error recovery techniques.

32.1

Introduction and Scope of Communications Challenge

In principle, an unmanned aircraft (UA) or unmanned air vehicle (UAV) can operate in several modes: (a) Man-in-the-loop. The pilot on the ground has direct real-time control of all aircraft systems including engine and aerodynamic control surfaces. This is the typical mode of operation for radio-controlled (RC) aircraft operated by the hobbyist. (b) Man-on-the-loop (sometimes called “semi-autonomous”). Onboard automation, such as an autopilot, ensures stable and controlled flight in accordance with flight plans and other directives (including real-time flight commands) received from a pilot on the ground. (c) Autonomous. The aircraft operates without direct real-time human control and responds automatically to changes in its operating environment or aircraft state (although a human may optionally monitor aircraft operations in real time). Of course, a single aircraft might be capable of all three modes of operation. For example, the nominal operating mode for most of a flight might be “man-on-theloop,” but with the capability for “man-in-the-loop” for launch and recovery, and “autonomous return to base” in the event of control link failure. While an aircraft capable of fully autonomous operation from launch to recovery might be a desirable theoretical goal, the practical reality is that the aviation industry – and society at large – does not yet trust the state of the art in automation to allow fully autonomous operation as a normal matter. The operating assumption is that, in the general fault-free case, a certified human pilot will have responsibility for the UA while it is operating (RTCA 2007, p. 11). This phrasing allows for at least two modes of operation: (a) man-in-the-loop (direct real-time control of the aircraft flight trajectory by the pilot); and (b) man-on-the-loop (semi-autonomous with the UA autopilot executing a flight plan previously commanded by the pilot). Of course, with the pilot not physically on the aircraft, a communications link must be provided. In addition, an autonomous mode will be needed for the (extremely rare) event of a control link failure. It should also be noted that policy guidance for very small UA (micro UA) might be less stringent but is yet to be established. This chapter will generally refer to a UA, which is synonymous with a UAV, as part of an “unmanned aircraft system” (UAS) as shown in Fig. 32.1. The UAS consists of the UA and the pilot. These two elements are tied together by control link interactions that provide telemetry and status on the “return link” (UA to pilot) and telecommands on the “forward link” (pilot to UA). For a line-of-sight (LOS) system, the return link and forward link are synonymous with “downlink” and “uplink,” respectively. In a beyond-line-of-sight (BLOS) system, both the return link and the forward link comprise an uplink and downlink to/from a satellite. For example, the

32 Problem of UAV Communications

717

Fig. 32.1 Internal data flows for an unmanned aircraft system (UAS)

UAS Telecommands

UA

PILOT Telemetry

forward link comprises a satellite uplink from the pilot control station (or some other ground-based satellite gateway terminal to which the control station is connected) to the satellite, and a downlink from the satellite to the UA. The UA and its pilot must solve all the problems of a more traditional (manned) aircraft. Specifically, the UAS as a whole must be able to “aviate, navigate, and communicate” – three verbs that are drilled into every pilot. The first, aviate, refers to managing and controlling the aircraft, ensuring stable flight and avoiding obstacles both in the air and on the ground. The second, navigate, refers to knowing one’s current position and flight plan. The third, communicate, refers to voice and data communication with air traffic control, other aircraft in the airspace (which may or may not be actively coordinating with the UA), and associated systems on the ground. In a traditional manned aircraft, only the last required a “technical” means of communication – typically a radio, although light signals and even hand signals have been used in the past. The first two (aviate and navigate) involved “communication” only in the sense of a pilot observing his or her instruments, making adjustments to those instruments, and manually controlling the aircraft itself (either directly, or indirectly via the autopilot). In the case of a UA, all three activities now involve technical means of communication since the intimate connection between the pilot and his or her UA must now be extended over longer distances – from hundreds of meters to thousands of kilometers. These new communications links must be highly reliable and exhibit very high levels of integrity in order to provide the same level of positive control that is achieved by a pilot actually flying in a plane. In order to quantify the levels of performance required, the aviation community relies on a paradigm called “required communications performance” (RCP) which consists of four performance metrics (ICAO 2008): 1. Availability. The probability that an operational communication transaction can be initiated when needed. 2. Continuity. The probability that an operational communication transaction can be completed within the communication transaction time. 3. Integrity. The probability that communications transactions are completed within the communication transaction time with undetected error. 4. Latency and communication transaction time. Latency is the maximum time to deliver a one-way message (or a two-way message plus response) for a given communications system. Communications transaction time is the time for completion of the operational communication transaction (including human response time) after which the initiator should revert to an alternative procedure.

718

S.B. Heppe

“Availability” is simply the long-term average probability that a communications link exists and is able to support a needed communication transaction, at an arbitrary instant of time when a human (or automated system) attempts to send a message. If there is a single radio link connecting the pilot and UA, “availability” would be the probability that both the ground radio and airborne radio are in working order; that the radio link is not corrupted by rain, foliage, jamming, or airframe blockage; and that the radio channel is not over-subscribed by other users. Note that random bit errors on an otherwise fault-free channel are excluded from this definition. “Continuity” is a conditional probability that a communication transaction can be completed (within the necessary time limit), given that the communications facility was “available” at the start of the transaction. Continuity is typically affected by short-term “soft failures” such as multipath fading and airframe blockage. If one ignores the case where a communications link is “failed” at the start of a transaction but “recovers” and successfully delivers a message within the required timeframe, then the probability of actually delivering an arbitrary message is Prfdeliveryg = (availability) (continuity). Values from 0.9 to 0.999 are typical. Particular requirements for UA communications are identified below. The general term “reliability” can be used as a shorthand term encompassing both availability and continuity. Integrity is the probability of an undetected error in a message (sometimes defined as the probability of experiencing an undetected error in any message over a defined span of time, such as an hour). Typical values range from 105 to 1010 or even lower. This definition actually quantifies a “loss of integrity,” but it is easier to cite a single exponent than a large number of nines (e.g., either 5 nines or 10 nines, or even more). Transaction expiration time is the time allowed to complete a transaction before an unsafe condition would result. Typical values have traditionally fallen in the range of seconds to hundreds of seconds. However, for linking a pilot to a UA for real-time perception of the airborne environment and real-time control, more stringent values are needed (fractions of a second to seconds). For one-way transactions such as delivery of telemetry, transaction expiration time is equivalent to the end-to-end latency of the data link. For two-way transactions, such as the delivery of a particular telemetry message indicating a potential collision threat, plus an uplink command executing an avoidance maneuver, transaction expiration time includes the two-way latency of the data link plus the pilot response time. Before detailing the numerical requirements for communications performance, and engineering methods to achieve that performance, it is useful to review the types of communication that must be supported. Figure 32.2 illustrates electromagnetic (radio and radar) interactions between the UAS (the UA and its pilot) and external entities. Some of these have an obvious impact on UAS communications. For example, in the navigation domain, many UA will have GPS and VOR receivers (as examples). These must be controlled and managed by the pilot via the telecommand forward link, and their output data must be delivered to the pilot by the telemetry return link. These data flows are generally low bandwidth (i.e., only limited amounts of data per second) and repetitive (thus, occasional message losses are not critical). Similarly, ATC ground surveillance involves various airborne subsystems which

32 Problem of UAV Communications

719

ATC Ground Surveillance Transponder ADS-B ADS-R ADS-C FIS-B TIS-B

Navigation GPS, VOR, DME, …

ATC ACL ACM AMC ATSA-ITP COTRAC D-ATIS DCL D-FLUP DLIC D-OTIS D-RVR D-TAXI FLIPNT NOTAM VOLMET 4DTRAD

Clearances, Status, Flight Plan Requests (Voice and Data)

UAS Telecommands

UA

PILOT Telemetry

Sense and Avoid TCAS, radar, ADS-B, Wx

Payload data

Cooperative & Non-cooperative Users User/Payload Operator/Cntl

Other NAS

Party line Users (voice)

AIS, Dispatches, Flight Planning

Owner/Operator/ Mission Controller

Fig. 32.2 UAS external interactions

must be controlled by the pilot (e.g., the radar transponder and automatic dependent surveillance – broadcast transmitter), and in some cases their output must be returned to the pilot as well (e.g., traffic information service – broadcast (TIS-B) receivers). It should be noted that some data can flow directly to or from the pilot without passing through the UA. The sense and avoid data streams (for cooperative and non-cooperative targets) are some of the most critical since aircraft move quickly and pilots generally have only a few seconds to make key decisions when things go wrong. These data streams place stringent constraints on the availability, continuity, latency, and integrity of the telecommand and telemetry data links of medium and large UA. Typical airborne equipment in this domain includes traffic alert and collision avoidance (TCAS), the ADS-B receiver, traffic radar, and weather radar. The data from these sensors must flow down to the pilot as telemetry data with very high reliability and integrity, and very low latency. Pilot telecommand responses to the aircraft, necessary to control the aircraft trajectory, avoid obstacles, and ensure safety of flight, must be returned to the aircraft with similar levels of reliability, integrity, and latency. The “air traffic control” (ATC) data flows comprise the traditional voice and data interactions between a pilot and the air traffic control system. The end-points of these interactions are the pilot on one side and the air traffic controller (or other ATC system element) on the other side. Typically for a manned aircraft, a radio on the aircraft would communicate with a radio at the desired ATC facility (such as an airport tower or en route center) or alternatively with a remote radio facility that was networked through ground infrastructure to ATC. This is shown graphically as the radio communications links labeled “A” and “B” in Fig. 32.3. These links

720

S.B. Heppe

A ATC

C1

C2

B

GND Networks

UA Control Station (UACS) D

Radio link Non-radio link

Fig. 32.3 Data link options for ATC communications

are usually shared on a party-line basis by many aircraft in a single ATC sector. For a UA, one could insist that ATC communications continue to be routed to (through) the aircraft as indicated by the links labeled “C1” and “C2” in the figure. This has the advantage that ATC deals with a UA just as it deals with a manned aircraft; the link C1 is equivalent to the link A or B from the standpoint of ATC. This would also maintain the party-line nature of the ATC/UA link since manned aircraft in the vicinity would be able to hear both the ATC and the relayed UA pilot’s transmissions. However, the voice (or data) on the aircraft must be routed to/from the pilot which requires a second radio link C2 in a different frequency band (i.e., to prevent cosite interference caused by simultaneous transmission and reception). C2 could be the same radio link used for UA telecommand and telemetry or a different link. Either way, the tandem combination of C1 and C2 will have lower end-to-end link availability, continuity, integrity, and worse latency, than either C1 or C2 alone. Furthermore, since there is no way for the UA to determine autonomously when a voice signal is coming from ATC and intended for its pilot, versus being transmitted by (or to) another pilot, the link C2 must actually carry all the traffic on link C1 – even though only a small fraction of this traffic is intended for the UA and its pilot. Hence, the bandwidth penalty is significant – each UA would effectively consume the equivalent spectrum of an entire ATC radio channel (albeit in a different band), instead of only a small fraction of a channel as is typical today.

32 Problem of UAV Communications

721

An alternative is to recognize that the UA pilot is actually on the ground and potentially could be accessed without any reliance on radio at all. This is indicated by landline “D” in the figure. Since landlines and other forms of terrestrial communication can be made more robust compared to radio links, this approach has advantages in terms of end-to-end availability, continuity, integrity, and latency (in addition to the savings in terms of bandwidth). However, this so-called “wired ATC network” approach requires an enhancement to the switching and networking capability of the ATC system. A third possibility is to rely on a radio link between ATC (or a remote radio tower) and the pilot in the UA control station (i.e., from one radio tower to another in the figure), thereby avoiding any change to the ATC switching system but nevertheless bypassing the UA itself, thereby avoiding the need for a relay. This approach is particularly convenient when the UA control station (the UACS) is in the same geographic area as, and has line-of-sight visibility to, the ATC facility or its remote radio tower. In reality, all three approaches should be accommodated in a mature architecture since some missions may not allow the pilot control station to easily access terrestrial infrastructure (e.g., disaster relief, remote/oceanic operations, etc.). However, the “wired ATC network” approach should be emphasized due to its benefits in terms of availability, continuity, integrity, latency, and spectrum utilization efficiency. Returning to Fig. 32.2, the party-line (voice) communication to other NAS users refers to the current situational awareness afforded by each pilot listening to the voice interaction between ATC and other aircraft in the vicinity. This may become less important in the future as ATC shifts to data link instead of voice, but party-line voice communications can be maintained (to the extent used) by suitable voiceactivated switching capability between landlines and radios. All of the external data flows addressed so far, in relation to Fig. 32.2, involve communications related to the safety of flight. These external data flows, along with the internal telecommand and telemetry data flows, are jointly identified as “control and non-payload communications” (CNPC). To the extent that radios are used for these flows (recall that ATC voice and data would ideally rely on landlines), the frequency channels used must be protected from a regulatory standpoint and reside in “aeronautical safety spectrum.” This is addressed further below. The owner, operator, and mission controller data flows, in Fig. 32.2, connect the pilot to the aircraft owner/operator or mission commander and do not normally involve radio communication in protected aeronautical spectrum. Finally, payload data (and control and status of the payload) must be addressed in the context of a complete system. However, this is a special category of communication since it does not involve safety of flight. Certain payloads, for example, video cameras, can involve very large data rates and bandwidths but must be handled outside of “aeronautical safety spectrum” if they are used as a mission payload. However, if used for ensuring safety of flight (e.g., for separation assurance or takeoff and landing support), they would fall within the CNPC “safety of flight” domain.

722

32.2

S.B. Heppe

Detailed CNPC Performance Requirements

The international community has been working for several years to estimate the future bandwidth and performance requirements of UA CNPC. This work has proceeded within the RTCA (a not-for-profit corporation that functions as a Federal Advisory Committee for the U.S. Federal Aviation Administration), EUROCAE (a European body also dedicated to aviation standardization), the International Civil Aviation Organization (ICAO), and the International Telecommunications Union (ITU). As of this writing (2012), a consensus is beginning to emerge. This consensus can be understood in terms of the bandwidth needs of an individual UA, the aggregate bandwidth needs of all UA in the airspace (requiring additional assumptions on the number, type, and deployment of UAS), and the RCP performance parameters needed to ensure safety of flight. Bandwidth Requirements for a Single Large UA. The key CNPC data flows illustrated previously in Fig. 32.2, which must be supported in aviation-protected spectrum, can be grouped into three broad categories: (1) command and control, (2) relay of ATC voice and data, and (3) sense and avoid. These involve forward link as well as return link communications which vary as a function of phase of flight and type of flight control employed (manual control, with the pilot directly commanding the control surfaces in real time, versus “automatic” (man-on-theloop) with the pilot directing an onboard autopilot in terms of, for example, speed, climb rate, turn rate, heading, or a detailed flight plan). The estimated data throughput requirements (bits per second) are tabulated in Fig. 32.4 for a large,

COMMAND AND CONTROL PHASE OF FLIGHT AND OPERATING MODE (MANUAL OR AUTOMATIC)

CONTROL UPLINK (UL)

ATC RELAY

NAVAIDS

DOWNLINK (DL)

UL

DL

VOICE RELAY UL and DL

SENSE AND AVOID

DATA RELAY UL

DL

TARGET TRACKS

AIRBORNE WEATHER RADAR

VIDEO

DL

DL

DL

MAXIMUM FOR ANY PHASE OF FLIGHT [1]

4600

7600

670

1100

4800

50

60

9100

28 000

270 000

AVERAGE FOR MANUAL OPERATION

1700

3200

670

870

4800

24

31

9100

8700

270 000

AVERAGE FOR AUTOMATIC OPERATION

440

650

140

190

4800

24

31

9100

8700

270 000

690

1200

250

330

4800

24

31

9100

8700

270 000

OVERALL AVERAGE

References: RTCA SC-203 CC008 (May 2009) (relying in part on ITU-R, 2009. Table 16) NOTES: 1. Flight phases include departure, en route, and arrival. Pre-flight and post-flight phases are excluded. 2. Data throughput values include overhead and are rounded to two significant figures. 3. Target tracks assumes 60 targets in local environment of the UA. 4. Video is only needed intermittently, for some aircraft. 5. Overall average assumes 80% of aircraft are operating automatically in each flight phase; 20% operating manually.

Fig. 32.4 Estimated peak and average non-payload throughput requirements (bps) for a typical large UA over a 4 h flight

32 Problem of UAV Communications

723

well-equipped UA (smaller UA might not support all listed data flows). Note: The “uplink” and “downlink” terminology of Fig. 32.4 can be converted to “forward link” and “return link,” respectively, in order to generalize for both LOS and BLOS communication architectures. The bottom row indicates average requirements for a large UA assuming 80 % of such aircraft are operated “automatically” (man-on-the-loop). All values include link overhead for network control, error correction, and authentication. Examining Fig. 32.4, the peak uplink (forward link) data throughput requirement is roughly 10 kbps (summing the uplink contributions in the top row), and the average data throughput requirement is roughly 5.8 kbps (summing the uplink contributions in the bottom row). A significant element in both of these values is ATC voice relay (4.8 kbps) which, as noted earlier, might be handled via landline without any RF spectrum impact at all – for at least some users. As a consequence, the future CNPC data link must be able to accommodate an uplink (forward link) data throughput of at least 10 kbps per UA (this might be required for some UA for some periods of time), but careful design might allow a large number of UA to share the spectrum resource with each UA consuming on average only 5.8 kbps of system capacity if the UA is used to relay ATC voice communication or only 1 kbps of system capacity if a “wired ATC network” is employed to connect ATC directly with the pilot, bypassing the UA. Similarly, for the downlink (return link), the peak requirement is roughly 320 kbps (again driven by the terminal area approach phase), and the average is about 290 kbps. These are large numbers, but several caveats are important: (a) The 270 kbps for “sense-and-avoid video” (intended to enhance situational awareness) is assumed to represent an intermittent requirement. It is not clear, at this time, if video is intermittently required for some or all classes of UA or only desirable. Some UA might not have this subsystem, and the associated data burden, at all. (b) Weather radar is also a significant data load (the two values in the table reflect compressed versus uncompressed data depending on phase of flight) but might only be installed on a small fraction of the medium and larger UA. It seems likely that a pilot in a ground-based UA control station could access realtime weather products via landline or satellite broadcast, without requiring a dedicated data stream from his or her own UA. (c) Target tracks (sense-and-avoid downlink indications of nearby traffic, including potential “intruders” that represent a risk of collision) are arguably one of the most critical data flows in the entire system. The data volume of this component is estimated at roughly 9.1 kbps. It is imperative that the pilot have a complete and accurate understanding of nearby traffic. However, even here, there is a potential for capacity savings through, for example, “reporting by exception” as suggested at the end of this chapter. (d) As with the uplink assessment, a portion of the downlink is based on relay of ATC voice (4.8 kbps) which, as noted earlier, might be handled via landline without any RF spectrum impact at all – for at least some users.

724

S.B. Heppe

As a result of these considerations, the future CNPC data link(s) must be able to accommodate up to 320 kbps per UA on the return link; however, some UA, some of the time, may require as little as 1.5–2 kbps if they do not use video for sense and avoid, do not have a weather radar installed and operational, rely on target reporting by exception, and rely on a wired ATC network for ATC voice to/from the pilot. Because of the wide range of data throughput requirements for a single UA, it is imperative that a multi-access sharing system be implemented which minimizes wasted spectrum and allows for efficient allocation of radio channel resources to individual UA. Bandwidth Requirements for a Regional Population of UA. The international community is currently attempting to allocate RF spectrum for UA CNPC data links. This requires an assessment of aggregate bandwidth requirements considering the needs of individual UA; the extent to which small, medium, and large UA participate in the data flows indicated above; the population distribution of small, medium, and large UA; the number of UA that might be operational at any one time; and the spatial distribution of these UA both geographically and in altitude (this affects the number of UAS that are mutually visible to one another as well as their associated ground stations). Furthermore, the total bandwidth required for a regional population of UA depends on whether the aircraft are supported by (a) line-of-sight (LOS) radios between the UA and the pilot as illustrated in Fig. 32.3 or (b) beyond-line-of-sight (BLOS) radios using satellite relay. As a first step, the participating members of ITU-R Working Party 5B estimated the peak areal density of UA in the 2030 timeframe. Two methods were used. The first method estimated areal density for all aircraft and then assumed that 10 % of these would be UA. A second method attempted a bottom-up analysis by projecting the number and types of UA missions (military and civilian, governmental, and commercial) that could be implemented by 2030, assuming regulatory guidelines were in place. These two methods resulted in expected areal densities of between 8.6 and 10.4 UA per 10,000 km2 (ITU-R 2009, Tables 5 and 7). This is equivalent to 0.00086–0.00104 UA/km2 . These are average densities which do not account for variations associated with human population density or particular mission requirements (military or otherwise). The total expected population of UA was also subdivided into three broad aircraft classes: small, medium, and large. These three classes were assigned certain operational characteristics, such as typical operating altitude and avionics equipage and typical operating regime. Working from the more conservative (larger) areal density noted above, the consensus view is that, for a uniform geographic distribution, small UA operating below 300-m altitude can be expected to represent an areal density of 0.0008 UA/km2 , medium UA operating between 300 and 5,500 m can be expected to represent an areal density of 0.0002 UA/km2 , and large UA operating above 5,500 m can be expected to represent an areal density of 0.00004 UA/km2 . The altitude range affects channel sharing and frequency reuse across geographic domains, and avionic equipage affects flight operations. For example, small UA were assumed to operate only in daylight at low altitude, within visual range of the pilot (and the UACS), and have only the most basic

32 Problem of UAV Communications

725

flight controls and sensors and therefore have only limited need for data transfer in aviation-protected spectrum. ITU-R WP5B also considered the impact of nonuniform geographic distribution of UA – for example, due to disaster relief such as floods and hurricanes and normal geographic asymmetries due to airport operations. While these factors can lead to a localized increase in areal density by as much as a factor of 10, the group determined that such nonuniformities could be accommodated by relying on spectrum resources in neighboring areas. Thus, total spectrum requirements could be based on the uniform distribution. Based on the analyses performed, ITU-R WP5B concluded (ITU-R 2009) that the total aviation-protected spectrum requirements for UA CNPC data links are: • 34 MHz for a terrestrial LOS system supporting all UA • 56 MHz for a regional-beam satellite BLOS system supporting medium and larger UA only (small UA are assumed to be too small to support the antenna needed for a satellite BLOS system) The satellite-based system requires more spectrum, despite supporting only a subset of the UA, because it is comparatively less able to rely on frequency reuse. Within a terrestrial LOS system, bandwidth requirements would be split between uplink and downlink with UACS to UA data flows consuming 4.6 MHz and UA to UACS data flows consuming 29.4 MHz. Within a satellite-based BLOS system comprising multiple satellites and a degree of frequency reuse, the forward links from the UACS to the satellite, and from the satellite to the UA, each consume 4.1 MHz and the return links from the UA to the satellite, and from the satellite back down to the UACS, each consume 24.05 MHz. RCP Thresholds for CNPC Data Flows. In addition to the individual and aggregate bandwidth requirements, which affect user channelization and total system capacity, it is important to understand the required communications performance in terms of availability, continuity, integrity, and transaction expiration time (including latency). Figure 32.5 provides guidance on the RCP levels recommended for pilotto-controller voice and data communications, based on ICAO (2008). This provides

RCP Type RCP 10

Transaction Continuity Availability Integrity* Time (sec) (per flight hr) (per flight hr) (per flight hr) Usage 10

0.999

0.99998

10-5

Controller voice intervention supporting separation assurance in a 5 nmi radius environment Controller routine communication in a 5 nmi radius environment – data

60

0.999

0.9999

10-5

RCP 120

120

0.999

0.9999

10-5

Controller voice intervention supporting separation assurance in a 15 nmi radius environment

RCP 240

240

0.999

0.999

10-5

Controller voice intervention supporting separation assurance in a 30/50 nmi radius environment

RCP 400

400

0.999

0.999

10

RCP 60

-5

Controller voice intervention supporting separation assurance outside a 30/50 nmi radius environment

*This is actually the probability of a loss of integrity during an arbitrary flight hour.

Fig. 32.5 ICAO guidance on RCP for controller/pilot voice and data communication

726

S.B. Heppe

information related to some, but not all, of the data flows associated with UA CNPC. The reader may note that continuity and availability levels are relatively high (typically 0.999–0.9999 and even 0.99998 for voice availability in an RCP 10 environment associated with a 10-s transaction expiration time) but that the transaction times are on the order of tens to hundreds of seconds. This implicitly allows for some “soft failures” where a link may be interrupted or interfered with for a period of time but then recovers. RTCA Special Committee 203 (RTCA/SC-203) has been working to identify additional requirements, beyond those identified by ICAO, specifically necessary to support UA CNPC data links. Final consensus has not been achieved, but some preliminary results are available in working papers within the committee and can be reported here. Readers are cautioned that further refinement may occur. With regard to controller/pilot voice and data communications for UA, RTCA/ SC-203 Working Group 2 has tentatively determined that the technically achievable round-trip latency for UA is likely to be up to a second longer than the technically achievable round-trip latency for manned aircraft. For ATC voice communications, the current capability for manned aircraft is about 770 ms (95 %) according to RTCA (2003). An additional second of delay for UA will still easily satisfy the ICAO guidance summarized above in Fig. 32.5. However, the situation for ATC data communications is less clear. In 1995, the FAA produced a report entitled Air Ground Data Link VHF ACARS Preliminary Test Report (Rehmann and Mestre 1995). The report concluded that the average round-trip message delay associated with ACARS (an existing air/ground data link system used by airlines and ATC) fell within the range of 10–20 s, with 5 out of approximately 2,300 messages lost. While the authors noted that Aeronautical Radio Inc. (ARINC), the operator of the ACARS, did not endorse the tests performed, these results would fall short of the most stringent ICAO guidance for data communications, even without an additional second of delay. Luckily, it is possible to improve the observed performance of ACARS with an enhanced (or a new) data link system. The working group has recommended further analysis to validate the ability to meet controller/pilot data link requirements for future UA operations. While controller/pilot data link communications represent an obvious issue for future UA operations, the real driving consideration for RCP, in the context of UAS, appears to be the interaction between a pilot and his or her UA in a time-critical safety-related scenario – specifically, a loss of separation between two aircraft flying in opposite directions on a collision course (see Fig. 32.6). In the case of an unmanned aircraft flying beyond visual range of the UACS and without the benefit of an autonomous collision avoidance system, the UA must sense the intruder, alert the pilot through a reliable return link, and respond to pilot forward link commands in order to avoid the collision. The detection of an intruding aircraft, and determination of a collision threat, could be based on TCAS, ADS-B, TIS-B, radar, or even visual detection with a suitable processor and software. The critical data flows are the telemetry containing the target track for the intruding aircraft and the telecommand containing the pilot’s command for the avoidance maneuver.

32 Problem of UAV Communications

727

Detection Range

Conflict Avoidance Period Airspace Within Which A Collision Can Occur

Collision Avoidance Period

Fig. 32.6 Collision avoidance scenario (RTCA 2011)

The entire timeline for this collision avoidance scenario spans only a few tens of seconds. This places high demands on communications link availability, continuity, and latency. RTCA/SC-203 has addressed this scenario with a two-pronged analytic approach: (a) First, identify RCP performance levels that appear adequate to achieve the same levels of safety as demonstrated by manned aircraft today. (b) Second, develop a strawman communications architecture, with sufficient engineering detail, to verify that these performance levels can be achieved in practice. As a preliminary matter, the working group recognized that the short timeline for the collision avoidance scenario places a stringent upper limit on communication latency. Simulations indicate that round-trip delays beyond a few seconds (exclusive of pilot reaction time) can adversely affect the outcome. Furthermore, the working group has recognized that the round-trip communication transaction time cannot be less than roughly 500 ms for a LOS communications system, or 1,100 ms for a BLOS communications system, due to the inherent message assembly and transmission times and propagation delays (RTCA 2009a). Hence, there is a strong incentive to deliver each message reliably on the first attempt. While message repetition (multiple transmissions separated in time) is not precluded as a viable communications technique, it should not be the primary means employed to achieve the necessary levels of link reliability. A precise latency requirement has not been identified, but the working group is proceeding under the assumption that the round-trip latency requirement for 95 % of the transactions must be on the order of 2 s or less. Turning to availability and continuity, two methods were used to identify total probability of message delivery, which could then be used to derive RCP performance levels for availability and continuity (RTCA 2011). Method #1 relied on historical midair accident rates for general aviation (Part 91) and transport category aircraft (Part 121). It then applied assumptions relating to the likelihood of experiencing a potential midair collision scenario and the likelihood that a pilot could avoid a collision by observation and piloting. The UA CNPC data link must

728

S.B. Heppe

have sufficient performance to not significantly degrade the pilot’s ability to avoid the collision. Method #2 relied on the probability of catastrophic failures noted in FAA/AC 23.1309 for Class I and Class III aircraft. Class I (single reciprocating engine, gross weight 6,000 pounds) was used as a surrogate for large UA (Part 121 equivalent operation in Method #1). These two methods resulted in comparable performance values for total probability of message delivery: 99.8 % for smaller and medium UA equivalent to general aviation aircraft operating under Part 91 and 99.999 % for larger UA equivalent to transport category aircraft (Part 121 operations). These are stringent values for a radio communication link – not easily achieved without careful engineering. Furthermore, the reader’s attention is drawn to the fact that larger UA have substantially more challenging requirements than smaller UA. The RCP performance parameters of availability and continuity may be traded off against one another to some degree but must be set so as to jointly satisfy the required overall delivery probability for a particular class of aircraft (either 99.8 or 99.999 %), keeping in mind that Prfdeliveryg D (availability) (continuity): With regard to message integrity, ICAO has already identified a target level for ATC/pilot voice and data communications as noted above in Fig. 32.5. This level is no more than one undetected message error per 105 flight hours. This can be converted to a per-message probability of undetected error (an undetected message error rate, or UMER) by considering the maximum number of messages that can be delivered over a span of time. For example, if the eventual UA CNPC data link supports 25 messages per second, there would be the potential for 90,000 messages per hour and required UMER = 1.1  1010  1010 . This level of integrity may also apply to relatively noncritical pilot/UA control interactions such as routine telemetry and some classes of telecommand which do not have an immediate impact on flight operations. However, this conjecture requires validation. There is also the question of integrity for pilot/UA interactions that are associated with extremely time-critical safety messages (such as track reports for potential intruder aircraft) and uplink commands that affect aircraft trajectory (such as avoidance maneuvers) or changes to flight plans stored in the UA. These have not been addressed by ICAO, but the RTCA/SC-203 has associated these messages with a potential catastrophic loss in the event of an undetected message error and tentatively assigned a required integrity level of no more than one undetected message error per 109 flight hours for larger UA. This is based on the historical midair collision rate of 2:5  109 per flight hour for Part 121 operations, the threshold level of catastrophic failure for Class III aircraft of 108 per flight hour from FAA AC 23/25.1309, and the likelihood of a catastrophic hazard (109 / as listed in FAA (2008). As with the integrity level for ATC voice and data communications, this can be converted to a UMER by considering the number of

32 Problem of UAV Communications

729

message opportunities per hour. If the eventual UA CNPC data link supports 25 messages per second, the derived UMER = 1014 . It should be noted that the analytic framework and RCP performance values and bounds described above are tentative and have not yet been validated by any regulatory body. The ITU based its analysis of bandwidth needs on three classes of UA (small, medium, and large), and these bandwidth requirements are currently being used to identify suitable RF spectrum resources. In slight contrast, the RTCA has tended to analyze RCP performance requirements relative to just two classes – a class of “smaller” UA which is somewhat related to the small UA of the ITU but is more directly associated with performance requirements derived from those of general aviation (Part 91 operations) and a class of “larger” UA which is somewhat related to the medium and large UA of the ITU but is more directly associated with performance requirements derived from those of transport category aircraft (Part 121 operations). The ultimate classification of UA and UAS operations, for regulatory purposes, will be agreed within ICAO based on consensus among the world’s civil aviation authorities; this work has yet to be completed. The performance requirements noted above should be viewed as likely representative of the range of requirements that may ultimately be applied to various classes of UA, recognizing that the specific classifications, breakpoints between classes, and associated requirements are subject to refinement and validation.

32.3

Frequency Bands for CNPC Communications

As noted in ITU-R (2010), current UAS use a wide range of frequency bands for control of the UA in segregated airspace. Systems operate on frequencies ranging from VHF (72 MHz) up to Ku-band (15 GHz) for both line-of-sight (LOS) and beyond-line-of-sight (BLOS). The factors driving the choice of frequency are related to limiting the size, weight, and power of the airborne data link equipment – particularly antennas and power amplifiers – as well as satisfying the data rates required. Many LOS systems operate in frequency bands allocated to the mobile service, or the aeronautical mobile service, but are not accorded extra protection as a safety service such as bands allocated to the aeronautical mobile (route) service (AM(R)S) or the aeronautical mobile satellite (route) service (AMS(R)S). Furthermore, many LOS and BLOS systems share the control link and the payload return link on a common carrier. This is desirable from the standpoint of minimizing radio size, weight, and power but mixes different data types in a way that would not be allowed in safety spectrum. In the future, considering that UA will operate in civil nonsegregated airspace, and that the UA CNPC data link will help ensure the safety of life and property, the RF spectrum employed for CNPC will need to be designated as a safety service. Figure 32.7 lists some existing aeronautical bands, between 100 and 5,150 MHz, which were examined by RTCA/SC-203 for UAS. Other aeronautical or aeronautical-related bands exist but were excluded from this list because either (a)

730

S.B. Heppe

Frequency Band

Allocated Use (typical systems)

108 - 112 MHz

ARNS (ILS LOC)

112 - 117.975 MHz

ARNS (VOR, GBAS)

117.975 - 137 MHz

AM(R)S (VHF voice and data)

960 - 1215 MHz

ARNS (DME, TACAN, Radar Transponders)

1525 - 1559 MHz

AMS(R)S

1610 - 1660 MHz

AMS(R)S

2900 - 3100 MHz

Radiolocation, Radionavigation

5000 - 5150 MHz

ARNS (MLS), Radionavigation Satellite, AMS(R)S, FSS (but limited to MSS feeder links)

Fig. 32.7 Potential aeronautical or aeronautical related bands for UA CNPC data links

they are already allocated to systems which would not be suitable for sharing with UAS or (b) they are at frequencies inappropriate for UAS use. After considering numerous factors including existing congestion in each band, potential for sharing, available bandwidth, link range, capacity, propagation characteristics, and air and ground cosite compatibility (among others), the list was narrowed to two candidates: portions of the 960–1,024-MHz ARNS band for a LOS system; and the 5,030– 5,091-MHz band for either a LOS or BLOS system (or both) (RTCA 2009b). While the 960–1,024-MHz subband currently lacks an AM(R)S allocation as would be required for a LOS system, this was viewed as achievable from a regulatory standpoint. As for the 5,030–5,091 MHz band, the only incumbent user is the microwave landing system (MLS) which is deployed in very limited numbers. Technical analyses indicated that both LOS and BLOS (satellite) CNPC systems could effectively share the band with MLS. While other candidate bands could potentially be used on a case-by-case basis, they would not be suitable for meeting any significant portion of the projected bandwidth requirements of the UAS community (34 MHz of terrestrial spectrum and 54 MHz of satellite spectrum). In a similar vein, the ITU-R has identified portions of the 960–1,164 and 5,030– 5,091-MHz bands as good candidates to support control links for UAS without causing harmful interference to incumbent services and systems (ITU-R 2010). As noted by the ITU, the 960–1,164-MHz band is highly favorable for UAS control links. Rain losses are negligible and free-space path losses are low enough to permit reliable long-range LOS communication between relatively low-power radios using omnidirectional and medium-gain antennas. While much of the band is heavily used, substantial subbands (960–976 and 1,151–1,156 MHz) are not used by airborne NAVAID transmitters and contain no fixed ground-based assignments in some countries. In such countries, 10.4 MHz of spectrum could be made available, which would furnish small UA with badly needed access to protected spectrum and provide

32 Problem of UAV Communications

731

UA of all types with band diversity that is essential for reliable UA CNPC. While the identified spectrum (10.4 MHz) is less than the 34 MHz identified as needed for terrestrial/LOS links, it “would suffice to meet all UAS CNPC requirements except for backup links, video, and downlinking of airborne weather-radar data in some countries.” RTCA/SC-203 has also noted (RTCA 2009b) that Inmarsat and Iridium (satellite systems operating in the 1.5-/1.6-GHz bands) could be considered as existing BLOS systems able to support UA CNPC links. Inmarsat operates in a band with a clear AMS(R)S allocation, has a total bandwidth of 14 MHz available on a worldwide basis, and is suitable for BLOS UA CNPC. Iridium also operates in an aeronautical band and has been used for BLOS UA CNPC by at least one UAS. But its first-generation system has rather limited per-channel bandwidth which limits its suitability for the full range of UAS (however, this may be overcome in its secondgeneration system). RTCA/SC-203 has also considered the potential for a new AMS(R)S allocation at either Ku-band or Ka-band that could support BLOS UA CNPC with a wideband geostationary satellite system. At Ku-band, the 12/14 GHz commercial fixed satellite service (FSS) band and the 13.25–13.4 and 15.4–15.7-GHz bands are all attractive. The latter two bands are allocated to the aeronautical radio navigation service (ARNS) on a primary basis, although the lower band is limited to Doppler navigation systems and both bands have other coprimary allocations. Nevertheless, the existing ARNS allocation indicates the potential to add another aeronautical safety service such as AMS(R)S. Similarly, the 19.7–20.2 and 29.5–30.0-GHz bands (Ka-band) appear potentially suitable for UAS BLOS CNPC links, with the upper 50 MHz of each of these bands appearing to be particularly attractive given other existing users, allocations, and interference considerations. SC-203 noted that “The close proximity between these bands and those used by Ka Band equipped military UAS makes it possible for the military to use the same equipment in both segregated and non-segregated airspace by just tuning to different frequencies without the need to change antennas or modems. This is a very attractive feature of the proposed K/Ka band segments” (RTCA 2009b). As is clear from the preceding discussion, it is challenging but not impossible to assemble the necessary spectrum resources for UAS CNPC data links. Of course, a single UAS could operate within the currently allocated bands using less than a MHz of spectrum, and the first several UAS deployed would not perceive any bandwidth constraints. However, uncontrolled or uncoordinated proliferation of UAS with differing CNPC data link formats, protocols, and waveforms could lead to relatively inefficient use of the spectrum and the potential for unsafe conditions brought about by interference. Therefore, there is a strong incentive to define a common data link standard that could efficiently support the entire community. The work of defining such a standard is likely to take place initially in various advisory bodies such as RTCA and EUROCAE, with an eventual standard being finalized by consensus within the International Civil Aviation Organization (ICAO).

732

32.4

S.B. Heppe

Technical Challenges for CNPC Communications

As noted previously, the required communications performance (RCP) thresholds for UA CNPC data links are stringent. The information flows for sense-and-avoid and pilot control, in particular, must be extremely reliable, have low latency, and have very high integrity, in order to ensure safety of flight. While numerical requirements have not yet been formally adopted, the emerging consensus at RTCA indicates that availability and continuity must satisfy message delivery probabilities on the order of 99.8–99.999 %, latency must be on the order of 1–2 s, and undetected message error rates (integrity) for at least some information flows (for some UA) must be on the order of 1010 –1014 per message. In order to meet these requirements, the system designer must address hardware failures that could permanently disable a link. Hardware failures primarily affect link availability and continuity (not latency or message integrity) and can be mitigated through careful design and system redundancy using well-understood techniques. These will not be further discussed. In addition, the system designer must address “soft failures” such as airframe shadowing, multipath fading, rain attenuation, and random bit errors introduced on the channel, which lead to short-term outages and potential undetected errors. These short-term outages are actually the dominant concern for the most stringent RCP thresholds identified. For example, airframe shadowing or excess rain attenuation, which persists for several seconds, could prevent a pilot from recognizing a safety threat and taking appropriate and timely action. The key impairments, of interest to the UA system designer, are addressed below in the context of the “link budget” which provides a quantitative framework for understanding link performance. The Link Budget as a Framework for Understanding Soft Failure Issues. All radio communication links operate by transmitting energy in a particular physical format (e.g., a “waveform”) from the transmitter to the receiver. Generally, for digital data links, one or more particular parameters of a sinusoidal radio wave are adjusted among a finite number of possibilities from one transmitted symbol to another. For example, the amplitude, frequency, or phase of the sinusoid might be adjusted. These alternatives generally go by the names of amplitude shift keying (ASK), frequency shift keying (FSK) and phase shift keying (PSK). The number of different symbols that can be sent is usually a power of 2; for example, a system with two possible phase states is called binary PSK or BPSK. A system with eight possible frequencies is called 8ary FSK or 8-FSK. Literally scores of different waveforms have been invented over the years. However, all of them involve the transmission of energy from a transmitter to a receiver. The desired signal is generally received at a very low signal level (picowatts, or even less) and in the presence of natural and man-made RF noise from the environment, as well as RF noise generated internal to the receiver itself. This RF noise tends to corrupt the desired signal and prevent accurate decision-making as to the symbol or symbols actually sent by the transmitter. Hence, a primary task of the communications engineer is to ensure that the amount of received energy in the desired signal, during a given symbol

32 Problem of UAV Communications

733

interval or set of symbol intervals, is sufficient to allow accurate reception of the vast majority of transmitted messages. The analytic framework used to predict reliable reception is called the “link budget.” A link budget is similar to a household financial budget, where one starts with monthly take-home pay or revenue, subtracts known fixed expenses as well as an allowance for variable expenses, and checks to see if everything can be covered – hopefully with a little extra “margin” that can be saved for a rainy day (i.e., when variable expenses turn out to be larger than expected). In the case of an RF link budget, one starts with the power transmitted by the transmitter, accounts for known fixed losses as well as an allowance for variable losses, and checks to see if there is sufficient power at the receiver (or, in other words, sufficient energy in each received symbol or set of symbols) to allow reliable reception given the expected RF noise – hopefully with a little extra “margin” that can be used if losses are greater than expected or the transmitter is somewhat weaker than expected.

Parameter

Units

Transmit power

Downlink

40

40

dBi

8

0

dB

1

1

47

39

dBm

Transmit antenna

Uplink

gain1

Cable and filtering loss Transmit EIRP

dBm

FSPL (1 GHz, 75 nmi)

dB

135

135

Additional losses2

dB

23

23

gain1

dBi

0

8

Cable and filtering loss

dB

1

1

Received signal power

dBm

–112

–112

Receiver noise power

dBm

–126

–126

Available Carrier-to-noise ratio (C/N)

dB

14

14

Theoretical C/N

dB

6

6

Implementation losses

dB

2

2

Safety margin

dB

6

6

Required C/N

dB

14

14

Excess margin3

dB

0

0

Receive antenna

1 Variability

of UA antenna gain due to airframe blockage or reflections is taken into account in the airframe loss part of the “additional losses” row.

2 Additional

losses comprise atmospheric losses, airframe losses and multipath losses. These are convolved together and the quoted value provides for 99.8% link availability (0 dB margin, 99.8% of the time).

3 With

double or triple ground diversity, the excess margin could increase by 9 to 14 dB.

Fig. 32.8 Simplified link budget for line-of-sight UA CNPC data link at L-band (Based on RTCA (2011) and Wilson (2011))

734

S.B. Heppe

A simplified sample link budget, showing the major categories of fixed and variable losses that must be considered for a LOS UA CNPC link, is illustrated in Fig. 32.8. This link budget is used here only as an example – many of the values are particular to the specific design and assumed operating scenario; other assumptions as to operating band, waveform, and operating scenario, would lead to different numbers. Also, a BLOS link budget must consider four links instead of just two (the UA uplink and downlink to the satellite and the ground station uplink and downlink to the satellite). Finally, the reader should note that the various gains and losses are multiplicative, but the link budget converts everything to a scaled logarithm called a “decibel” (dB). This allows the gains and losses to be added and subtracted as in a financial budget. Looking to the figure, the first line is the actual transmitter power in dB (milliwatts) or dBm. The numerical value of 40 dBm is equivalent to a 10 Watt transmitter. This link budget assumes that both the ground-based radio (for the uplink) and the airborne radio (for the downlink) have the same transmitter power. The second line is the antenna gain. The ground antenna is assumed to have a moderately directive antenna (8 dB relative to isotropic, or 8 dBi) which “focuses” the RF energy into a beam with roughly 8 dB more power than if the energy were radiated equally in all directions (8 dB is roughly a factor of 6:1). The airborne antenna is assumed to be nominally isotropic, so it has no directive gain relative to an isotropic antenna. However, the airborne antenna is mounted on the aircraft and can easily be shadowed by the airframe. As a result, the actual gain experienced from one moment to the next can vary significantly from this value. Generally it will be worse. This is accounted for in the “additional losses” line below. The third line accounts for transmitter cable and filtering losses. Finally, the fourth line accumulates the first three into a “figure of merit” for the transmitter – its “effective isotropic radiated power” (EIRP). This is used for the subsequent calculations. The free-space path loss (FSPL) row accounts for the weakening of the RF signal as it propagates from the transmitter to the receiver. For a radio link at roughly 1-GHz (L-band), and a 75-nmi path, the loss is 135 dB. This is a nominal value assuming no big obstructions and no refractive losses. Further data and calculation methods are provided in ITU-R (2007). For UA operating at long range and low altitude (expected to be a common scenario), the design engineer should specifically account for diffraction losses which can add on the order of 15–20 dB of loss for an aircraft close to the radio horizon (ITU-R 2007; CCIR 1991). The “Additional losses” row accounts for several time-varying effects which are critical to an understanding of the UA CNPC data link design. These include airframe shadowing, antenna variability, multipath fading, and atmospheric and rain attenuation (especially relevant at 5 GHz and higher). Since these different effects are physically independent, a statistical analysis is needed to determine the “allowance” that must be provided in order to guarantee a given level of performance. For the sample link budget provided here and based on other data and design assumptions as discussed below (including at least two antennas on the aircraft but a single ground antenna), a convolution of the probability distributions indicated that the combined losses would be no worse than 23 dB, 99.8 % of the

32 Problem of UAV Communications

735

time. Thus, a CNPC data link attempting to achieve an availability of 99.8 %, with these assumptions, should allow for 23 dB of variable losses. The next three lines account for the receive antenna gain, cable and filtering losses at the receiver, and final received power level (dBm) at the input to the firststage amplifier, assuming the variable losses noted above were at the threshold value. The aircraft antenna gain (0 dBi) is again a nominal value that does not account for the expected variation already handled in the “additional losses” row. The assessed received signal power is 112 dBm (6  1015 W). The “receiver noise power” line accounts for ambient and receiver-internal RF noise across the channel bandwidth of the receiver. This is usually based on measured data for the environment and measured or predicted data for the receiver equipment chain including the antenna, cables, filters, mixers, and firststage amplifier (at least), referenced to the same reference point as the received carrier power (in this case, the input to the first-stage amplifier). The sample link budget indicates an equivalent noise power of 126 dBm. The difference between the received signal power, or carrier power (C), and the receiver noise power (N) is the “carrier-to-noise power ratio” (C/N). This value (14 dB for the sample link budget) must be large enough to allow for reliable operation, allowing the messages to be received without error (at least most of the time). Note that C/N refers to the integrated power (for the desired signal and the noise) over the receiver channel bandwidth. Another important metric is ratio of the energy in a single transmitted symbol or bit (abbreviated Es or Eb / and the noise power density per Hz of bandwidth (abbreviated N0 /, called Es /N0 or Eb /N0 , respectively. All of these metrics represent a “signal-to-noise ratio” (SNR), but because of the definitional differences, any reference to “SNR” should be precisely defined as one of these particular metrics for a particular application or calculation. The “theoretical C/N” is the C/N for a particular waveform and data link protocol (with its synchronization features, authentication, and forward error correction) needed to achieve the desired message error rate, before accounting for implementation losses (hardware imperfections). For the waveform and forwarderror-correction coding scheme assumed here, this value is 6 dB. To this theoretical C/N value, it is necessary to account for the receiver implementation losses (such as improper timing, inaccurate carrier tracking, amplifier nonlinearities, filter distortion) and also allow for a “safety margin” accounting for all additional effects and link impairments beyond the ability of the engineering team to predict. In this case, the combination of the theoretical C/N, implementation losses, and safety margin (6 + 2 + 6 = 14 dB) exactly matches the available C/N assuming variable link losses at the 99.8 % availability level. Hence, the link closes 99.8 % of the time as required, taking into account all the known variables (the engineering team could say they succeeded in the link design), but there is no “excess margin” beyond the safety margin. Considering the link budget as a whole, it is clear that the “additional losses” line is a key driver. The primary physical factors contained in this line are now addressed.

736

S.B. Heppe

Fig. 32.9 Typical top-mounted antenna performance on a midsized aircraft (RTCA 2011) (Ref: Colby et al. 1996)

Airframe Shadowing and Installed Antenna Performance. It is convenient – but not accurate – to assume that an omnidirectional antenna mounted on an aircraft will actually deliver uniform gain in all directions (or even uniform gain over one hemisphere, such as an upper hemisphere for a top-mounted antenna or a lower hemisphere for a bottom-mounted antenna). Figure 32.9 illustrates typical performance for a midsized manned aircraft with a top-mounted, nominally omnidirectional L-band antenna (either horizontal, vertical, or right-hand circular polarized) operating at 1,575.42 MHz, relatively close to the potential new AM(R)S allocation around 1 GHz currently being considered for UAS CNPC data links. The left-hand panel of Fig. 32.9 indicates significant gain variation within 15ı of the local horizontal plane and very poor performance (as much as 25 dB below nominal) below the aircraft. For an aircraft that is a significant lateral distance from the serving ground station, any bank “away from” the ground station would expose the bottom of the aircraft to the ground station and create a significant drop in antenna performance. The right-hand panel (measured at a 7ı depression angle roughly associated with long-range communications to a ground station, for an aircraft in straight-and-level flight) indicates that typical gain is roughly 5 dB but typically varies by several dB for different azimuth angles relative to the aircraft centerline. Furthermore, more than 5 % of the azimuthal pattern exhibits an additional 5 dB loss, and there are narrow nulls (e.g., exactly off the tail) where antenna gain is as much as 10 dB worse than nominal (if “nominal” is taken to be 5 dB relative to an ideal isotropic antenna). Clearly, these variations must be accounted for if the system is intended to achieve 99.8 % or greater link availability. The loss in performance (as much as 25 dB in this case) can be partially mitigated by installing, and relying on, two or more CNPC data link antennas strategically placed on the UA. For example, top, and bottom-mounted antennas, or antennas located at each wingtip, can be used to ensure that the worst-case losses are never

32 Problem of UAV Communications

737

experienced simultaneously by both (or all) antennas. Multiple antennas on the ground can also be used – particularly if the UA CNPC architecture relies on a shared nationwide infrastructure with multiple radio cells (sectors) each with its own antenna. For three widely separated ground antennas surrounding a UA, spanning more than 180ı of arc as perceived by the UA, at least one ground antenna will always have a nominally unobstructed view to at least one antenna on the aircraft. These techniques are both called “antenna diversity” or “space diversity” (because one node or “station” has two or more antennas at different locations in space). The improvement due to antenna diversity is illustrated in Fig.32.10 for a particular UAV with a carbon composite body and a 3-m wingspan. The upper panel is for a single L-band antenna installed in a winglet (i.e., at the end of a wing, in an RF-transparent housing). The five traces are for different elevation or depression angles relative to the local horizontal, but the key point of this plot is the worst-case performance for a single antenna, experienced at any angle, at roughly 28 dBi. This would imply a required link margin of at least 28 dB, just for this impairment alone, relative to a link designed for a nominal 0-dBi antenna. The lower panel indicates that diversity antennas can reduce the required margin to about 13 dB. If multiple antennas are available on the ground as well, arranged so as to avoid a condition where all of them are on a “line of symmetry” relative to the airborne antennas, or all of them can view only one of the airborne antennas (this requires at least three widely separated ground antennas surrounding the aircraft and spanning more than 180ı of arc as seen from the UA), the margin requirement can likely be reduced to roughly 10 dB or less (RTCA 2011). This is an 18-dB improvement over the no-diversity case. Multipath Fading. Multipath refers to portions of the transmitted signal which are reflected off of objects or terrain in the environment and which arrive at the receiving antenna nearly “out-of-phase” relative to the direct signal. This can significantly reduce the received signal strength. Close to complete cancellation (fading) can occur for short periods of time. Figure 32.11 illustrates typical fading on an L-band air/ground data link for an aircraft flying at 500 m, communicating with a ground station antenna 25 m above ground level. The large-scale lobing evident in Fig. 32.11, with signal strength varying by 10–20 dB, is due to so-called ground-bounce multipath and is relatively predictable for a particular environment and flight profile (as indicated by the red trace which is a mathematical model) (ICAO 1999). The fine-scale fading is due to other effects such as reflections off of individual terrain features, buildings, small bodies of water, and the airframe itself. As with the antenna effects illustrated earlier, these multipath fades must be accounted for if the system is intended to achieve 99.8 % link availability or greater. ITU-R (2007) provides a general method to estimate the probability of fade depth for multipath fading, but does not provide any framework to assess correlation time (the time that the fade depth is relatively constant). This turns out to be a significant factor for CNPC link availability and continuity. Also, while there is a fair amount of experimental data for air/ground data links at VHF, there is relatively little experimental data for air/ground data links (on flight

738

S.B. Heppe

Prob.That Gain < Abscissa

Cumulative Gain Profile – One Installed Antenna 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 –35

100% level (cumulative probability = 1.0)

–30

–25

–10 –15 –20 Absolute Gain (dBi)

–5

–20 deg

+10deg

–10 deg

+20deg

0

5

0 deg

Prob. Gain < Abscissa

Cumulative Gain Profile – Left /Right Diversity Antennas 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 –35

–30

–25

–20

–15 –10 –5 Absolute Gain (dBi)

0

5

Fig. 32.10 Antenna performance with and without diversity

profiles that would be representative of typical UA operations) at L-band and above. While some estimates of expected UA CNPC link performance can be made by extrapolating from lower frequencies and compensating for known differences at the higher frequencies, further work is needed in this area. Lacking precise information on the correlation time of multipath fading for typical UA operating scenarios at the expected operating frequencies, one can nevertheless make a first estimate of required fade margin by calculating the cumulative fade distribution for a given environment, using approved techniques as described, for example, in ITU-R (2007). Typical values at L-band are 18 dB of fading (at the 99.8 % availability level) for an aircraft at 75-nmi range and 10,000-ft altitude, and a ground antenna at 100 ft, and 22 dB of fading (for 99.8 % availability) if the aircraft is at 5,000-ft altitude. At shorter range, for example, 30 nmi, multipath

32 Problem of UAV Communications

739

fading may be on the order of 10 dB less severe – but this is still a significant impairment. As with airframe and antenna variation, multipath effects can be partially mitigated by multiple data link antennas on the aircraft or on the ground (or both). Widely separated ground antennas can partially mitigate ground-bounce multipath (the large-scale variation in Fig. 32.11), and antenna diversity on the ground as well as the aircraft can partially mitigate the short-term fluctuations simply because the reflected signals are unlikely to arrive at all the available antennas with the same relationship of amplitudes and phase shifts. For example, with two data link paths exhibiting uncorrelated statistics, the pair will deliver 99.8 % link availability if each individual link is engineered to provide roughly 95 % link availability. At long range and low altitude, this can reduce necessary fade margins by roughly 10 dB (e.g., to the range of 7–10 dB). Additional diversity paths provide additional gain, although improvements quickly become marginal. Furthermore, if multiple messages are sent in quick succession (i.e., within the correlation time of the fading, so that several messages in succession might be corrupted), a change in channel operating frequency from one message to the next can frequently provide for independent fading statistics. This can be beneficial in systems that rely on selective retransmission (or even multiple transmissions for all messages) to enhance overall link availability. Preliminary evidence and analysis indicate that a slow frequency-hopping system operating within the contiguous bandwidth expected for a LOS system (i.e., on the order of 10 MHz or greater) will provide the necessary statistical independence from one message to the next. Atmospheric and Rain Attenuation. All radio signals are attenuated by atmospheric gases, water vapor, and liquid rain. At L-band and below, this attenuation is generally very small and can be ignored or compensated by a relatively modest margin allowance. However, at 5 GHz and above and especially at Ku- and Ka-band

20 Frequency = 1000 MHz UA Altitude = 500 m GS Altitude = 25 m Diffuse K = 18 dB

10

Signal Strength (dB)

0 –10 –20 –30 –40 –50 –60

Simulation. 10 Hz Duration = 1000s

0

0.5

1

1.5

Distance (m)

2

2.5 x 104

Fig. 32.11 Typical multipath fading (See also ICAO 1999 for a comparison of measured and modeled fading on typical air-ground VHF data links)

740

S.B. Heppe

frequencies currently under consideration for at least some BLOS links, both clear sky and rain attenuation can be significant. Of course, it does not rain all the time (the worldwide average is only about 5 % at a random location), and some aircraft might actually fly above the rain layer for most of their mission. But for aircraft flying in the rain (e.g., during takeoff and landing and for low-altitude operations) and using a data link frequency of 5 GHz or higher for safety communications, link outages caused by rain must be explicitly considered. There are several widely used models for attenuation caused by atmospheric gases, water vapor, and rain. ITU-R (2007) offers a good starting point; this ITU recommendation draws on several other ITU recommendations and provides a good road map for analysis. Generally speaking, atmospheric gases contribute a fairly predictable and constant attenuation for a given operational scenario, whereas the rain models provide a statistical understanding of the percentage of time that a particular attenuation level is exceeded. This directly affects the assessment of link availability. Surprisingly, it turns out that continuity (the problem of a link failing during a message transmission period) can be substantially ignored at the availability levels (and probability of message delivery) relevant to UA CNPC. This is because rain fades tend to persist for tens of seconds or longer, with fade levels changing significantly only over seconds (not milliseconds) as a consequence of the physics of rain. Specifically, the volume of space, through which the RF signal travels, can be thought of as an extended tube or ellipsoid which is partially filled and emptied by the rain. An intense rain burst will tend to “fill” the radio path with water, whereas a lull in the rain will tend to allow the radio path to empty out. Because raindrops fall at a characteristic rate and because the radio path has a significant length and girth, it takes a finite amount of time (typically measured in seconds) to significantly change the amount of liquid water contained in the tube or ellipsoid. Antenna diversity (space diversity) is generally not as effective at combating rain attenuation on BLOS links since the rain fade on two data links supported by a single aircraft tends to be correlated, even if multiple satellites are used for these two data links (i.e., two antennas on the UA pointing to different satellites). Convolution of Time-Varying Effects. As noted earlier, once the characteristics of antenna performance, airframe shadowing, multipath fading, and rain attenuation are understood for a particular LOS or BLOS CNPC architecture and operational scenario, the four effects should be treated as statistically independent and convolved to derive an aggregate distribution of fade probability, to avoid an overly conservative link design. For the case of a typical L-band CNPC data link operating an LOS path over 75 nmi, intended to achieve an overall availability of 99.8 % with dual-link diversity, these separate effects would have a combined margin requirement in excess of 30 dB. In contrast, a proper statistical analysis indicates that the actual requirement is only 23 dB. The reader is cautioned that these numbers are illustrative, and particular UA operating scenarios should be separately validated to ensure that required link performance can be achieved. Lost-Link Considerations. Despite all the engineers’ best efforts, there is always a possibility that the CNPC data link (including all its diversity paths) will be

32 Problem of UAV Communications

741

permanently lost. This could affect the telemetry link or the telecommand link or both. Potential causes are: hardware failures on the UA, hardware failures at the UACS, hardware failures in the ground infrastructure, widespread jamming or interference, and pilot error (e.g., commanding the UA to fly beyond communications coverage). If ATC voice and data communications were being routed through the UA at the time of failure (instead of being routed directly between ATC and the UA control station through a wired ground network), the ATC data flows would be affected as well (note: this is another reason to favor the wired ground network, and only rely on ATC data flows relayed through the UA when no other alternative exists). When a lost-link event occurs, procedural methods must be in place to ensure safety of flight for the affected UA, safety of flight for other aircraft in the airspace, and safety for people and property on the ground. There is as yet no worldwide consensus on how to handle a UA with a lost CNPC link. However, several reasonable procedures can be identified now based on currently available technology and operational procedures: • First, “heartbeat messages” on uplink and downlink can be used to track link performance and quickly identify a lost-link event when it occurs (time out thresholds are currently TBD). If found to be operationally beneficial, summary statistics could even be echoed back to the other node (e.g., the UA could transmit down to the UACS, as part of its telemetry stream, the statistics of the heartbeat messages it observed on the uplink). • If the UA detects a failed CNPC data link, it could autonomously set a unique radar transponder code which would serve to alert ATC, as well as other aircraft in the airspace, that it was suffering a lost link. This would be similar to the existing procedure for radio failure on a manned aircraft. A lost forward link is directly observable by the UA. A lost return link, which leaves the forward link intact, is also possible although less likely. This can be detected if the UACS is echoing back return link heartbeat statistics or if the UACS has the ability to send a dedicated command indicating a lost return link. Hence, it is always possible for the UA transponder to be properly set. Suitable transponder codes remain to be defined but are currently being considered. Special messages could also be broadcast via future (or upgraded) ADS-B systems. • If controller-pilot communications are supported with a wired ground network, these data flows will likely remain intact even if the UA CNPC RF data link fails. Therefore, the pilot can inform ATC that the UA has suffered a lost link. Similarly, if the UA has reset the transponder code, this will be apparent to ATC, and the controller could then contact the pilot. • The UA should have a default flight plan, stored in its onboard computer, which will be autonomously executed in the event of a lost CNPC data link. This flight plan might be updated from time to time during a particular flight but should in any case comprise standardized procedures (following guidelines preapproved by aviation regulatory authorities). For example, this might involve a short period following its current flight plan while broadcasting its condition via its

742

S.B. Heppe

radar transponder and ADS-B transmitter (this allows ATC to clear the airspace immediately around the UA), followed by a climb to a preapproved altitude, followed by execution of a new flight plan such as “circle for X minutes waiting for link to be reestablished” or “return to base” or “fly to alternate recovery site” or “fly to previously identified remote area and self-destruct.” If the UA is equipped with ADS-B, the flight plan being executed could potentially be broadcasted via this medium so that ATC, and surrounding aircraft, would be aware of the UA’s intent. It should be noted that some UA operational today already implement such a lost-link flight plan strategy as part of their standard operational doctrine (absent the broadcast via ADS-B).

32.5

A Strawman Architecture for CNPC Communications

Potential architectures for UA CNPC communications have evolved within RTCA, the ITU-R, and other groups. The strawman architecture described in this section (the “strawman”) is one such solution based on the work in RTCA and is presented here simply to illustrate the types of solutions that could be envisioned for a future system. Other options are certainly possible and will need to be explored before a final conclusion can be developed into a consensus standard for the UAS industry to adopt. With regard to frequency bands, the strawman incorporates LOS links at L-band and C-band and BLOS links at C-band and either Ku-band or Ka-band (or both). Precise band edges are TBD, but current international efforts indicate that it will be possible to assemble at least 34 MHz of LOS spectrum and 56 MHz of BLOS spectrum by relying on this approach. Note that C-band is shared between LOS and BLOS; the strawman assumes that the BLOS system will operate with uplink and downlink channels at the edges of the allocated C-band spectrum (likely the 5,031–5,090-MHz band) with a central gap between the two. This central gap is necessary to avoid self-interference by BLOS users who will be transmitting and receiving simultaneously. The gap is then filled with the C-band portion of the LOS subsystem. The exact split between the LOS and BLOS spectrum at C-band is TBD and could potentially vary from one geographic region to another. Alternative bands may also be used by individual UAS – for example, VHF for LOS connectivity and L-band (using either Inmarsat or Iridium) for BLOS connectivity. But the strawman concept does not focus on these alternatives. The LOS subsystem is assumed to rely on airborne antennas which are nominally omnidirectional. However, as noted above, each airborne antenna is assumed to suffer significant shadowing and nulling due to its interaction with the airframe. Thus, at least two antennas and radios are assumed for even the smallest UA. The transmitters on the aircraft, and the transmitters and antennas at the ground stations, must be sized to provide sufficient transmit power and antenna gain to close the link against the expected fixed and variable impairments as discussed above. Engineering trade-offs among these elements (i.e., the airborne and ground transmitters and the ground antennas) are expected within the framework of the

32 Problem of UAV Communications

743

strawman design and the associated link budget analyses. Phased-array or “smart antenna” technology, with signal detection and characterization and multichannel beam-forming capability, is expected to be used in the ground stations in order to track and provide sufficient ground antenna gain to multiple UA at arbitrary locations in a given coverage sector. An additional advantage of this technology is that multipath fading on the downlink could potentially be eliminated or reduced to very small levels. Diversity techniques are heavily used in order to meet anticipated (but not yet validated) RCP requirements. The UA population is assumed to be sub-divided into two groups based primarily on size but possibly including other factors such as speed and operating regime. The first group is comprised of “smaller” UA which must satisfy a message delivery probability of 99.8 % within 1–2 s in order to satisfy presumed safety requirements. The second group is comprised of “larger” UA which must satisfy a message delivery probability of 99.999 % within 1–2 s. The first group is expected to have relatively small CNPC data throughput requirements, is not expected to operate BLOS because of their small size, and may rely almost exclusively on L-band LOS links. At least two antennas and radios are assumed to exist on each of these smaller UA in order to provide support for the necessary level of diversity to achieve 99.8 % message delivery probability with feasible transmit power and omnidirectional airborne antennas (but even more antennas and radios may be installed, if required). The second group is expected to have larger CNPC data throughput requirements as well as more stringent RCP requirements will use either LOS or BLOS (or both) and will rely on a dual-band system offering at least four-way diversity. For example, a larger UA could rely on L-band and C-band for LOS links, with at least two antennas and radios in each band (a total of at least four antennas and radios), or some combination of C-band, Ku-band, or Kaband BLOS links (for aircraft that are so equipped), so as to provide connectivity to at least two satellites simultaneously with each satellite link engineered to deliver at least 99.8 % message delivery probability (thereby achieving 99.999 % delivery probability for the diversity pair). Larger UA with BLOS capability are also expected to support LOS capability for short-range and low-altitude operation where terrestrial infrastructure exists to support them. This also partially mitigates the problem of overcoming rain attenuation at low altitude (i.e., for takeoff and landing operations), where the satellite link (if used) might encounter a significant amount of rain attenuation along the radio path. With regard to LOS ground stations, a regional or nationwide deployment is assumed in order to provide basic coverage (similar to VHF voice radio and ATC radar today) as well as to provide adequate space diversity to meet expected RCP requirements. A notional deployment is illustrated in Fig. 32.12. In this concept, 12 frequency subbands are assigned to nominally hexagonal sectors and reassigned such that co-channel interference between sectors using the same subband is minimized (note: this is for sectors that have relatively high-altitude limits; a dedicated frequency plan for smaller UA might use a “four-color mapping” instead of a “twelve-color mapping”). Within the strawman concept, the total number of ground stations is equal to the total number of sectors (an actual deployment could

744

S.B. Heppe Sector Mapping

Ground Stations

12 6

Sectors (Freqs) 3

7 1

9

3

11 8

5

7

6

8

10 4

7

6

3

11

2 12

9

1

3 8

5

4

12

1 11

10

2

10

6 9

1 11

5

3 8

2

10

Fig. 32.12 Notional sector and ground station deployment for LOS subsystem (RTCA 2011)

have multiple ground stations supporting a single sector), but the ground stations are offset from the centers of the sectors so that they coincide with half of the hexagonal vertices as shown. With this approach, each sector is served by three widely separated ground stations which guarantee the required level of space diversity to meet anticipated requirements. For example, if the white star in the figure is a UA, it is served by three ground stations each roughly “one sector radius” away. However, the reader may observe that if the UA moves slightly to the left in the figure, it will be slightly less than one sector radius from each of two ground stations, but slightly more than one sector radius from the two “next nearest” ground stations. Thus, in order to compensate for this excess path loss while still providing nominal access to at least three ground stations, the ground stations could be engineered to have slightly enhanced antenna gain relative to the baseline link budget. For example, in the link budget of Fig. 32.8, the ground station antenna gain could be increased from 8 dB to approximately 10 dB (although other engineering accommodations could be made as well). For the C-band LOS subsystem, the baseline antenna gain is substantially higher, but the same general technique would apply. While the radio communications waveform, message format, and multiple access technique are yet to be decided, the strawman assumes that the airborne radios are multichannel radios and can continuously monitor suitable pilot signals from all in-view ground stations through each available antenna and radio on the aircraft (recall that smaller UA will have at least two antenna/radio pairs, most likely at L-band, and larger UA will have at least four antenna/radio pairs, with at least two at L-band and at least two at C-band). This multi channel receiver architecture is similar to terrestrial cellular systems widely deployed today. The UA’s reception performance is continuously updated for each antenna and radio, for each ground station. This allows the UA to manage handoffs and select one of its available radios for each downlink transmission. Based on the channel statistics observed for the

32 Problem of UAV Communications

745

ground station pilot signals, the UA will associate itself with a particular sector and ground station (this could also be commanded by the ground system in some cases). This access process involves authentication of the UA in the system and assignment of a low-rate or high-rate channel (or subset of a channel) based on UA data link requirements. When a smaller UA is required to transmit, it selects the antenna/radio that has recently exhibited the best uplink performance from any of the ground stations in view, and transmits through this antenna/radio pair on its assigned channel (i.e., the one associated with its “primary” CNPC data link). In a similar vein, each ground station monitors all traffic channels in the system – not only its “assigned subset.” This is a little different from cellular systems today, but a receiver bank is relatively inexpensive, and the benefits are enormous (as will now be explained). When a ground station receives a message on a channel outside its primary responsibility, it reads the header and routes the message to the primary ground station to which it was addressed (or a suitable concentrator node that communicates with a regional grouping of ground stations). In this way, each ground station acts as a potential “diversity receiver” for each UA downlink transmission, thereby achieving the diversity goals of the system without increasing spectrum utilization (spectrum is expensive; receivers are cheap). The primary ground station (or a suitable concentrator node – which in the limit could even be the UA control station) merges the return link messages from each UA into a correlated stream by selecting the first copy of each message received and discarding later copies. The correlated stream is routed to the UACS (if the concentrator node is not the UACS itself). Return link message reception statistics are gathered at each ground station for each UA and shared among the regional ground stations or merged at a concentrator/controller node (which could be the UACS if this is the concentrator node for downlink traffic). In principle, this sharing or merging of data allows for optimum selection of a ground station for forward link transmissions to each UA (i.e., the ground station with greatest probability of message delivery). The selection algorithm has not been defined or agreed, but options include (a) deterministic use of the “primary” until a handoff is executed (this either fails to optimize message delivery probability or requires frequent handoffs which incurs its own overhead penalty), (b) selection of an optimum ground station with uplink transmission on the UA’s preassigned frequency (this requires the non-primary ground station to transmit on the primary’s channel, while the primary remains quiet) or (c) selection of an optimum ground station with uplink transmission on a channel associated with the selected ground station (this requires additional control traffic and a time delay to alert the UA that an unexpected frequency channel should be monitored for an imminent message). Each of these options has advantages and disadvantages that are yet to be assessed and traded off. For larger UA that operate at both L-band and C-band in order to achieve 99.999 % message delivery probability within the LOS subsystem, critical messages are duplicated at L-band and C-band, and parallel processes to those described above are implemented at C-band. Hence, there will be a minimum of two message transmissions (one in each frequency band) and nominally four or more received

746

S.B. Heppe

copies. It is assumed that each ground station location will host both an L-band and a C-band subsystem. But even though the ground stations host dual-frequency systems, multipath will be statistically independent between the two frequencies. UA manufacturers can also strive for statistical independence of antenna/airframe directive gain patterns as well (although the extent to which this can be achieved in practice is TBD). For wideband and less critical traffic, such as video to enhance situational awareness, and weather radar, and some low-rate control traffic, only the C-band subsystem is used by the larger UA’s. The achieved message delivery probability for these information flows is 99.8 %. By limiting video and weather radar traffic to C-band, the relatively scarce spectrum resource at L-band is conserved. The waveform and message format are assumed to include forward-errorcorrection coding which both improves message delivery probability at a given transmitter power level and offers a degree of integrity (i.e., protection against undetected error). As noted earlier and depending on RCP requirements ultimately assessed, the inherent level of integrity offered by the system can be augmented with additional CRC checks, message cross-checks, and selective message repeats (and cross-checks) in order to meet RCP requirements. A selective message repeat can be achieved in less than 500 ms, thereby preserving low latency, and slow frequency hopping from message to message will mitigate multipath fading with statistically independent fade statistics on the repeated message. Overall, the strawman architecture serves as an existence proof that stringent RCP requirements potentially levied on future UAS CNPC data links will be achievable with existing technology at acceptable cost. Excursions from this strawman, and other CNPC architectures entirely, could be considered and compared on the basis of performance and cost, as the community moves toward a consensus on a preferred approach. With regard to the BLOS subsystem that could be used by some larger UA, the strawman architecture is somewhat less complete due to uncertainty regarding eventual frequency band support. However, it is clear that 99.8 % message delivery probability on each individual satcom link can be achieved within reasonable design guidelines, for flight above certain threshold altitudes that are dependent on regional rain statistics and time of year. Hence, it is clear that larger UA could operate “at altitude” using a BLOS system for critical CNPC data. A message delivery probability of 99.8 % can be achieved with a single BLOS link, and 99.999 % can be achieved with two BLOS links (one operating at Ku- or Ka-band and the other operating at L- or C-band). Another possibility is to operate with a single BLOS system plus a LOS subsystem in high-density domestic airspace, in order to achieve 99.999 % message delivery probability in this airspace and revert to the single BLOS system in remote areas where a lower RCP requirement could be justified. It should be noted that even in domestic airspace, a hybrid LOS and BLOS system could prove beneficial for certain users. For example, a suitably designed BLOS subsystem could support CNPC data flows in a protected frequency band (AMS(R)S), along with high-data- rate payload data flows in a neighboring

32 Problem of UAV Communications

747

frequency band without special protection. Such a system, combined with an L-band or C-band LOS link for critical CNPC data flows, could prove to be particularly cost effective. Reducing Downlink Data Load: Reporting by Exception. It is desirable to minimize the data load associated with each UA – if it can be done safely – since this allows for a greater number of UA to share a given spectrum resource. As noted earlier, in many cases, the sense-and-avoid downlink data – the target tracks for potentially threatening nearby aircraft – dominate the bandwidth used. In the most basic case, each UA transmits a track report on every aircraft it can “sense,” and most of these aircraft are reported by a large number of UA. Hence, there is redundancy on the return link, and the system engineer should consider ways to minimize this redundancy. With this thought in mind, assume that the ATC ground infrastructure broadcasts a local or regional correlated surveillance picture on a “traffic information service-broadcast” (TIS-B) forward link, accessible by all UA in the airspace, and also makes this information available to all UA pilots via terrestrial means (i.e., an authenticated internet stream via the “wired ATC network”). This is a single data flow which requires much less bandwidth than the aggregate of all the sense-and-avoid return links of all the UA in the airspace. The pilot could indicate to the UA that he or she is receiving this data feed – such an indication requires only one very short forward link message every second. The UA can compare its own target track data to that contained in the correlated surveillance picture received from ATC via the TIS-B forward link. If they match within predefined accuracy bounds, for aircraft within a suitable “range of interest” from the UA, and if the UA knows that the pilot has also received this same picture, no return link reporting is necessary (except for a small heartbeat message indicating that the system is operational). If there is a discrepancy, either in an estimated target position or in the detection of a target not reported by ATC, only these differences need to be reported. Of course, if the UA fails to receive the TIS-B forward link or fails to receive an indication that the pilot has received it, the full target list would be reported. This would be a normal event in remote areas where no ground network exists. Of course, in remote areas, there are few aircraft in the airspace and little competition for spectrum resources. As may be seen, reporting by exception can significantly reduce bandwidth requirements for the system without any loss in safety (i.e., since the fallback modes are fully functional) but at the expense of greater ground system complexity.

32.6

Conclusion

In order to ensure an adequate level of safety for unmanned aircraft operating in shared airspace, the communication link between the unmanned aircraft and the ground-based pilot must carry certain safety-critical data and must be extremely robust and reliable. Initial estimates of channel loading and required communications performance have been developed, and candidate frequency bands for control and non-payload communications have been identified. While further work is needed to fully characterize the communications channel and achieve international consensus

748

S.B. Heppe

on frequency allocations and other regulatory aspects, the current consensus view is that a safe and workable implementation appears likely to be achievable through a combination of data link diversity and other error correction and error recovery techniques. The precise set of techniques that will be employed, and the extent to which those techniques will be mandated in the system and subsystem performance standards remains to be worked out.

References A general model for VHF aeronautical multipath propagation channel. ICAO/AMCP/ WG-D//WP6, Presented by Arnaud Dedryvere, Prepared by Benoˆıt Roturier and Beatrice Chateau, Jan 1999 Air traffic organization safety management system manual, Version 2.1, May 2008 CCIR, Handbook of Curves for Radio Wave Propagation Over the Surface of the Earth (International Telecommunication Union, Geneva, 1991) G. Colby et al., Test Results of the Joint FAA/DoD Investigation of GPS Interference, ION (1996) Guidance Material and Considerations for Unmanned Aircraft. RTCA DO-304 March 2007 International Civil Aviation Organization, ICAO Manual on Required Communications Performance. Document 9869 AN/462 (2008) International Telecommunications Union, Propagation data and prediction methods required for the design of terrestrial line-of-sight systems. ITU-R Recommendation P.530–12 (2007) International Telecommunications Union, Characteristics of unmanned aircraft systems and spectrum requirements to support their safe operation in non-segregated airspace. ITU-R M.2171 (2009) International Telecommunications Union, Results of studies of the AM(R)S allocation in the band 960-1 164 MHz and of the AMS(R)S allocation in the band 5 030-5 091 MHz to support control and non-payload communications links for unmanned aircraft systems. ITU-R M.2205 (2010) Next Generation Air/Ground Communication System (NEXCOM) Safety and Performance Requirements. RTCA/DO-284, 23 Jan 2003 A. Rehmann, J.D. Mestre, Air ground data link VHF airline communications and reporting system (ACARS) preliminary test report. FAA Technical Center, Feb 1995 UAS control and communications link performance – latency. RTCA issue paper SC203CC009 UAS CC Performance vD 21Apr09 (2009a) UAS Spectrum, RTCA issue paper SC203-CC011 UAS Spectrum vN 16April09, 2009, and October 2009b revisions UAS control and communications link performance – availability and continuity. RTCA issue paper SC203-CC016 UAS CC Availability vK 12Jul2011 (2011) W.J. Wilson, Strawman design for terrestrial unmanned aircraft control links, in Proceeding of the 2011 Integrated Communication, Navigation and Surveillance (ICNS) Conference (IEEE, Piscataway, 2011)

Cognitive Networking for UAV Swarms

33

Christian Wietfeld and Kai Daniel

Contents 33.1 Introduction and Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.2 Design Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.2.1 Key Building Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.2.2 Basic Principles of Context-Aware UAV Swarm Control . . . . . . . . . . . . . . . . . . . . . . . 33.2.3 Reference Scenarios and Key Performance Indicators (KPIs) . . . . . . . . . . . . . . . . . . 33.3 Modular Service Platform for Unmanned Aerial Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.3.1 Requirements for a Flexible UAS Service Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 33.3.2 Key Building Blocks of the Service Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.3.3 Validation and Key Learnings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.4 Model-Based Development of Cognitive Steerings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.4.1 Multi-scale-Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.4.2 Validation by Experiments and Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.4.3 UAV-Specific Channel Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.5 Communication-Aware Networking Algorithms for UAV Swarms . . . . . . . . . . . . . . . . . . . . . . 33.5.1 Inter-swarm Behavior Controlled by Microscopic Steerings . . . . . . . . . . . . . . . . . . . . 33.5.2 Wide-Area Mobility Controlled by Macroscopic Steerings . . . . . . . . . . . . . . . . . . . . . 33.5.3 An Integrated Steering Approach Using Communication-Aware Potential Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.5.4 Interference-Aware Positioning of Aerial Relays (IPAR) . . . . . . . . . . . . . . . . . . . . . . . 33.6 Performance Comparisons of Different Steering Algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.6.1 Aerial Sensor Network Scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.6.2 Ad-Hoc Aerial Relay Network Scenario Using Channel-Aware Potential Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.6.3 Optimal Aerial Relay Positioning Using IPAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.6.4 Analyzing the Impact of Role-Based Connectivity Management . . . . . . . . . . . . . . . 33.7 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

750 751 751 752 753 754 754 755 761 762 762 763 764 766 766 769 770 771 771 771 773 773 776 777 777

C. Wietfeld () • K. Daniel Communication Networks Institute, Faculty of Electrical Engineering and Information Technology, TU Dortmund University, Dortmund, Germany e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 32, © Springer Science+Business Media Dordrecht 2015

749

750

C. Wietfeld and K. Daniel

Abstract

A new generation of lightweight and small Unmanned Aerial Vehicles (UAVs) enables the design of networked aerial robotic systems for a wide range of applications. In this book chapter challenges and solution approaches with respect to self-organized and robust multi-UAV systems are discussed. In order to achieve a flexible deployment of Unmanned Aerial Systems (UAS), a novel service-oriented system architecture is introduced. The design aspects of mobility control algorithms addressing the competing requirements of communication reliability and spatial distribution of UAV swarms are demonstrated using two reference scenarios with diverging requirements: aerial sensor networks and ad-hoc aerial relay networks. The novel solution approaches presented in this chapter rely on agent-based control of UAV swarms operating autonomously in dynamically changing environments. Different communicationaware algorithms for microscopic as well as macroscopic mobility control are presented, such as Cluster Breathing, Smart Cube, Potential Fields, and Role-Based Connectivity Management. For the positioning of UAV relays, the Interference-Aware Positioning of Aerial Relays (IPAR) algorithm is introduced. The system design and optimization of cognitive networking for UAS requires a dedicated multi-scale simulation environment, which includes a detailed physical channel model and real-world experiments. The chapter discusses how the network and application task-specific performance requirements are met with the proposed mobility control algorithms even in the cases of temporarily unavailable communication links. The self-healing capabilities therefore allow for reliable networking and control of aerial robot swarms in diverse use cases, such as emergency response, environmental monitoring, and ad-hoc network provisioning.

33.1

Introduction and Related Work

The advances in embedded systems technologies have enabled broad research and first deployments of networked sensors/actors to fulfill monitoring and control tasks (Akyildiz et al. 2002). While many of these cyber-physical systems incorporate mobile components and therefore rely on wireless communications, they are typically ground-based. Hence, they are limited in their scope of operation and movement. The research in the area of networked aerial robotic systems focuses on enabling teams of communicating UAVs (Michael et al. 2006; Frew et al. 2008) to autonomously fulfill a given task (How et al. 2009; Chung et al. 2011), such as the exploration of chemical or nuclear plumes (Spears et al. 2009; White et al. 2008; Daniel et al. 2009; Corrigan et al. 2008). In addition to exploration tasks, teams of flying robots may also provide an ad hoc relay network to serve ground-based users in case of network overload or outage situations (Cheng et al. 2008; Pengcheng et al. 2011).

33 Cognitive Networking for UAV Swarms

751

In contrast to many ground-based networking applications in which the mobility of the communication nodes is mandated by the application scenario, the mobility of aerial robot swarms can be actively influenced. These controlled mobility algorithms aim to meet both application task requirements and communication-related performance indicators (Wang et al. 2006; Rooker et al. 2007; Stump et al. 2008; Akkaya et al. 2010; Ghaffarkhah et al. 2011). While exploration tasks may, for example, suggest spreading out in different directions, the need for continuous data transfer may rather demand for a behavior that keeps the swarm close together. Given a known and static environment, it is possible to steer the swarm with preplanned schemes, which fulfill the competing goals. However, in many real-life scenarios, the swarm needs to be prepared for the unexpected, in particular when disaster management applications or component failures need to be considered. To enhance the robustness and flexibility of the robot swarm, current research seeks to distribute the control intelligence across the different platforms, leading to an autonomous swarm behavior controlled by software agents implemented on each UAV (How et al. 2009; Chung et al. 2011). This book chapter provides an insight into promising new solution approaches for cognitive networking capabilities to solve the trade-off between application taskand communication-related goals. As a specific contribution of this book chapter, cognitive networking algorithms are discussed as part of a generic service platform concept supporting diverging UAS application services. In the following chapter, the key design challenges in terms of building blocks, reference scenarios, and key performance indicators are presented. Then a new modular service platform for UASs is presented, which leverages embedded Web services technologies to allow for the flexible deployment of multipurpose UAV swarms. Before cognitive networking algorithms are discussed in detail, a model-based development approach, which encompasses analytical modeling, multi-scale system simulation as well as experiments, is introduced. Finally, various new communication-aware mobility steerings are presented and discussed based on performance evaluation results.

33.2

Design Challenges

33.2.1 Key Building Blocks The generic design of a swarm-based aerial robotic system considered in this book chapter (cf. Fig. 33.1) consists of UAVs that are equipped with sensors/actors and act as mobile communication nodes at the same time. Depending on the use case, the aerial robotic swarm connects via base stations to ground-based devices, such as mission control entities, databases, and network gateways. In this chapter the focus lies on the air-to-ground (A2G), air-to-air (A2A), and air-to-user (A2U) links, while the choice of the back-end ground connections is left out of scope. Different communication technologies for A2G, A2A, and A2U links can be considered: • Local area networks (in particular IEEE 802.11a/b/g/n, p and s) • Cellular networks such as HSPA, IEEE 802.16e, and LTE

752

C. Wietfeld and K. Daniel Agent-controlled UAV equipped with sensors

Networked robot swarms with self-healing capabilities

Agent-controlled UAV equipped with relay (e.g. LTE)

A Air-to-Air Movement

Ai Air-to-Airr Interfering Cellular Base Station

Air-to-Ground Plume

Air-toUser

Serving Cellular Base Station Sensor Data Sink

Aerial Sensor Network

Network Gateway

Users Ground Network Coverage

Movement

Ad-hoc Aerial Relay Network

Fig. 33.1 Architecture and reference tasks for networked aerial robot swarm

The primary task of the communication network is to transfer application taskrelated data, namely, sensor data or user data. At the same time, the swarm control data required to steer the members of the swarm are communicated via the network. To achieve maximum robustness, aerial communication nodes may act as relays in order to provide multiple hop connections and meshing capabilities, as described in Lloyd et al. (2007) and Jung et al. (2010).

33.2.2 Basic Principles of Context-Aware UAV Swarm Control Due to the dynamically changing, unpredictable system environment leading to potentially unreliable wireless communication links, an autonomous behavior of the swarm is desired. The agent-based control of the robot swarm proposed in this book chapter can be distinguished in two main steering types: • The microscopic steering controls the behavior of the individual UAV so that a coherent, interconnected swarm is maintained (also called cluster). At the same time, the members of the swarm should not interfere with each other and achieve a maximum impact related to the given swarm task. The microscopic steering is of high importance for the A2A links, as it will self-configure and determine the availability of communication links within the swarm. • The macroscopic steering controls the swarm according to a given task through a larger-scale scenario, namely, to explore an unknown area or to move from point A to point B. The macroscopic steering will typically be the main influencing factor for the availability of A2G and A2U links. The key contribution of this book chapter is that both steering types are combined by an autonomous, agent-based decision process on each UAV. Thereby the robot swarm can perform its application task independently of a central ground station. Self-healing capabilities allow that UAVs, which are cut off from the swarm due to obstacles and channel impairments, will automatically reintegrate into the swarm.

33 Cognitive Networking for UAV Swarms

753

33.2.3 Reference Scenarios and Key Performance Indicators (KPIs) To illustrate the concrete challenges and performance indicators of cognitive networking in aerial robot swarms, two application areas are introduced: the exploration of tropospheric plumes as an example for a monitoring task and the ad hoc network provisioning for outage compensation in terrestrial networks (cf. Fig. 33.1).

33.2.3.1 Scenario 1: Plume Exploration In the case of a chemical or nuclear accident, disaster managers require information about the current and future development of a plume to decide on potential counter measures for protecting inhabitants and rescue personnel. The aerial sensor system should provide as fast as possible, but not necessarily in real-time, accurate information about the concentration of harmful substances. This chapter discusses in particular the following key performance indicators: • Spatial Exploration Ratio (SER): The SER is determined by dividing the scenario in a regular spaced grid. If a grid element is accessed by a UAV and once the information was successfully transferred to the data sink, this grid element is counted as being covered. The sum of all covered grid elements divided through the total number of grid elements delivers the SER value. • Cluster Separation Ratio (CSR): Among the variety of communication-related parameters to be considered, the CSR value serves as an indicator to assess the communication performance: the time span, in which one or more UAVs are not connected to the main part of the swarm (the so-called cluster), is determined and related it to the overall mission time. The key challenge for the control of the aerial robot swarm in this scenario is to allow for continuous communication within the swarm by meeting a minimum CSR target (e.g., CSR < 1%) while at the same time as much space as possible is explored (indicated by a high SER value).

33.2.3.2 Scenario 2: Ad-Hoc Aerial Relay Network The handling of temporary network outages or overload situations is a major challenge for network operators. Aerial network provisioning is a very interesting option to quickly provide ad hoc network coverage, even if parts of the ground-based infrastructure are not operational or overloaded. The considered key success criteria for the networked swarm are: • Ground Network Coverage (COV): The COV is determined by cumulating all coverage areas of each individual UAV excluding overlaps, obstacles, and isolated UAVs that are not connected to the swarm. • Throughput (TP): The amount of communication traffic handled by the system in terms of TP is determined by summing up the raw data rates of all individual A2U communication connections within the swarm.

754

C. Wietfeld and K. Daniel

Table 33.1 Overview of challenges and proposed solution approaches Challenges and design New channel-aware swarm control algorithms Scenario considerations Microscopic Macroscopic Aerial sensor network

Ad-hoc aerial relay network

Trade-off between CSR and SER: loose connectivity among the UAVs leads to a high SER, but at the same time the CSR may grow too high Trade-off between TP and COV: a stable TP can be achieved by reducing the COV to adapt to bad channel conditions

Cluster Breathing (CB), Smart Cube (SC), C.-A. Potential Fields (CAPF)

Cluster Repelling Walk (CRW), Role-Based Connectivity, Return and Release

Communication-aware Potential Fields (CAPF) Interference-aware positioning of relays (IPAR)

The scenario-specific challenge will typically be to achieve a minimum throughput to match the requirements of specific communication service (e.g., TP > 5 Mbit/s) while at the same time the COV value should be maximized to serve as many users as possible. Table 33.1 summarizes the challenges and design considerations addressed in this book chapter. It provides an outlook of the corresponding solution approaches discussed later on. Before the proposed swarm control algorithms are discussed, the overall system architecture is introduced in the next section.

33.3

Modular Service Platform for Unmanned Aerial Systems

The research for context-aware communication within Unmanned Aerial Systems (UAS) has focused not only on the mobility control algorithms (see in subsequent sections) but also on the development of a generic and modular service architecture, which supports heterogeneous, multipurpose UAS. In the following section the requirements and solution approach for a generic UAS service platform are introduced.

33.3.1 Requirements for a Flexible UAS Service Architecture In order to serve different use cases, an Unmanned Aerial System should build on a flexible software and communication architecture. Several architecture approaches have been presented and implemented to address multi-sensor, multi-networktechnology and multi-vehicle systems for unmanned autonomous systems (Tisdale et al. 2006; Dias et al. 2008; Love et al. 2009; Elston et al. 2009, 2011). As an

33 Cognitive Networking for UAV Swarms

755

evolution and alternative to these approaches, in this chapter an architecture building on embedded Web services technologies to ease the integration of highly diverse services beyond sensor tasks is introduced. While service-oriented architectures (SoA) are state of the art for IT services, in the embedded and real-time control systems area, SoA architectures are still an emerging topic. The key benefit of adopting the SoA concept for embedded systems is that the complexity and heterogeneity of different hardware and software components is hidden by well-defined Web Services and clearly structured XML data formats. This allows for the reuse of software components and flexible integration of new components. Recent developments have overcome the concern regarding the overhead introduced by SoA for resource-constraint systems. Prominent examples for the adoption of SoA concepts for embedded systems are the Device Profiles for Web services (DPWS) (Discroll et al. 2009) as well as Efficient XML (EXI) (Schneider et al. 2011). Research projects such as the European MORE project (Wolff et al. 2007) have shown the feasibility of using DPWS for geo-monitoring services and remote control of embedded devices. Other areas, in which SoA-based embedded Web services gain importance, are distributed energy systems and Smart Grid control (Wietfeld et al. 2011). The design of a future-proof UAS service platform should therefore leverage the concept of a service-oriented architecture while at the same time the real-time constraints of Unmanned Aerial Systems shall be taken into account. The service architecture should be flexible to support diverse aerial services ranging from sensor-based monitoring to communication relaying. Furthermore, different types of Unmanned Aerial Vehicles (quadcopters, fixed wing, tiltwing) as well as different wireless networking technologies (ranging from Wi-Fi to public cellular networks) have to be taken into account for the system architecture design. Appropriate system capabilities are required to allow for autonomous operation in unknown environments and for providing robustness against component and node failures. End users shall be able to operate the system with minimum training, allowing them to focus on the application task to be fulfilled. Therefore, automatic self-configuration and health management shall be supported. Finally, the implementation needs to comply with real-time, safety and security requirements. Hence, state-of-the-art, highly efficient Web services–based concepts, such as optimized IP transport, DPWS, and EXI, are considered to meet the resource constraints of embedded HW/SW platforms. Within the research project AVIGLE (Rohde et al. 2010), an example UAS service platform fulfilling these requirements has been designed, implemented, and validated by real-life experiments. Key building blocks addressing these requirements are introduced in the next section.

33.3.2 Key Building Blocks of the Service Architecture As a proof of concept of the flexibility of the architecture, two diverging UAS use cases are considered in the following: a 3D Visual Exploration Service (as an

756

C. Wietfeld and K. Daniel

Service Control Air (SCA)

Air-to-Air Comm.

Service Control Air

Cognitive Mobility Control

Air-to-Ground Comm.

Task Assignment & Mission Planning

Service Control Ground (SCG)

3D Visual Exploration (3DVE)

Aerial Radio Access Network (ARAN)

…

Application Services

Focus of this chapter

Fig. 33.2 Key components of UAS service platform

example for a generic Sensor Exploration Service) and an Aerial Radio Access Network Service. These reference application services are linked with two core UAS control services groups: the Service Control Ground (SCG) and the Service Control Air (SCA); see Fig. 33.2. The core services provide the generic control functionality to operate the UAS independent of a specific application task, such as navigation services. In addition, communication services are an essential part of the system architecture: on one hand local buses provide onboard communication (e.g., intra-SCA), while wide-area wireless communication services allow for the inter-UAV communication as well as the link between UAVs and the ground-based functionalities. It should be mentioned at this stage that the architecture is worked out in this chapter in great detail for Unmanned Aerial Systems, but it is flexible enough to cater also other types of autonomous vehicles, such as ground-based or underwater vehicles. In the following, the key services of a UAS are introduced in detail.

33.3.2.1 Service Control Air (SCA) The Service Control Air (SCA) provides all functionalities to enable the operation of multiple, even heterogeneous UAVs within a UAS. A basic assumption is that each UAV has implemented specific flight control mechanisms, which allows it to be steered by the SCA to a specific position. Therefore, the detailed local control

33 Cognitive Networking for UAV Swarms

Multicopter Tiltwing

System Monitoring & Logging Service (SMLS) Geo-Positioning Service (GPOS)

GPS GALILEO

Mobility Control Service (MCS)

Flight Control Access Service (FCAS)

…

757

Obstacle Map Service (OMS) Service Control Air (SCA)

UAV Information Dispatch Service (UIDS)

Radio Propagation Map Service (RPMS)

CAPF IPAR …

Scout Intermediate Relay

Dynamic Role Mgmt. Service (DRMS)

…

…

802.11a .. s

Mobile Comm. Mgmt. Service (MCMS)

Payload Management Service (PMS)

Comm. Relays

UMTS / LTE …

Sensors

Service Control Ground (Air-to-Ground)

Other SCAs of AUS (Air-to-Air)

… User Equipment (Air-to-User)

CAPF-Communication-Aware Potential Fields IPAR–Interference-Aware Positioning of Aerial Relays Focus of this chapter

Fig. 33.3 Overview of Service Control Air (SCA)

algorithms and components, such as the individual control of the UAV engines, are out of scope for the SCA. The following services are essential parts of the SCA (cf. Fig. 33.3): • In line with the idea of service-orientation, the complexity of the UAVs’ capability to stay in the air and fly into a given direction is hidden through a corresponding Flight Control Access Service (FCAS). The parameters of this service are either global geographical positions or directions. • The main user of the FCAS is the Mobility Control Service that implements different strategies to achieve predefined performance indicators. The mobility control of a UAS is context depending, while the context is defined by the application task, the communication requirements within the system, and the physical environment (topology, weather, etc.). The in-depth discussion of the appropriate design of the Mobility Control Service is covered in subsequent sections, as different applications task demand for specific Mobility Control Services. Example algorithms are the Communication-Aware Potential Fields (CAPF) and the Interference-Aware Positioning of Aerial Relays (IPAR). • Another essential service is the Payload Management Service (PMS), which provides the access to different components carried by the UAV to fulfill a given application task. Example payload components are optical sensors (such as infrared cameras), environmental sensors (such as gas or radioactivity detectors), and wireless communication relays, to allow for ground coverage and provisioning of A2U links.

758

C. Wietfeld and K. Daniel

• Via the Payload Management Service, component-specific software libraries are integrated. The modular approach allows for the integration of future payloads, such as new sensor types or robotic components. • The Geo-Positioning Service (GPOS) provides accurate information about the current position of the UAV. This information is required both for mobility control as well as for application services, which link sensor data with geo-positions. This function will typically leverage satellite-based positioning technologies, such as GPS, GALILEO, and combinations thereof. The accuracy of the positioning information provided by the existing Global Positioning System (GPS) as well as the future GALILEO system is typically in the range of several meters which is sufficient for most application scenarios (Nieh¨ofer et al. 2011). Due to the inaccuracies of satellite-based positioning, additional sense-and-avoid capabilities are required to reliably avoid collisions with obstacles and within the swarm. If needed, additional positioning technologies can be integrated via the GPOS (e.g., for indoor usage). • The Dynamic Role Management Service (DRMS) manages the UAV’s application task capabilities and the task assignment. Through this service available capabilities are offered to the system, while at the same time specific task assignments within the overall UAS may require the deactivation of certain components. For example, a resource-consuming sensor can be switched off, in case the UAV serves only as a communication relay at a given point in time. The Dynamic Role Management Service therefore interacts with the Payload Management Service and the Mobility Control Service. • Finally, Mobile Communication Management Services (MCMS) are available to allow for the transmission of UAS internal control data as well as payload data, such as pictures, videos, or ground user traffic. Similarly to the Payload Management Service, different wireless communication technologies can be integrated via the service-oriented approach. The MCMS allows integrating mobility management functionality to realize handovers and dynamic mesh routing. The CMS interacts with the MCS as mobility control often needs to take into account communication-related parameters (such as Radio Signal Strength parameters). Throughout the system operation, the CMS may gain information to feed the Radio Propagation Map Service. The Mobility Control Service is supported by additional services within the SCA, such as: • Obstacle Map Service (OMS): provides geo-referenced information about the physical characteristics of the environment for proactive, wide-area collision avoidance, in addition to near-distance sense-and-avoid systems of the UAV. The Obstacle Map Service may use either stored map data as well as dynamically generated data gathered by the UAS during exploration tasks. • Radio Propagation Map Service (RPLS): provides geo-referenced information about the channel characteristics of ground base stations for proactive planning to avoid communication outages between UAVs and the ground station. Similar to the above mentioned Obstacle Map Service, pre-stored data as well as dynamic data can be used.

33 Cognitive Networking for UAV Swarms

759

Support functions such as a System Monitoring and Logging Service (SMLS) as well as an Information Dispatch Service (IDS) ensure the reliable operation of the system.

33.3.2.2 Service Control Ground (SCG) The Service Control Ground (SCG, cf. Fig. 33.4) is the counterpart of the Service Control Air and therefore contains various corresponding services, such, the Mobile Communication Management Service (MCMS), System Monitoring and Logging Service (SMLS), UAS Obstacle Map Service (UOMS), UAS Radio Propagation Management Service (UPMS), and UAV Payload Management Service (UPMS). While the Service Control Air will be present in multiple instances (one per UAV), the standard system will comprise one SCG instance. The UAV Payload Management Service serves as global repository for the different service capabilities of the various UAVs within the UAS. Due to resource constraints in the UAV, the aerial instances of the map services OMS and RPLS will contain a subset of the information handled in the ground-based instance (UOMS and UPMS). For robustness, the Service Control Ground can be realized in a redundant, physically distributed way, but this consideration is out of scope in this presentation of the architecture. There are a number of services, which are specific to the ground-based system, compared with the aerial control components:

Service Control Airs of Swarm (Air-to-Ground)

802.11a...s UMTS / LTE

Mobile Comm. Mgmt. Service (MCMS)

Dynamic Task Assignment Service (DTAS)

System Monitoring & Logging Service (SMLS)

Mission Preplanning & Safety Service (MPPS)

… Swarm Information Dispatch Service (SIDS)

Service Control Ground (SCG)

UAS Obstacle Map Service (UOMS) UAS Radio Propagation Map Service (UPMS)

3DVE Service.

UAV Payload Mmgt. Service (UPMS)

Application Service Registry (ASR)

Comm. Relays

ARAN Service … 3DVE– 3D Visual Exploration ARAN– Aerial Radio Access Network

Optical Sensors

UAS Application Services

Focus of this chapter

Fig. 33.4 Overview of Service Control Ground (SCG)

…

760

C. Wietfeld and K. Daniel

• The Dynamic Task Assignment Service (DTAS) provides a user interface toward the end user and allows for the specification of the application details to be fulfilled by the UAS. This service is context sensitive, that is, it must consider the payload components of the UAVs, the flight capabilities of the UAVs, the environmental conditions, etc. • The DTAS interacts closely with the Mission Pre-Planning and Safety Service (MPPS), which will – based on the task requirements specified by the end user and the available UAV characteristics – breakdown the overall in tasks to be fulfilled by the UAVs’ Mobility Control Service. The MPPS leverages information provided by OMS and RPLS. • The Application Service Registry (ASR) manages the interaction with different application services of the UAS. Different application service will require access to different payload components and mobility control strategies. The ASR therefore interacts with the UPMS, DTAS, and MPPS in order to match the application task requirements with the available service capabilities. • The Swarm Information Dispatcher Service (SIDS) efficiently routes any data between the ground-based services and the aerial components, may it be control information or payload data. It comprises scheduling and prioritization algorithms to master potential overload situations.

33.3.2.3 Example UAS Services In the following, two service types in line with the reference scenarios defined in Sect. 33.2 are introduced (cf. Fig. 33.5). 3D Sensor Exploration Service (3DSE) The 3D Sensor Exploration Service (3DSE) serves as an example for a distributed aerial sensor network application. The UAVs are equipped with sensors, such as

Service Control Ground

Service Control Ground

3D Point Generation

3D Mesh Generation Gas Sensor Analysis

3D Sensor Exploration Service (3DSE)

Texture Generation

Interference-Aware Dynamic Network Planning

VoIP Gateway

Visualization

….

Network Virtualization

Secure 3rd Party Network Interworking Situational Awareness System

AAA

Aerial Radio Access Network (ARAN)

Push-to-X Services

Internet Gateway ….

Focus of this chapter 3rd party networks

Fig. 33.5 Example application services for communication-aware UAV operations

33 Cognitive Networking for UAV Swarms

761

gas sensors or cameras as a payload, and deliver geo-referenced gas sensor data or aerial photographs of the earth’s surface from different perspectives. This service may comprise very specialized functionality, such as 3D point generation, 3D mesh generation, and texture generation, in order to finally visualize the results for the users of the service (visualization). A detailed description of the image processing and virtualization approach is provided by Rohde et al. (2010). An example for an aerial gas sensor application is described in Daniel et al. (2011). Potential users of the 3DSE are rescue organizations, which require detailed information of a city environment after an earthquake or after a large-scale chemical incident. Using an up-to-date 3D map as part of situational awareness systems allows efficient planning and optimization of rescue actions. But the results of this service can also be leveraged within the UAS itself, because it provides current data for the mission planning through the obstacle and radio propagation services.

Aerial Radio Access Network Service (ARAN) The Aerial Radio Access Network Service (ARAN) provides ad-hoc ground network coverage by aerial relays. It requires communication relays as payload on the UAVs’ side and specific mobility control algorithms. Examples for standard functionalities are AAA services, a VoIP gateway, and an Internet gateway. UAS-specific service functionalities are related to: • Interference-Aware Dynamic Network Planning which needs to take into account both coverage requirements as well as interference mitigation. This aspect is addressed in subsequent sections. • Push-to-X services which are required to support IP-based, multimedia group communications. These services are of specific importance for ad-hoc communication of emergency response organizations (Subik et al. 2011). • Secure Third Party Network Interworking is required to integrate the Aerial Radio Access Network with ground-based networks, such as a public cellular network, which shall be complemented with the aerial network to compensate network outages. • Network virtualization functionalities are required in case the aerial network shall be used by multiple network operators. In this case, the available system capacity needs to be shared amongst the different networks based on service contracts. Therefore, scheduling and monitoring of Quality of Services parameters are part of this building block. The network virtualization is subject of ongoing research and out of scope for this chapter.

33.3.3 Validation and Key Learnings The implementation of the proposed architecture has been carried out on embedded platforms (such as Vortex86DX and ARM Cortex-A8) and validated for flexibility and efficiency in experiments (see below). It has shown the embedded Web services approach is very suitable for the integration of services, such as different sensor types or communication relays. Nevertheless, in case of specialized sensors (such

762

C. Wietfeld and K. Daniel

as IR cameras), the implementation of software connectors to integrate proprietary drivers with the generic architecture requires significant effort. Therefore, the benefits of a Web services-based approach can be only fully leveraged, when the suppliers of the UAV payload components will provide a Web service compliantinterface off the shelf. Due to the usage of highly optimized binary EXI coding, the actual size and processing times of data messages to be exchanged within the system were similar to a comparable, non-Web services-based implementation. For real-time critical parts of the system, the data message structures have been hardcoded in order to avoid encoding delays. In summary, the embedded Web services approach has proven to be feasible for UAS, as it provides a framework for a common, well-established formal description of service characteristics and data formats and at the same time allows to trade flexibility with efficiency by using highly efficient data encoding techniques.

33.4

Model-Based Development of Cognitive Steerings

33.4.1 Multi-scale-Simulation Environment For the performance evaluation of cognitive UAV networks and the corresponding steering algorithms, it is essential to consider multiple system components and characteristics (cf. Fig. 33.6): • The mobility model delivers the dynamically changing position of the UAVs. • The wireless channel model’s output – the Radio Signal Strength – is derived from the current position of the UAV taking into account a realistically modeled

UAV Mobility Model POS

Sensor Model

POS

DIR

UAV Channel Model

Fig. 33.6 Example information flows in multi-scale system simulation environment

OMNET / NS-3

DATA UAV Protocol Model

RSSI

MATLAB

COV

COV DATA DIR POS RSSI

Context-Aware Control Algorithms

Grid Cell Coverage Info Payload + Error Info Steering Direction Current Position of UAVs Radio Signal Strength Indicators

33 Cognitive Networking for UAV Swarms

763

system environment (topology, obstacles, interferers, antenna characteristics, etc.). • The sensor model deployed with the UAV allows to determine the spatial coverage of the sensor based on its current position. In case of a plume exploration task, it is assumed for simplification that once the UAV has entered a grid cell, sensor measurement values for the complete grid cell are instantly available. • The communication protocol models deliver detailed information about the timing and length of data transmissions. To address these different system aspects, a multi-scale simulation environment is required, which combines various, specialized simulation models and real-life HW/SW implementations into a complete system model (Daniel et al. 2011). Figure 33.6 provides an overview of major information flows between the various subsystem models and the control algorithms to be investigated. A special focus of the performance analysis is placed on the dynamic meshing capabilities, as short-life routing paths have to be reconfigured and re-established as fast as possible to avoid separations within the aerial swarm network. Subsequently, MANET and VANET protocols have been integrated into the simulation environment to allow the analysis of the interdependencies among channel, mobility, and meshing capabilities. With regard to an appropriate trajectory calculation, a UAV-specific kinematic movement model is utilized: while the agent-based steering control algorithms mandate desired future positions of the UAV, the physical characteristics (mass, acceleration) need to be taken into account to achieve realistically modeled swarm behavior. The system component models are developed and validated based on dedicated experiments (e.g., channel measurements), which are introduced in the next section.

33.4.2 Validation by Experiments and Measurements Although the multi-scale simulation models provide a powerful and realistically modeled environment to evaluate the system, experiments are essential to validate both dedicated system component models as well as the overall system performance. In the research projects AirShield (Daniel et al. 2011) and AVIGLE (Goddemeier et al. 2012), different types of UAVs, ranging from multi-copter to tilt-wings and blimps, are deployed to validate mobility control algorithms. Swarm control data (telemetry) and payload-related data (gas sensor values) were successfully transmitted and routed over a Wi-Fi (A2A mesh) and UMTS-HSPA (A2G) network (cf. Fig. 33.7). With respect to the relevance of the mesh protocol for the air-to-air links, several experiments in an EMC anechoic room have been conducted and analyzed with regard to the 802.11 performance in an interfered environment (Daniel et al. 2010). Furthermore, the impact of interferences on the reliability of air-to-ground links has also been analyzed for LTE and Mobile WiMAX by means of different network and channel emulators (e.g., R&S CMW 500, EB Propsim C8). The air-to-ground

764

C. Wietfeld and K. Daniel

Gas sensors

A2A links A2G links

Embedded PC

Communication Subsystem

Fig. 33.7 Example experiments carried out in the AirShield project: validation of A2G and A2A links (left), validation of sensor data transfer (middle), gas sensor and communication payload (right)

availability of cellular networks in higher altitudes (up to 500 m) has been surveyed with a captive balloon in urban and suburban environments. Since the channel characteristic is a crucial system property for UAV networks, results of recent studies (Goddemeier et al. 2012) are presented in the following section.

33.4.3 UAV-Specific Channel Models The UAV channel is influenced by numerous factors such as antenna characteristics, attenuation, multipath propagation, and interference. Extensive work has been performed on ground-based wireless systems as well as satellite systems, which can be partly applied for the UAV channel. However, the UAV channel bears certain specifics due to its different geometry. The air-to-air (A2A) channel is characterized by a dominant direct path and additional interference due to reflections from the ground (Allred et al. 2007). The characteristics of the air-to-ground (A2G) channel are highly dependent on the used communication technology in use. Leveraging publicly available cellular networks are useful to provide wide-area air-to-ground coverage without additional infrastructure. Therefore, a dedicated UAV channel model for A2G links is of specific interest for the system modeling. Base station antennas in cellular network systems usually show directional characteristics providing reliable ground coverage for the end users. Figure 33.8 depicts the specific characteristics of the height-dependent UAV channel, based on measurement data, ray-tracing results, and a specifically developed analytical channel model (Goddemeier et al. 2012). The channel can be characterized by 4 zones: at low heights up to 10–20 m, shadowing effects are observed (Zone I), while in heights up to approx. 50 m, a strong direct path interferes with reflections of the ground respectively the roof tops (Zone II). At heights above 50 m, the UAVs antenna leaves the main antenna lobe and a reflected path from

33 Cognitive Networking for UAV Swarms

765

−60 Raytracing

RSSI[dBm]

−80

Measurements

New UAV-specific analytical channel model

−100

I −120

II 50

III

IV

150

250

350

height [m] height IV

III

II I

distance

Fig. 33.8 UAV channel characteristics: air-to-ground link

the ground dominates (Zone III). Finally, for altitudes above 300 m, the received power for direct paths as well as reflected paths decreases. Consequently, no reliable communication is possible (Zone IV). The results illustrated in Fig. 33.8 have been validated with channel measurements in real-life UMTS/GSM networks in Germany

766

C. Wietfeld and K. Daniel

by utilizing a captive balloon (h < 500 m). The key conclusion from this synopsis of a far more detailed analysis (Goddemeier et al. 2012) is that the exploration at heights above approx. 250 m requires communication-aware aerial relay nodes with meshing capabilities to be able to extend the operational range by establishing multihop communication links.

33.5

Communication-Aware Networking Algorithms for UAV Swarms

The networked UAV systems discussed in this book chapter rely on autonomously controlled mobility. Recent research has come up with several mobility control strategies which have often been inspired by nature or found an illustrative counterpart in nature. An in-depth survey of connectivity-dependent algorithms to control the position of mobile sensor nodes is provided in (Younis et al. 2008). Examples of relevant work related to robot mobility control have been presented by Wang et al. (2006), Rooker et al. (2007), Stump et al. (2008), Elston et al. (2009), Han et al. (2009), Akkaya et al. (2010), and Ghaffarkhah et al. (2011). This chapter focuses in particular on algorithms addressing the interdependency and desired equilibrium between coverage and connectivity requirements in UAV swarm missions.

33.5.1 Inter-swarm Behavior Controlled by Microscopic Steerings For the distributed control of a swarm, in Reynolds (1987) three basic rules are established: separation, alignment, and cohesion. Based on the physical distances between neighbors, the behavior of the swarm is controlled to avoid collisions and to ensure harmonized movements for the computer-based modeling of biological swarms and herds. This fundamental work has inspired also the research on controlled mobility for robot swarms; see, for example, Cao et al. (1997) and Ryan et al. (2004). In the context UAVs performing jointly an exploration task (see scenario 1), the coverage of the swarm shall be maximized (leading to separation) while at the same time connectivity between the UAVs needs to be preserved (leading to cohesion). In Daniel et al. (2011) a corresponding channel-quality-aware mobility control algorithm, which introduces the so-called Cluster Breathing (CB) rules, has been proposed and studied. This algorithm considers the received power from neighboring vehicles, the radio Signal Strength Indicator (RSSI) as space and time-variant communication parameter (cf. flow chart shown on top of Fig. 33.9). As the RSSI will vary due to the dynamically changing channel characteristics, the swarm will “breath” to a certain degree. Similar to other highly dynamic control processes in mobile radio networks (such as power control or handover decisions), the impact of short-term fading must be compensated appropriately, for example, by introducing medium-term averaging of RSSI values and hysteresis. To achieve a better spreading of the UAVs when fulfilling a swarm task, such as a plume

33 Cognitive Networking for UAV Swarms

767

1 Measure RSSI to neighbour UAVs

Start

RSSI > RSSI_flee 2

no

yes

Flee: Move away from neighbour UAVs

−70

yes

Seek: Move towards neighbour UAVs

no RSSI – Measured Received Signal Strength Indicator RSSI _ opt–target RSSI value RSSI _ flee–RSSI threshold leading to flee movement RSSI _ seek–RSSI threshold leading to a seek movement

Position Hold relative to neighbour UAVs

300

1.0 improvement of channel quality

1

−80

0.75

2

−85

RSSI

flee drop in channel quality 0

t=0 1

flee

100 2

swarm volume t=50s

20

40 Time [s]

0.5 60

UAV

seek

200 y [m]

−75

−90

flee

seek

normalized swarm volume

RSSI of example UAV connection [dBm]

RSSI < RSSI_seek

0

0

100

2D presentation of 3D mobility behavior 200

300

x [m]

Fig. 33.9 Example of cognitive steering due to changing channel characteristics (Cluster Breathing): Flow Chart, RSSI vs. Swarm Volume Chart, Spatiotemporal 2D Chart Of Swarm Behavior

detection, the number of required neighbors can be limited, for example, to two or three. Figure 33.9 shows an example of the behavior of an aerial robot swarm: after the robots have been separated according to the given RSSI boundaries, an area with stronger impact of fading is entered at 10 s (1), which leads automatically to a reduction of the distances between the robots respectively of the volume created by the network nodes. With improved channel quality (2), a flee movement is initiated, until the target RSSI value is reached. Despite careful selection of the control parameters, it might nevertheless happen that the connection to one or more UAVs is completely lost. These situations require cluster-self-healing capabilities: one option is, for example, that the UAVs of the swarm move back to the position, where the connection was lost. In some cases, it might be intended to split up the swarm for a given time to fulfill the swarm task more efficiently. This special situation is discussed below in more detail (cf. Sect. 33.6.4). The Cluster Breathing (CB) steering as described above is able to meet the communication-related goals on a microscopic level, but has the drawback that it does not exploit the full potential regarding the local exploration coverage of the swarm: the UAVs within the swarm will not or very rarely move into the edges of a cube, even if this would be possible from a communication perspective.

768

C. Wietfeld and K. Daniel

A static configuration can be ruled out, as it will not be possible to react on the dynamically changing physical channel characteristics. Consequently, a more advanced algorithm has been developed, namely, the Smart Cube algorithm. The Smart Cube (SC) (Daniel et al. 2011) approach aims to optimize the 3D distribution of the UAV agents within a given cube by using Voronoi tessellation (Ammari and Das 2010) among other techniques. This algorithm provides a better local coverage in case of spatial exploration tasks. To cater for RSSI variations, the Smart Cube concept adapts the size of the cube to be covered dynamically according to current physical channel conditions. However, the performance evaluation of larger-scale scenarios has shown that the overall performance of given exploration tasks is strongly impacted by the macroscopic behavior, which might efficiently compensate for deficiencies in the local coverage of the swarm. In order to maintain the connectivity of the UAV swarm, the maximum distances between the nodes needs to be limited. Instead of using distances the Smart Cube uses a virtual cube which size is determined by an RSSI threshold (cf. Fig. 33.10a)

a

b virtual cube@ −80 dBm

ClusterCentroid

c

d

Fig. 33.10 Achieving better spatial distribution through Smart Cube algorithm. (a) Initial state. (b) Voronoi tesselation. (c) Lloyd movement. (d) Improved topology

33 Cognitive Networking for UAV Swarms

769

for achieving communication awareness. Within this cube a Voronoi tessellation is done by each agent (cf. Fig. 33.10b). On basis of the Voronoi tessellation, the spatial volume difference between the single Voronoi cells can be evaluated. This information is utilized by each agent to iteratively increase or decrease the owned space in order to achieve an equal volume distribution (cf. Fig. 33.10c). This algorithm is known as Lloyd or k-means-algorithm (Du et al. 2006). The Lloyd algorithm itself is not communication aware. For avoiding any separation, the virtual cube restricts possible the target positions to an area where the RSSI between the nodes will be sufficient.

33.5.2 Wide-Area Mobility Controlled by Macroscopic Steerings The macroscopic behavior of the swarm is highly dependent on the task to be fulfilled by the swarm. In case of an exploration task (see reference scenario 1), the swarm needs to move as quickly as possible through a given volume, bearing in mind the key performance indicators, in particular exploration efficiency. In completely static and homogenous scenarios, a deterministic movement of a fixed configuration of UAVs working row by row through the scenario can be considered. Such a mechanism is neither communication-aware nor would it allow for a dynamic and self-organizing adaption to complex scenarios and UAV outages. From an exploration perspective, the well-known multiple traveling salesman problem (mTSP) provides a solution approach to find the shortest path to visit all relevant grid cells within a known volume. This centralized algorithm can serve as a best-case reference in terms of exploration efficiency (Daniel et al. 2011), but it operates completely independent from communicationrelated boundary conditions as well as from application data gathered during the exploration. Subsequently, it is necessary to develop a situation-aware behavior, which takes an eventually dynamically changing and unknown environment into account: • The Self-Repelling Walk (SRW) presents a known algorithm to steer the behavior of individual mobile agents, avoiding intersections with their own paths as much as possible. In its basic form, this approach can serve as a reference for an autonomous behavior of the individual UAVs within the swarm, but it takes neither communication requirements nor cooperative behavior into account. • The Cluster Repelling Walk (CRW) applies the principles of the SRW to a complete communication cluster being superimposed by microscopic steerings at the same time. On a macroscopic level the swarm moves synchronously while being still able to compensate for changing channel environments through microscopic control algorithms (cf. Fig. 33.11). An alternative to the combination of different microscopic and macroscopic steerings can be provided by a novel approach, namely, Communication-Aware Potential Fields (CAPF), which is described in more detail below.

770

C. Wietfeld and K. Daniel

2500 Macroscopic steering path of UAV cluster centroid

2000

UAV cluster

Z-Axis

1500

1000

500

0 2500 2000

2500 2000

1500 Y-Axis

1500

1000 1000

500

500 0

X-Axis

0

Fig. 33.11 Exploration by macroscopic steering (e.g., Cluster Repelling Walk)

33.5.3 An Integrated Steering Approach Using Communication-Aware Potential Fields The Communication-Aware Potential Fields (CAPF) algorithm expands the potential fields approach used to guide UAVs. Originally (Howard et al. 2002), the Potential Fields have been introduced to allow the mobile agents to derive attracting and repelling forces from a given environment. Thereby obstacles can be associated with a Potential Field creating repelling forces, while points of interest trigger attracting forces. The combination of potential forces with connectivity targets has been further evolved, for example, by Zou et al. (2003), Wang et al. (2005), Garetto et al. (2007) and Tan et al. (2008). In Goddemeier et al. (2011) UAVs have been associated with potential fields, thereby creating respective attracting as well as repelling forces (corresponding to low and high RSSI values when applying the Cluster Breathing method). In case of aerial sensor networks, the CommunicationAware Potential Fields (CAPF) are used for microscopic control combined with exploration-oriented macroscopic steerings. In case of the ad-hoc relay network

33 Cognitive Networking for UAV Swarms

771

scenario, the CAPF approach is suited to provide both micro- as well as macroscopic steering behavior: the area to be provisioned with a wireless communication service can, for example, be associated with a potential field generating an attracting force. In addition, base stations required to ensure connectivity with other networks induce attracting forces. In case of several UAVs providing the ground-based communication service, repelling forces between the UAVs are introduced to ensure a sufficient spacing to avoid interference and allow for maximum ground coverage.

33.5.4 Interference-Aware Positioning of Aerial Relays (IPAR) The introduction of terrestrial relays and femto-cells to increase the spectral efficiency of broadband cellular networks has gained much attention in recent years, for example, in Pabst et al. (2004) and Andrews et al. (2011). A recently proposed system concept is based on a swarm of UAVs acting as mobile relays to offload traffic to surrounding terrestrial cells with free resources (Rohde et al. 2012). Thereby cell outages as well as cell overload situations can be compensated. The optimal position of the aerial relays is highly dependent of the inter-cell interference (ICI) introduced by the mobile relays: in case that a mobile relay is operated in the vicinity too close to a ground-based station (cf. Fig. 33.1), the mobile relay will cause harmful interference resulting in reduced traffic capacity. Therefore, the specific Interference-Aware Positioning of Aerial Relays (IPAR) algorithm was introduced (Rohde et al. 2012). The IPAR algorithm’s goal is thus to find positions that yield a maximized throughput of the cellular network for the aerial relays while considering interference and available traffic load. The IPAR algorithm uses an interference and load aware analytical model, inspired by Elayoubi et al. (2008), to evaluate the system performance (fitness) and iteratively improves an initial population of relaying scenarios through mutation and recombination of the relay positioning. Example results for the positioning of UAVs leveraging the IPAR algorithm are presented below (cf. Sect. 33.6.3).

33.6

Performance Comparisons of Different Steering Algorithms

In the following, the performance of different algorithms presented above is presented for the two reference scenarios: aerial sensor network and aerial relay networks.

33.6.1 Aerial Sensor Network Scenario In this section different algorithms to address the aerial sensor network scenario are discussed. Figure 33.12 shows example results related to the Spatial Exploration Ratio (SER – percentage of grid cells visited by the swarm) and the Cluster

772

C. Wietfeld and K. Daniel Comparison of algorithms with 5 UAVs

100

0.1

1

CRW + Smart Cube* 50

Szenario Size: 1,5 km x 1,5 km x 1,5 km Number of UAVs: 5

100

90 120 180 Time [min]

240

300

20 0

100 80

1

60 40

10 3 100

60

CB SC CAPF Cluster Repelling Walk

Variation of number of UAVs

0.1

*Cluster Separation Ratio < 1%

0

80

40 SRW

CRW+Cluster Breathing*

10

100

60

mTSP

Cluster Separation Ratio (CSR) [%]

Spatial Exploration Ratio (SER) [%]

mTSP

0

Target area: SER>70%, CSR 10, which has the intended meaning that all AVs, except av1, should always be more than 10 m away from av1. This formula contains the variable x, the sort AV, the object av1, the feature XYDist, the predicates ¤ and >, and the constant value 10, besides the logical symbols. To evaluate such a formula (see Sect. 37.7.3 for details), an interpretation of its symbols must be given. Normally, their meanings are predefined. However, in the case of reasoning over streams, the meaning of features cannot be predefined since information about them becomes incrementally available. Instead their meaning has to be determined at runtime. To evaluate the truth value of a formula, it is therefore necessary to map feature symbols to streams, synchronize these streams, and extract a state sequence where each state assigns a value to each feature. In a system consisting of streams, a natural approach is to syntactically map each feature to a single stream. This is called syntactic integration. This works well when there is a stream for each feature and the person writing the formula is aware of the meaning of each stream in the system. However, when systems become more complex and when the set of streams or their meaning changes over time, it is much harder for a designer to explicitly state and maintain this mapping. Therefore, automatic support for mapping features in a formula to streams in a system based on their semantics is needed. This is called semantic integration. The purpose of this matching is for each feature to find one or more streams whose content matches the intended meaning of the feature. This is a form of semantic matching between features and contents of streams. The process of matching features to streams in a system requires that the meaning of the content of the streams is represented and that this representation can be used for matching the intended meaning of features with the actual content of streams. The same approach can be used for symbols referring to objects and sorts. It is important to note that the semantics of the logic requires the set of objects to be fixed. This means that the meaning of an object or a sort must be determined for a formula before it is evaluated and then may not change. It is still possible to have different instances of the same formula with different interpretations of the sorts and objects. The goal is to automate the process of matching the intended meaning of features, objects, and sorts to content of streams in a system. Therefore, the representation of the semantics of streams needs to be machine readable. This allows the system to reason about which stream content corresponds to which symbol in a logical formula. The knowledge about the meaning of the content of streams needs to be specified by a user, even though it could be possible to automatically determine this in the future. By assigning meaning to stream content, the streams do not have to use predetermined names, hard-coded in the system. This also makes the system domain independent, which implies that it could be used to solve different problems in a variety of domains without reprogramming.

908

P. Doherty et al.

The approach to semantic integration in DyKnow uses semantic web technologies to define and reason about ontologies. Ontologies provide suitable support for creating machine readable domain models (Horrocks 2008). Ontologies also provide reasoning support and support for semantic mapping which is necessary for the integration of streams from multiple UASs. The Web Ontology Language (OWL) (Smith et al. 2004) is used to represent ontologies. Features, objects, and sorts are represented in an ontology with two different class hierarchies: one for objects and one for features. To represent the semantic content of streams in terms of features, objects, and sorts, a semantic specification language called S SLT has been defined (Heintz and Dragisic 2012). This is used to annotate the semantic content of streams. Finally, a semantic matching algorithm has been developed which finds all streams which contain information relevant to a concept from the ontology, such as a feature. This makes it possible to automatically find all the streams that are relevant for evaluating a temporal logical formula. These streams can then be collected, fused, and synchronized into a single stream of states over which the truth value of the formula is incrementally evaluated. By introducing semantic mapping between ontologies from different UASs and reasoning over multiple related ontologies, it is even possible to find relevant streams distributed among multiple UASs (Heintz and Dragisic 2012).

37.6.6 ROS-Based Implementation The ROS-based implementation of DyKnow consists of three main parts: a stream processing part, a stream reasoning part, and a semantic integration part. Each part consists of a set of components. There are three types of components: engines, managers, and coordinators. An engine takes a specification and carries out the processing as specified. A manager keeps track of related items and provides an interface to these. A coordinator provides a high-level functionality by coordinating or orchestrating other functionalities. An overview of the parts and the components is shown in Fig. 37.25. This diagram corresponds to the DyKnow component in Fig. 37.2. The design is very modular as almost every component can be used independently. The stream processing part is responsible for generating streams by, for example, importing, merging, and transforming streams. The Stream Manager keeps track of all the streams in the system. Streams can either be generated by a stream processing engine or by some external program. A stream processing engine takes a stream specification and generates one or more streams according to the specification. The semantic integration part is responsible for finding streams based on their semantics relative to a common ontology. The Stream Semantics Manager keeps track of semantically annotated streams, where an annotation describes the semantic content of a stream. The Ontology Manager keeps track of the ontology which provides a common vocabulary. The Semantic Matching Engine finds all streams whose semantic annotation matches a particular ontological concept. The semantic

Streams

37 A Distributed Hybrid Deliberative/Reactive Architecture

909

Stream Processing Stream Manager Semantic Integration

Streams

Stream Processing Engine

Streams

Stream Semantics Manager

Ontology Manager

Semantic Matching Engine

Stream Reasoning Engine

Stream Reasoning Coordinator

Stream Reasoning

Fig. 37.25 The components of DyKnow

integration part is used by the stream reasoning part to find the relevant streams in order to evaluate a logical formula. The stream reasoning part is responsible for evaluating temporal logical formulas over streams as described in Sect. 37.7.3. A stream reasoning engine takes a logical formula and a stream of states and evaluates the formula over this stream. A Stream Reasoning Coordinator takes a logical formula, finds all the relevant streams needed to evaluate the formula, creates a stream specification for generating a single stream of states from all the relevant streams, and instructs the stream reasoning engine to evaluate the formula over the stream as it is generated by a stream processing engine.

37.6.7 A Traffic Monitoring Example As a concrete example of the use of DyKnow, Fig. 37.26 provides an overview of how part of the incremental processing required for a traffic surveillance task can be organized as a set of DyKnow knowledge processes. At the lowest level, a helicopter state estimation component uses data from an inertial measurement unit (IMU) and a global positioning system (GPS) to determine the current position and attitude of the UASTL RMAX. A camera state estimation component uses this information, together with the current state of the pan-tilt unit on which the cameras are mounted, to generate information about the current camera state. The image processing component uses the camera state to determine where the camera is currently pointing. Video streams from the color and thermal cameras can then be analyzed in order to generate vision percepts representing hypotheses about moving and stationary physical entities, including their approximate positions and velocities. The data originating from the PFC and the PPC is provided by the Data Interface through the Platform Server (Sect. 37.4.2.1).

910

P. Doherty et al.

Symbolic reasoning

Chronicle Recognition

[0, 20]

B

[10, 20]

Qualitative spatial relations (close, behind, same_road, …)

D

Qualitative Spatial Reasoning

Geographical Information System

A

Car objects

[5, 10]

C [0, 10]

[10, 10]

E

Temporal Logic Progression

Anchoring

Sensor processing

Vision objects Color camera Thermal camera

Image Processing Camera state

Vehicle State Estimation IMU

GPS

Vehicle state

Camera State Estimation Pan-tilt unit

Source Computational unit Stream

Fig. 37.26 Potential organization of the incremental processing required for a traffic surveillance task

Symbolic formalisms such as chronicle recognition (Ghallab 1996) require a consistent assignment of symbols, or identities, to the physical objects being reasoned about and the sensor data received about those objects. Image analysis may provide a partial solution, with vision percepts having symbolic identities that persist over short intervals of time. However, changing visual conditions or objects temporarily being out of view lead to problems that image analysis cannot (and should not) handle. This is the task of the anchoring system, which uses progression of formulas in a metric temporal logic to incrementally evaluate potential hypotheses about the observed objects (Heintz et al. 2013). The anchoring system also assists in object classification and in the extraction of higher-level attributes of an object. For example, a geographical information system can be used to determine whether an object is currently on a road or in a crossing. Such attributes can in turn be used to derive relations between objects, including qualitative spatial relations such as beside.car1 ; car2 / and close.car1 ; car2 /. Concrete events corresponding to changes in such attributes and predicates finally provide sufficient information for the chronicle recognition system to determine when higher-level events such as reckless overtakes occur.

37.7

The Deliberative Layer

Conceptually, the deliberative layer includes all high-autonomy functionalities normally associated with rich world models and deep reasoning capabilities. Functionalities commonly associated with the deliberative layer are automated task

37 A Distributed Hybrid Deliberative/Reactive Architecture

911

planners, motion planners, execution monitoring systems, diagnosis systems, and knowledge bases with inference engines, among others. The temporal latencies associated with the decision cycles for these functionalities are often far slower than for functionalities which exist in the reactive and control layers. The decision cycles for the different layers are necessarily asynchronous where each of the sets of functionalities run concurrently. The interfaces between the deliberative layer and the reactive and control layers are essential for transitioning output from deliberative functionality into useful processes at the lower layers of the architecture. In some sense, much of what happens is a form of dynamic compilation which results in transforming qualitative goal-directed assertions into actual actuation commands associated with the control layer of the architecture which contribute to the achievement of these goals. A deliberative functionality, motion planning, and the dynamic compilation of its output have already been described in Sect. 37.5. This section will describe two additional deliberative functionalities central to high autonomy in the HDRC3 architecture. Section 37.7.2 describes an automated task-based planner, TALplanner, and its integration in the HDRC3 architecture. Section 37.7.3 describes a novel execution monitoring system based on specifying monitoring queries in terms of temporal logical formulas and evaluating these formulas online in real time. Due to the complex nature of these functionalities, an open research issue is how one might provide verification and validation techniques for the use and integration of these functionalities, in addition to the formal specification of high-level missions. The section therefore begins with a brief presentation of a Temporal Action Logic (TAL) which provides a formal basis for specifying the semantics of TSTs, TALplanner, and the execution monitor (Sect. 37.7.1). Additionally, in Sect. 37.7.4, it is shown that high-level missions can in fact be specified in terms of TSTs. Consequently, high-level missions are not only provided with a formal semantics, but due to the relation between these specifications and TSTs, there is a natural means of coupling declarative specification of missions with their procedural execution.

37.7.1 Temporal Action Logic Temporal Action Logic, TAL, is a well-established non-monotonic logic for representing and reasoning about actions (Doherty et al. 1998; Doherty and Kvarnstr¨om 2008). This logic provides both clear intuitions and a formal semantics for a highly expressive class of action specifications which supports temporal aspects, concurrent execution, and incomplete knowledge about the environment and the effects of an action. Therefore it is a highly suitable basis for describing and reasonable the elementary actions used in realistic mission specifications. Here a limited subset of TAL is described, and the reader is referred to Doherty and Kvarnstr¨om (2008) for further details. TAL provides an extensible macro language, L(ND) , that supports the knowledge engineer and allows reasoning problems to be specified at a higher abstraction level than plain logical formulas. The basic ontology includes parameterized features

912

P. Doherty et al.

f .x/ that have values v at specific timepoints t, denoted by Œtf .x/ D O v, or over intervals, Œt; t 0 f .x/ D O v. Incomplete information can be specified using disjunctions of such facts. Parameterized actions can occur at specific intervals of time, denoted by Œt1 ; t2 A.x/. To reassign a feature to a new value, an action uses the expression R.Œtf .x/ D O v/. Again, disjunction can be used inside R./ to specify incomplete knowledge about the resulting value of a feature. The value of a feature at a timepoint is denoted by value.t; f /. The logic is based on scenario specifications represented as narratives in L(ND). Each narrative consists of a set of statements of specific types, including actiontype specifications defining named actions with preconditions and effects. The basic structure, which can be elaborated considerably (Doherty and Kvarnstr¨om 2008), is as follows: Œt1 ; t2 A.v/

.pre .t1 ; v/ H) post .t1 ; t2 ; v// ^ cons .t1 ; t2 ; v/

stating that if the action A.v/ is executed during the interval Œt1 ; t2 , then given that its preconditions pre .t1 ; v/ are satisfied, its effects, post .t1 ; t2 ; v/, will take place. Additionally, cons .t1 ; t2 ; v/ can be used to specify logical constraints associated with the action. For example, the following defines the elementary action fly-to: If a AV should fly to a new position .x 0 ; y 0 / within the temporal interval Œt; t 0 , it must initially have sufficient fuel. At the next timepoint t C1, the AV will not be hovering, and in the interval between the start and the end of the action, the AV will arrive and its fuel level will decrease. Finally, there are two logical constraints bounding the possible duration of the flight action. Œt; t 0 fly-to.av; x0 ; y0 / Œt fuel.av/  fuel-usage.av; x.av/; y.av/; x0 ; y0 / ! R.Œt C 1 hovering.av/ D O False/ ^ R..t; t 0  x.av/ D O x0 / ^ R..t; t 0  y.av/ D O y0 / ^ R..t; t 0  fuel.av/ D O value.t; fuel.av/  fuel-usage.av; x.av/; y.av/; x0 ; y0 /// ^ t 0  t  value.t; min-flight-time.av; x.av/; y.av/; x0 ; y0 // ^ t 0  t  value.t; max-flight-time.av; x.av/; y.av/; x0 ; y0 //  The translation function Trans translates L(ND) expressions into L(FL), a first-order logical language (Doherty and Kvarnstr¨om 2008). This provides a welldefined formal semantics for narratives in L(ND). This separation between the macro language and the base logic makes TAL highly extensible. When adding new constructs to the formalism, new expression types are defined in L(ND), and  Trans is extended accordingly, generally with no extension required in the base language L(FL). The L(FL) language is order-sorted, supporting both types and subtypes for features and values. This is also reflected in L(ND), where one often assumes variable types are correlated to variable names – for example, av3 implicitly ranges over AVs. There are a number of sorts for values Vi , including the Boolean sort B with the constants ftrue; falseg. V is a supersort of all value sorts. There are a number of sorts for features Fi , each one associated with a value sort

37 A Distributed Hybrid Deliberative/Reactive Architecture

913

dom.Fi / D Vj for some j . The sort F is a supersort of all fluent sorts. There is also an action sort A and a temporal sort T . Generally, t; t 0 will denote temporal variables, while ;  0 ; 1 ; : : : are temporal terms. L(FL) currently uses the following predicates, from which formulas can be defined inductively using standard rules, connectives, and quantifiers of first-order logic: • Holds W T  F  V, where Holds.t; f; v/ expresses that a feature f has a value v at a timepoint t, corresponding to Œt f D O v in L(ND). • Occlude W T  F , where Occlude.t; f / expresses that a feature f is permitted to change values at time t. This is implicit in reassignment, R.Œt f D O v/, in L(ND). • Occurs W T  T  A, where Occurs.ts ; te ; A/ expresses that a certain action A occurs during the interval Œts ; te . This  corresponds to Œts ; te A in L(ND). When a narrative is translated, Trans first generates the appropriate L(FL) formulas corresponding to each L(ND) statement. Foundational axioms such as unique names and domain closure axioms are appended when required. Logical entailment then allows the reasoner to determine when actions must occur, but the fact that they cannot occur at other times than explicitly stated is not logically entailed by the translation. This problem is handled in a general manner through filtered circumscription, which also ensures that fluents can change values only when explicitly affected by an action or dependency constraint (Doherty and Kvarnstr¨om 2008). The structure of L(ND) statements ensures that the second-order circumscription axioms are reducible to equivalent first-order formulas, a reduction that can often be performed through predicate completion. Therefore, classical first-order theorem proving techniques can be used for reasoning about TAL narratives (Doherty and Kvarnstr¨om 2008). For unmanned systems, however, the logic will primarily be used to ensure a correct semantics for planners, execution monitors, and mission specification languages and correlate this semantics closely to the implementation. Using TAL does not require theorem proving onboard.

37.7.2 Task Planning Using TALplanner When developing the architecture for a system capable of autonomous action execution and goal achievement, one can envision a spectrum of possibilities ranging from each behavior and task being explicitly coded into the system, regardless of complexity, up to the other extreme where the system itself generates complex solutions composed from a set of very primitive low-level actions. With the former end of the spectrum generally leading to more computationally efficient solutions and the latter end generally being far more flexible in the event of new and potentially unexpected tasks being tackled, the proper choice is usually somewhere between the two extremes. In fact, several different points along the spectrum might be appropriate for use in different parts of a complex system. This is also the case in the HDRC3 architecture: Elementary action nodes in TSTs provide a set of high-level actions such as “takeoff” and “fly-to point A,” but one also makes use

914 Fig. 37.27 Task planning called by TST executors for goal nodes

P. Doherty et al. TST Executor (for goal nodes)

DyKnow

TFPOP or TALplanner (Task Planning)

Motion Planner

Motion Plan Info (Trajectory Queries)

of automated planning techniques to compose such actions into plans satisfying a set of declaratively specified goals. The goals themselves will vary depending on the nature of a mission. For example, they could involve having acquired images of certain buildings or having delivered crates of emergency supplies to certain locations after a natural disaster. Support for planning is closely integrated in the HDRC3 architecture through the use of goal nodes in Task Specification Trees (Sect. 37.4.3). Any given TST can contain multiple goal nodes, each of which describes a simple or complex goal to be achieved at that particular point in the corresponding mission. When a goal node is assigned to a specific agent, the executor for that goal node can call DyKnow (Sect. 37.6) or use any other means at its disposal to acquire essential information about the current state of the world and the AV’s own internal state (Fig. 37.27). The node is then expanded by that agent through a call to an onboard task planner, which in turn can call the motion planner and Motion Plan Info functionalities discussed earlier (Sect. 37.5.1). The resulting concurrent plan is converted to a new “plan subtree” attached under the goal node. Since TST structures are inherently distributable, the entire plan can then be distributed to the participating agents. This is the principal means of invoking a task planning functionality. Note that the resulting TST directly represents a plan using sequence and concurrency nodes together with elementary action nodes corresponding to plan actions. Therefore, a plan can be executed at the appropriate time in the standard manner using TST executors. This obviates the need for a separate plan execution system and results in an integrated means of executing a mission, regardless of the techniques or combination of techniques that was used to generate it. The planner that is used to expand a goal node can be changed freely, and in a heterogeneous system different platforms may use different planners. Currently the planners most commonly used in the HDRC3 architecture are TALplanner (Doherty and Kvarnstr¨om 2001; Kvarnstr¨om 2005) and TFPOP (Kvarnstr¨om 2011), two domain-independent concurrent temporal planners that were developed explicitly for use in this architecture. These planners are similar in certain respects: Both have a semantics based on Temporal Action Logic (Sect. 37.7.1), and both use domainspecific control formulas in this logic to guide the search for a solution, leading to very rapid plan generation. However, they use different search spaces and plan

37 A Distributed Hybrid Deliberative/Reactive Architecture

915

structures, and they differ in the expressivity of the temporal constraints they permit. TALplanner will be described here, while the reader is referred to Kvarnstr¨om (2011) for further information about TFPOP. TALplanner. TALplanner (Doherty and Kvarnstr¨om 1999; Kvarnstr¨om and Doherty 2000; Kvarnstr¨om 2005) is a forward-chaining temporal concurrent planner where planning domains and problem instances are specified as goal narratives in a version of TAL extended with new macros and statement types for plan operators, resource constraints, goal specifications, and other types of information specific to the task of plan generation. These macros serve to simplify planning domain specifications, but retain a basis in standard TAL through an extended translation function. For example, plan operators are converted into standard TAL action-type specifications, and actions in a plan are represented as standard timed action occurrences of the form Œt1 ; t2  A.v/. Control Formulas. In addition to providing a declarative first-order semantics for planning domains, TAL is also used to specify a set of domain-specific temporal control formulas acting as constraints on the set of valid solutions: A plan is a solution only if its final state satisfies the goal and all control formulas are satisfied in the complete state sequence that would result from executing the plan, which can be viewed as a logical model. The use of control formulas serves two separate purposes. First, it allows the specification of complex temporally extended goals such as safety conditions that must be upheld throughout the (predicted) execution of a plan. Second, the additional constraints on the final solution often allow the planner to prune entire branches of the search tree – whenever it can be proven that every search node on the branch corresponds to a state sequence that violates at least one control rule. Formulas can then be written to guide the planner towards those parts of the search space that are more likely to contain plans of high quality, in terms of time usage or other quality measures. As an example, consider three simple control rules that could be used in an airplane-based logistics domain. First, a package should only be loaded onto a plane if a plane is required to move it: if the goal requires it to be at a location in another city. Regardless of which operator is used to load a package, one can detect this through the fact that it is in a plane at time t C 1, but was not in the same plane at time t. 8t; obj; plane; loc: [t] :in(obj, plane) ^ at(obj, loc) ^ [t+1] in(obj, plane) ! 9loc0 [ goal(at(obj, loc0 )) ^ [t] city of(loc) D 6 O city of(loc0 ) ] Second, if a package has been unloaded from a plane, there must be a valid reason for this: It must be the case that the package should be in the city where the plane has landed.

916

P. Doherty et al.

8t; obj; plane; loc: [t] in(obj, plane) ^ at(plane, loc) ^ [t+1] :in(obj, plane) ! 9loc0 [ goal(at(obj, loc0 )) ^ [t] city of(loc) D O city of(loc0 ) ] Third, if a package is at its destination, it should not be moved. 8t; obj; loc: [t] at(obj, loc) ^ goal(at(obj, loc)) ! [t+1] at(obj, loc) Surprisingly, such simple hints to an automated planner can often improve planning performance by orders of magnitude given that the planner has the capability to make use of the hints. Concurrent Plan Structure. Forward-chaining planning is most often used for sequential plans, where each new action is added immediately after the previous one in the plan. TALplanner uses a similar technique for concurrent plan generation, with a relaxed constraint on where a new action occurrence is placed. Definition 37.1. (Concurrent Plan). A concurrent plan for a goal narrative G is a tuple of ground fluent-free action occurrences with the following constraints. First, the empty tuple is a concurrent plan for G. Second, given a concurrent plan p D hŒ1 ; 10  o1 .c 1 /; : : : ; Œn ; n0  on .c n /i for G, its successors are the sequences that 0 add one new action occurrence ŒnC1 ; nC1  onC1 .c nC1 / and satisfy the following constraints: 1. Let G 0 D G [ fŒ1 ; 10  o1 .c 1 /; : : : ; Œn ; n0  on .c n /g be the original goal narrative G combined with the existing plan. Then, the new action onC1 .c nC1 / must be 0 applicable over the interval ŒnC1 ; nC1  in G 0 . This implies that its preconditions are satisfied, that its effects are not internally inconsistent and do not contradict the effects of the operator instances already present in the sequence, and that 0 the duration nC1  nC1 is consistent with the duration given in the operator specification. 2. 1 D 0: The first action starts at time 0. 3. nC1  n : The new action cannot be invoked before any of the actions already added to the plan. 4. nC1  max.10 ; : : : ; n0 /: There can be no gap between the time interval covered by the current plan and the start of the newly added action. Generating Concurrent Plans. The concurrent TALplanner algorithm will now be briefly described (Algorithm 3). The reader is referred to Kvarnstr¨om (2005) for further details. First, the planner conjoins all goal statements (line 2) and uses the control formulas in the goal narrative to generate a set of pruning constraints. These constraints allow control formulas to be verified efficiently when actions are incrementally added to a plan. Specifically, each control formula may result in an initial constraint to be tested initially, a set of operator-specific incremental

37 A Distributed Hybrid Deliberative/Reactive Architecture

917

Algorithm 3 Concurrent TALplanner Input: A goal narrative G . Output: A plan narrative entailing the goal Ngoal and the control formulas Ncontrol . 1 procedure V TALplanner-concurrent.G / G goal Conjunction of all goal statements 2  generate-pruning-constraints.G control / 3 hinit; incr; finali hinit; ;; 0; 0; hii hcond. queue, visited states, latest invocation time, tmax , plani 4 node hnodei Stack (depth first search) 5 Open 6 while Open ¤ hi do pop.Open/ Current plan candidate 7 hC; S; 0 ; max ; pi 8 G0 G [ p [ occlude-all-after.G ; 0 / No knowledge about future 9 for all constraints ˛ in C do Check queued constraints C C n f˛g Remove satisfied constraint 10 if TransC .G 0 / ˆ Trans ˛  then  Constraint violated 11 elsif TransC .G 0 / ˆ Trans :˛ then backtrack G 00 G [ p [ ftmax D max g Narrative with complete knowledge 12 Consistency check 13 if TransC .G 00 / ˆ false then backtrack current plan/ then backtrack 14 if 9s 2 S:better-or-equal.s; final state of  Goal + queued + final ctrl satisfied 15 if TransC .G 00 / ˆ Trans  ^ C ^ final then 16 return G 00 17 else Not a solution, but check children S [ ffinal state of current plang 18 S0 19 for all successor actions A D Œ ; 0  oi .c/ for p according to Def 37.1 do C [ incri Œ ; c Old conditions + incr control 20 C0 C 0 [ fprevail condition of Ag Add prevail condition 21 C0 22 push hC 0 ; S 0 ; ; max.max ; 0 /; hpI Aii onto Open 23 fail

constraints to be tested whenever an instance of a particular operator is added to a plan, and a final constraint to be tested in the final solution. The initial search node is created (line 4), as is a stack of open search nodes (for depth first search). As long as the search space has not been exhausted (line 6), the planner retrieves the topmost node in the stack of open nodes. This node consists of a queue of constraints that remain to be evaluated (C ), a set of visited states to be used in cycle checking (S ), the latest invocation time of any action in the current plan (0 ), the latest end time of any action in the current plan (tmax ), and the current plan candidate (p). Assuming a completely specified initial state and deterministic actions, the narrative G [ p would now contain complete information about the development of the world that would result from executing exactly the plan p and nothing else. This then corresponds to a unique infinite state sequence specifying the values of all TAL fluents at all points in time. However, the current search node must only be pruned if all possible extensions of p violate a control formula. Given the constraints placed on successors in Definition 37.1, no action added to p can be invoked earlier than 0 , and therefore, all effects of actions added to p must take place at 0 C 1 or later. Line 8 therefore generates the narrative G [ p [ occlude-all-after.G; 0 / which has additional formulas disclaiming knowledge about fluents after time t0 .

918

P. Doherty et al.

In lines 9–11, the planner determines whether a queued constraint is now definitely true (in which case it can be removed from the queue) or definitely false (in which case the planner has to backtrack). Then, the planner must verify certain conditions under the assumption that no additional actions are added. A narrative G 00 assuming complete information is therefore generated in line 12. If this is inconsistent (which in practice can be tested efficiently given the expressivity of TALplanner operators), the planner must backtrack (line 13). If a state that is better or equal has already been visited, the planner should also backtrack (line 14). The better-or-equal relation can, for example, take into account resource availability, where a state that is equal in all respects except that it provides more of a particular resource can be considered strictly better. In lines 15–16, TALplanner determines whether the current plan candidate satisfies the goal and all remaining pruning constraints. If this is the case, a solution is found and can be returned. If not, the node is expanded through the generation of a set of successors. Each successor may be given additional incremental pruning constraints, instantiated with the actual invocation timepoints and actual arguments of the corresponding action (line 20). Similarly, any prevail conditions (similar to preconditions but relating to the entire execution interval of an action) are added to the condition queue, to be tested when more information about future states is available.

37.7.3 Execution Monitoring Regardless of the effort spent modeling all possible contingencies, actions and plans may fail. Robust performance in a noisy environment therefore requires some form of supervision, where the execution of a plan is constantly monitored in order to detect and recover from potential or actual failures. For example, since an AV might accidentally drop its cargo, it should monitor the condition that whenever it carries a crate, the crate remains until the AV reaches its intended destination. This is an example of a safety constraint, a condition that must be maintained during the execution of an action or across the execution of multiple actions. The carrier can also be too heavy, which means that one must be able to detect takeoff failures where the AV fails to gain sufficient altitude. This can be called a progress constraint: Instead of maintaining a condition, a condition must be achieved within a certain period of time. While some of these constraints would best be monitored at the control level, there are also many cases where monitoring and recovery should be lifted into a higher-level execution monitor (De Giacomo et al. 1998; Fern´andez and Simmons 1998; Fichtner et al. 2003; Gat et al. 1990; Ben Lamine and Kabanza 2002). The execution monitoring system described here is based on an intuition similar to the one underlying the temporal control formulas used in TALplanner. As a plan is being executed, information about the surrounding environment is sampled at a given frequency by DyKnow (Sect. 37.6). Each new sampling point generates a

37 A Distributed Hybrid Deliberative/Reactive Architecture

919

new state which provides information about all state variables used by the current monitor formulas, thereby providing information about the actual state of the world as opposed to what could be predicted from the domain specification. The resulting sequence of states corresponds to a partial logical interpretation, where “past” and “present” states are completely specified whereas “future” states are completely undefined. Note that simply comparing the actual and predicted states and signaling a violation as soon as a discrepancy is found is not sufficient, because not all discrepancies are fatal – for example, if the altitude was predicted to be 5:0 m and the current measurement turns out to be 4:984 m. Also, some information might be expensive or difficult to sense, in which case the ground operator should be given more control over which information is actually gathered and used for monitoring. Sensing may even require special actions that interfere with normal mission operations. Finally, the richer the planner’s domain model is, the more it can predict about the development of the world. This should not necessarily lead to all those conditions being monitored, if they are not relevant to the correct execution of a plan. Therefore, most conditions to be monitored are explicitly specified, though many conditions can be automatically generated within the same framework if so desired. Execution Monitor Formulas in TAL . Execution monitor formulas are expressed in a variation of TAL where the high-level language L(ND) is augmented with a set of tense operators similar to those used in modal tense logics such as LTL (Emerson 1990) and MTL (Koymans 1990). Tense operators allow the expression of complex metric temporal conditions and are amenable to incremental evaluation as each new state is generated. This allows violations to be detected as early and as efficiently as possible using a formula progression algorithm, while the basis in TAL provides a common formal semantic ground for planning and monitoring. Three tense operators have been introduced into L(ND): U(until), Þ(eventually), and (always). Like all expressions in L(ND), these operators are macros on top of the first-order base language L(FL). Definition 37.2. (Monitor Formula). A monitor formula is one of the following: •    0 ,  <  0 , or  D  0 , where  and  0 are temporal terms. • !  ! 0 , ! < ! 0 , or ! D ! 0 , where ! and ! 0 are value terms. • f, where f is a Boolean fluent term (state variable term). • fD O !, where f is a fluent term and ! is a value term of the corresponding sort. •  UŒ; 0  , where  and are monitor formulas and  and  0 are temporal terms. • ÞŒ; 0  , where  is a monitor formula and  and  0 are temporal terms. • Œ; 0  , where  is a monitor formula and  and  0 are temporal terms. • A combination of monitor formulas using standard logical connectives and quantification over values. The shorthand notation  U   UŒ0;1/ , Þ   ÞŒ0;1/ , and    Œ0;1/  is also permitted.

920

P. Doherty et al.

Tense operators use relative time, where each formula is evaluated relative to a “current” timepoint. The semantics of these formulas satisfies the following conditions (see Doherty et al. (2009) for details): • The formula  UŒ; 0  (“until”) holds at time t iff holds at some state with time t 0 2 Œt C ; t C  0  and  holds until then (at all states in Œt; t 0 /, which may be an empty interval). • The formula ÞŒ; 0   (“eventually”) is equivalent to true UŒ; 0   and holds at t iff  holds in some state with time t 0 2 Œt C ; t C  0 . • The formula Œ; 0   is equivalent to : ÞŒ; 0  : and holds at t iff  holds in all states with time t 0 2 Œt C ; t C  0 . Example 37.3. Suppose that an AV supports a maximum continuous power usage of M , but can exceed this by a factor of f for up to  units of time, if this is followed by normal power usage for a period of length at least  0 . The following formula can be used to detect violations of this specification:  8av:.power.av/ > M ! power.av/ < f  M UŒ0;  Œ0; 0  power.av/  M /

t u

Note that this does not in itself cause the AV to behave in the desired manner. That has to be achieved in the lower-level implementations of the helicopter control software. The monitor formula instead serves as a method for detecting the failure of the helicopter control software to function according to specifications. Monitors and Actions. In many cases, conditions that should be monitored are closely related to the actions currently being executed. Two extensions are made to facilitate the specification of such formulas. First, monitor formulas can be explicitly associated with specific operator types. Unlike global monitor formulas, such formulas are not activated before plan execution but before the execution of a particular step in the plan, which provides the ability to contextualize a monitor condition relative to a particular action. An operator-specific monitor formula can also directly refer to the arguments of the associated operator. Example 37.4. Execution should be monitored when an AV attempts to pick up a box. Since the arguments of pickup-box include the av and the box, the following operator-specific monitor formula can be used: ÞŒ0;5000 Œ0;1000 carrying.av; box/ Within 5,000 ms, the AV should detect that it is carrying the box, and it should detect this for at least 1,000 ms. The latter condition protects against problems during the pickup phase, where the box may be detected during a very short period of time even though the ultimate result is failure. t u Second, a set of execution flags allow monitor formulas to query the internal execution state of an agent. This is useful when one wants to state that a certain fact

37 A Distributed Hybrid Deliberative/Reactive Architecture

921

should hold during the execution of an action or that an effect should be achieved during the execution of an action or before the execution of another action. For example, if an AV picks up a crate, it should sense the weight of the crate until it deliberately puts down the crate. An execution flag is an ordinary Boolean state variable which holds exactly when the corresponding action is being executed. By convention, this state variable will generally be named by prepending “executing-” to the name of the corresponding operator: The pickup-box operator is associated with the executing-pickup-box execution flag, which takes a subset of the operator’s parameters. TST executors for elementary action nodes can set this flag when execution starts and clear it when execution ends. Example 37.5. Consider the climb(av) operator, which should cause the AV to ascend to its designated flight altitude. Here, one may wish to monitor the fact that the AV truly ends up at its flight altitude. This can be achieved using the formula executing-climb.av/ U altitude.av/  7:0. t u When it is clear from context which operator is intended, the shorthand notation EXECcan be used to refer to its associated execution flag with default parameters: EXEC U altitude.av/  7:0. Example 37.6. Whenever an AV picks up a box, it should detect the box within 5,000 ms and keep detecting it until it is explicitly put down. Using an operatorspecific monitor formula for pickup-box: EXEC UŒ0;5000 .carrying.av; box/ U executing-putdown.av; box//

Automatic Generation of Monitor Formulas. The use of a single logical formalism for modeling both planning and execution monitoring provides ample opportunities for the automatic generation of conditions to be monitored. For example, one can automatically generate formulas verifying that preconditions and prevail conditions (which must hold throughout the execution of an action) are satisfied, that expected effects take place, and that bounds on action durations are not violated. Similarly one can automatically extract causal links, where one action achieves one condition that is later needed by another, and generate a formula verifying that this condition is never violated in the interval between these actions. Since there are pragmatic reasons for not generating all possible monitor formulas, automatic generation only takes place for those conditions that are flagged for monitoring. This provides the benefits of automatic formula generation while keeping the control in the hands of the domain designer. See Doherty et al. (2009) for details.

37.7.3.1 Recovery from Failures Any monitor formula violation signals a potential or actual failure from which the system must attempt to recover in order to achieve its designated goals.

922

P. Doherty et al.

Recovery is a complex topic, especially when combined with the stringent safety regulations associated with autonomous flight. Goals and constraints may be time dependent, making local repairs difficult: An AV might only be allowed to fly in certain areas at certain times. Also, recovering from a failure for one AV may require changing the plans of another AV: If heli1 fails to deliver a box of medicine on time, heli2 might have to be rerouted. Therefore, the main focus has been on recovery through replanning. Given that the planner is sufficiently fast when generating new plans, this does not adversely affect the execution of a fully autonomous mission. Having detected a failure, the first action of an AV is to suspend the execution of the current plan, execute an emergency break if required, and then go into autonomous hover mode. Currently, one takes advantage of the fact that the UAS Tech Lab RMAX and LinkQuads are rotor based and can hover. For fixed-wing platforms, this is not an option, and one would have to go into a loiter mode if the recovery involves time-consuming computation. This is followed by the execution of a recovery operator, if one is associated with the violated monitor formula. The recovery operator can serve two purposes: It can specify emergency recovery procedures that must be initiated immediately without waiting for replanning, and it can permit the execution system to adjust its assumptions about what can and cannot be done. For example, if an AV fails to take off with a certain carrier, the associated recovery operator can adjust the AV’s assumptions about how many boxes it is able to lift. This feeds back information from the failure into the information given to the planner for replanning. The implementation of a recovery operator can also detect the fact that the AV has attempted and failed to recover from the same fault too many times and choose whether to give up, try another method, remove some goals in order to succeed with the remaining goals, or contact a human for further guidance.

37.7.3.2 Execution Monitoring with Inaccurate Sensors Monitoring should not only maximize the probability that a failure is detected but also minimize the probability of false positives, where a failure is signaled but none has occurred. Some problems, such as those caused by dropouts and communication delays, can be ameliorated by extrapolating historical values and by delaying state generation slightly to allow values to propagate through the distributed system. Noise could be minimized through sensor value smoothing techniques and sensor fusion techniques. However, inaccuracies in the detected state sequence can never be completely eliminated. Therefore, the fact that state values may be inaccurate should be taken into consideration when writing monitor formulas. For example, the meaning of the condition  8av:speed.av/  T is that the sensed and approximated speed of an AV must never exceed the threshold T . Since a single observation of speed.av/ above the threshold might be an error or a temporary artifact, a more robust solution would be to signal a failure if the sensed speed has been above the threshold during an interval Œ0;  instead of at a single timepoint. This can be expressed as  ÞŒ0;  speed.av/  T : It should always be the case that within the interval Œ0;  from now, the sensed speed returns to being below the threshold.

37 A Distributed Hybrid Deliberative/Reactive Architecture

923

This formula is somewhat weak: It only requires that a single measurement in every interval of length  is below the threshold. An alternative would be to require that within  time units, there will be an interval of length  0 during which the AV stays within the limits: .speed.av/ > T ! ÞŒ0;  Œ0; 0  speed.av/  T /.

37.7.3.3 Formula Progression To promptly detect violations of monitor conditions during execution, a formula progression algorithm is used (Bacchus and Kabanza 1998). By definition, a formula  holds in the state sequence Œs0 ; s1 ; : : : ; sn  iff Progress.; s0 / holds in Œs1 ; : : : ; sn . In essence, this evaluates those parts of the monitor formula that refer to s0 , returning a new formula to be progressed in the same manner once s1 arrives. If the formula ? (false) is returned, then sufficient information has been received to determine that the monitor formula must be violated regardless of the future development of the world. For example, this will happen as soon as the formula  speed < 50 is progressed through a state where speed  50. Similarly, if > (true) is returned, the formula must hold regardless of what happens “in the future.” This will occur if the formula is of the form Þ  (eventually,  will hold), and one has reached a state where  indeed does hold. In other cases, the state sequence complies with the constraint “so far,” and progression will return a new and potentially modified formula that should be progressed again as soon as another state is available. Definition 37.3. (Progression of Monitor Formulas). The following algorithm is used for progression of monitor formulas. Note that states are not first-class objects in TAL and are therefore identified by a timepoint  and an interpretation I. Special cases for  and Þ can be introduced for performance. procedure Progress.; ; I/ if  D f .x/ D O v  if I ˆ Trans Œ  return > else return ? if  D :1 return :Progress.1 ; ; I/ if  D 1 ˝ 2 return Progress.1 ; ; I/ ˝ Progress.2 ; ; I/ if  D 8x: V // where x belongs to the finite domain X return c2X Progress.Œx 7! c; ; I/ if  D 9x: W // where x belongs to the finite domain X return c2X Progress.Œx 7! c; ; I/ if  contains no operator  tense  if I ˆ Trans  return > else return ? if  D 1 UŒ1 ;2  2 if 2 < 0 return ? elsif 0 2 Œ1 ; 2  return Progress.2 ; ; I/ _ .Progress.1 ; ; I/^ .1 UŒ1 1;2 1 2 // 15 else return Progress.1 ; ; I/ ^ .1 UŒ1 1;2 1 2 /

1 2 3 4 5 6 7 8 9 10 11 12 13 14

The result of Progressis simplified using the rules :? D >, .? ^ ˛/ D .˛ ^ ?/ D ?, .? _ ˛/ D .˛ _ ?/ D ˛, :> D ?, .> ^ ˛/ D .˛ ^ >/ D ˛,

924

P. Doherty et al.

and .> _ ˛/ D .˛ _ >/ D >. Further simplification is possible using identities such as ÞŒ0;   ^ ÞŒ0; 0    ÞŒ0;min.; 00 / . For this approach to be useful, it must be possible to progress typical monitor formulas through the state sequences generated during a mission using the often limited computational power available in an autonomous robotic system. The following evaluation uses one synthetic test where one can study complex combinations of time and modality. An earlier version of the actual DRC computer onboard a UASTL RMAX was used to run progression tests for formulas having a form that is typical for monitor formulas in many applications. State sequences are constructed to exercise both the best and the worst cases for these formulas. The evaluation shows that even with the worst possible inputs, complex formulas can be evaluated in less than 1 ms per state and formula on this CPU (1.4 GHz Pentium M). One example formula is  :p ! ÞŒ0;1000 Œ0;999 p, corresponding to the fact that if the property p is false, then within 1,000 ms, there must begin a period lasting at least 1,000 ms where p is true. For example, the previously discussed formula .speed.av/ > T ! ÞŒ0;  Œ0; 0  speed.av/  T / has this general form. To estimate the cost of evaluating this formula, it was progressed through several different state streams corresponding to the best case, the worst case, and two intermediate cases. A new state in the stream was generated every 100 ms, which means that all formulas must be progressed within this time limit or monitoring will fall behind. Figure 37.28 shows the average time required to progress a certain number of formulas through each state in a sequence. This indicates that 100 ms is sufficient for progressing between 1,500 and 3,000 formulas of this form on the computer onboard the UASTL RMAX, depending on the state stream. Similar results have been shown for formulas of different forms, such as  :p ! ÞŒ0;1000 Œ0;999 p and  :p ! ÞŒ0;1000 Œ0;999 p (Doherty et al. 2009).

37.7.4 High-Level Mission Specification The ability to clearly and concisely specify missions is fundamentally important for an unmanned system to be practically useful and requires a suitable mission specification language that should satisfy a variety of requirements and desires. The language should be comprehensible to humans and not only useful as an intermediate representation both generated and received by software. At the same time intuitions are not sufficient: A strict formal semantics must be available. There should also be a close connection to mission execution, allowing the actual semantics of the language to be used to specify the correct operation of the system and thereby facilitating the validation of the system as a whole. A principled foundation where these issues are considered, for single platforms (vehicles) as well as for fleets of homogeneous or heterogeneous platforms, is essential for these types of systems to be accepted by aviation authorities and in society. The mission specification language used in the HDRC3 architecture is designed to allow partial mission specifications with constraints, including resource

37 A Distributed Hybrid Deliberative/Reactive Architecture

925

average time to progress a state 160

p false 100ms then true 1000ms p false 1000ms then true 1000ms p false 100ms then true 4000ms p always true

140

time (milli seconds)

120 100 80 60 40 20 0

0

500

1000

1500 2000 number of formulas

2500

3000

3500

Fig. 37.28 Average progression time: always :p ! eventually always p

requirements and temporal deadlines. It extends the support for highly expressive elementary actions in Temporal Action Logic (TAL, Sect. 37.7.1) with temporal composite actions that also provide a formal semantics for the high-level structure of a mission (Doherty et al. 2012). The composite action constructs in the extended logic TALF correspond closely to the control structures supported by Task Specification Trees (Sect. 37.4.3). As shown in Fig. 37.29, missions can therefore be defined either in TALF or as TSTs and can be translated in either direction. The delegation functionality discussed in Sect. 37.8 can then be used to delegate a mission in the shape of a TST to one or more platforms for execution, resulting in a distributed task structure. Composite actions in TALF are characterized recursively through the general construct “ with VARS do TASK where CONS ”: Any composite action consists of a task TASK that should be executed in a context characterized by a set of variables VARS constrained by a set of constraints CONS . The TASK , in turn, can be an elementary TAL action or consist of a combination of composite actions using constructs such as sequential composition, parallel (concurrent) composition, conditional, (unbounded) loops, while-do, and a concurrent for-each operator allowing a variable number of actions to be executed concurrently. At the mission specification level considered here, each constraint definition can be as general as a logical formula in TAL, giving it a formal semantics. For pragmatic use in a robotic architecture, a wide class of formulas can be automatically transformed into constraints processed

926

P. Doherty et al.

Varying means of generating mission specifications Graphical interface for specific mission type

Text editor

Automated planner

Mission specification in Temporal Action Logic (TALF)

Mission specification as Task Specification Tree (TST)

TALF provides a formal semantics for Task Specification Trees

TSTs provide an executable, distributable structure for TALF

Delegation

Task Specification Tree, delegated and distributed Agent 2

S C

E E

E

Agent 7

Each node associated with a platformspecific executor

Fig. 37.29 Mission specification, translation, and delegation

by a constraint satisfaction solver, allowing a robotic system to formally verify the consistency of a (distributed) task through the use of (distributed) constraint satisfaction techniques. A composite action-type specification declares a named composite action. This is useful in order to define a library of meaningful composite actions to be used in mission specifications. Each specification is of the form Œt; t 0 comp.Nv/ A.t; t 0 ; vN / where comp.Nv/ is a composite action term such as monitor-pattern.x; y; dist/, consisting of an action name and a list of parameters, and A.t; t 0 ; vN / is a composite action expression where only variables in ft; t 0 g [ vN may occur free. A composite action expression (C-ACT), in turn, allows actions to be composed at a high level of abstraction using familiar programming language constructs such as sequences (AI B), concurrency (A jj B), conditions, and loops. The associated syntax is defined as follows:

37 A Distributed Hybrid Deliberative/Reactive Architecture

927

WWD Œ;  0  with xN do TASK where  TASK WWD Œ;  0  ELEM- ACTION - TERM j Œ;  0  COMP-ACTION-TERM j .C-ACT I C-ACT / j .C-ACT jj C-ACT / j if Œ then C- ACT else C- ACT j while Œ do C- ACT j foreach xN where Œ do conc C- ACT C- ACT

where xN is a potentially empty sequence of variables (where the empty sequence can be written as ),  is a TAL formula representing a set of constraints, ELEM- ACTION - TERM is an elementary action term such as fly-to.av; x; y/, COMP ACTION - TERM is a composite action term, and Œ is a TAL formula referring to facts at a single timepoint . For brevity, omitting “with xN do” is considered equivalent to specifying the empty sequence of variables, and omitting “where ” is equivalent to specifying “where TRUE.” Note that the I and jj constructs are easily extended to allow an arbitrary number of actions, as in .AI BI C I D/. Like elementary actions, every composite action C-ACT is annotated with a temporal execution interval. This also applies to each “composite sub-action.” For example, Œt1 ;t2  with av; t3 ; t4 ; t5 ; t6 do  Œt3 ; t4 fly-to.av; x; y/I Œt5 ; t6 collect-video.av; x; y/ where Œt1 has-camera.av/ denotes a composite action where one elementary action takes place within the interval Œt3 ; t4 , the other one within the interval Œt5 ; t6 , and the entire sequence within Œt1 ; t2 . The with-do-where construct provides a flexible means of constraining variables as desired for the task at hand. In essence, “Œt1 ; t2  with xN do TASK where ” states that there exists an instantiation of the variables in xN such that the specified TASK (which may make use of xN as illustrated above) is executed within the interval Œt1 ; t2  in a manner satisfying . The constraint  may be a combination of temporal, spatial, and other types of constraints. Above, this constraint is used to ensure the use of an av that has a camera rather than an arbitrary av. As the aim is to maximize temporal flexibility, the sequence operator (;) does not implicitly constrain the two actions fly-to and collect-video to cover the entire temporal interval Œt1 ; t2 . Instead, the actions it sequentializes are only constrained to occur somewhere within the execution interval of the composite action, and gaps are permitted between the actions – but all actions in a sequence must occur in the specified order without overlapping in time. Should stronger temporal constraints be required, they can be introduced in a where clause. For example, t1 D t3 ^ t4 D t5 ^ t6 D t2 would disallow gaps in the sequence above. Also, variations such as gapless sequences can easily be added as first-class language constructs if desired.

928

P. Doherty et al.

Formal Semantics. To define a formal semantics for TALF, the base logic L(FL) is first extended with support for fixpoint formulas. This is required for unbounded loops and unbounded recursion, which cannot be characterized in the first-order logic L(FL). A translation is then defined from TALF into the extended base logic L(FLFP ). As a result of this translation, a composite action is a theory in L(FLFP ). Questions about missions thereby become queries relative to an inference mechanism, allowing operators to analyze mission properties both during pre- and post-mission phases. This also provides a formal semantics for Task Specification Trees, which can be translated into TALF, and thereby for the execution system used in the HDRC3 architecture. Details regarding translations and the resulting formal semantics are specified in Doherty et al. (2012). Examples. In November 2011, a powerful earthquake off the coast of Japan triggered a tsunami with devastating effects, including thousands of dead and injured as well as extensive damage to cities and villages. Another effect, which became increasingly apparent over time, was the extensive damage to the Fukushima Daiichi nuclear plant which later resulted in a complete meltdown in three reactors. The exact level of damage was initially difficult to assess due to the danger in sending human personnel into such conditions. Here unmanned aircraft could immediately have assisted in monitoring radiation levels and transmitting video feeds from a closer range. Several composite actions that can be useful for the problem of information gathering in this situation will now be considered. The focus is on demonstrating the L(ND) composite action constructs, and some aspects of the actions below are simplified for expository reasons. Assume the existence of a set of elementary actions whose meaning will be apparent from their names and from the explanations below: hover-at, fly-to, monitor-radiation, collect-video, and scan-cell. Each elementary action is assumed to be defined in standard TAL and to provide suitable preconditions, effects, resource requirements, and (completely or incompletely specified) durations. For example, only an AV with suitable sensors can execute monitor-radiation. In the following composite action, an AV hovers at a location .xav ; yav / while using its onboard sensors to monitor radiation and collect video at .xtarg ; ytarg /: Œt; t 0 monitor-single.av; xav ; yav ; xtarg ; ytarg / Œt; t 0 with t1 ; t2 ; t3 ; t4 ; t5 ; t6 do . Œt1 ; t2 hover-at.av; xav ; yav / jj Œt3 ; t4 monitor-radiation.av; xtarg ; ytarg / jj Œt5 ; t6 collect-video.av; xtarg ; ytarg / / where Œtsurveil-equipped.av/ ^ radiation-hardened.av/ ^ t1 D t3 D t5 D t ^ t2 D t4 D t6 D t 0 The first part of the constraint specified in the where clause ensures that an AV involved in a monitoring action is equipped for surveillance and is radiationhardened (in addition to the conditions placed on monitor-radiation, which include

37 A Distributed Hybrid Deliberative/Reactive Architecture

929

the existence of radiation sensors). The temporal constraints model a requirement for these particular actions to be synchronized in time and for the AV to hover in a stable location throughout the execution of monitor-single. These constraints could easily be relaxed, for example, by stating that hovering occurs throughout the action but monitoring occurs in a subinterval. The following action places four AVs in a diamond pattern to monitor a given location such as a nuclear reactor at a given distance, counted in grid cells. The AVs involved are not specified as parameters to the monitoring action but are chosen freely among available AVs, subject to the constraints modeled by sub-actions such as monitor-single: Œt; t 0monitor-pattern.x; y; dist/ Œt; t 0with s1 ; : : : ; w4 ; av1 ; av2 ; av3 ; av4 do . .Œs1 ; s2  fly-to.av1 ; x C dist; y/I Œs3 ; s4  monitor-single.av1 ; x C dist; y; x; y// jj .Œu1 ; u2  fly-to.av2 ; x  dist; y/I Œu3 ; u4  monitor-single.av2 ; x  dist; y; x; y// jj .Œv1 ; v2  fly-to.av3 ; x; y C dist/I Œv3 ; v4  monitor-single.av3 ; x; y C dist; x; y// jj .Œw1 ; w2  fly-to.av4 ; x; y  dist/I Œw3 ; w4  monitor-single.av4 ; x; y  dist; x; y/// where

s3 D u3 D v3 D w3 ^ s4 D u4 D v4 D w4 ^ s4  s3  minduration

Four sequences are executed in parallel. Within each sequence, a specific AV flies to a suitable location and then monitors the target. The target must be monitored simultaneously by all four AVs (s3 D u3 D v3 D w3 and s4 D u4 D v4 D w4 ), while s4  s3  minduration ensures this is done for at least the specified duration. As flying does not need to be synchronized, the intervals for the fly-to actions are only constrained implicitly through the definition of a sequence. For example, the translation ensures that t  s1  s2  s3  s4  t 0 , so that each fly-to must end before the corresponding monitor-single. All grid cells must also be scanned for injured people. The following generic action uses all available AVs with the proper capabilities, under the assumption that each such AV has been assigned a set of grid cells to scan. An assignment could be generated by another action or provided as part of the narrative specification. For clarity, this includes several clauses (with do, where TRUE) that could easily be omitted: Œt; t 0  scan-with-all-uavs./ Œt; t 0  with do foreach av where Œtcan-scan.av/ do conc Œt; t 0  with u; u0 do Œu; u0  scan-for-people.av/ where TRUE where TRUE

As shown below, each AV involved in this task iterates while there remains at least one cell .x; y/ that it has been assigned (“owns”) and that is not yet scanned. In each iteration the variables .x 0 ; y 0 / declared in the nested with clause range over

930

P. Doherty et al.

arbitrary coordinates, but the associated where clause ensures that only coordinates that belong to the given AV and that have not already been scanned can be selected. Also in each iteration, tc is bound to the time at which the constraint condition is tested and u; u0 are bound to the timepoints at which the inner composite action is performed. The repeated use of u; u0 is intentional. The scan-cell action will occur over exactly the same interval as the enclosing composite action construct: Œt; t 0  scan-for-people.av/ Œt; t 0  with do while Œtc  9x; yŒowns.av; x; y/ ^ :scanned.x; y/ do Œu; u0  with x 0 ; y 0 do Œu; u0 scan-cell.av; x 0 ; y 0 /where Œtc owns .av; x 0 ; y 0 / ^ :scanned.x 0 ; y 0 / where TRUE

It is now possible to define a small mission to occur within the interval Œ0; 1000, where scanning may use the entire interval while the grid cell .20; 25/ is monitored at a distance of 3 cells and must terminate before time 300: Œ0; 1000.Œ0; 1000 scan-with-all-uavs./ jj Œ0; 300 monitor-pattern.20; 25; 3// It should be emphasized that in the expected case, the task of generating specifications of this kind would be aided by libraries of predefined domain-related actions as well as by user interfaces adapted to the task at hand. The structure and highlevel nature of the language remains important when ensuring that these tools and their output are both correct and comprehensible to a human operator inspecting a mission definition.

37.8

Collaborative Systems

Though the main focus of this chapter has been on the operation of a single unmanned aircraft, issues related to cooperation and collaboration are also essential for the successful use of such systems. At the cooperative level, the combination of an aircraft or other robotic platform and its associated software is viewed as an agent. Humans interacting with platforms through, for example, ground control stations and other interfaces are also considered to be agents. Taken together, these agents form a collaborative system where all participants can cooperate to perform missions. One aspect of a collaborative system is that all agents are conceptually equal and independent in the sense that there is no predefined control structure, hierarchical or otherwise. A consequence is that the control structure can be determined on a mission-to-mission basis and dynamically changed during a mission.

37 A Distributed Hybrid Deliberative/Reactive Architecture

931

When an unmanned system is viewed as an agent acting on behalf of humans, it is also natural to view the assignment of a complex mission to that system as delegation – by definition, the act of assigning authority and responsibility to another person, or in a wider sense an agent, in order to carry out specific activities. Delegation is therefore part of the foundation for collaboration in this architecture: A delegator can ask a contractor to take the responsibility for a specific task to be performed under a set of constraints. Informally, an agent receiving a delegation request must verify that to the best of its knowledge, it will be able to perform the associated task under the given constraints, which may for example, concern resource usage or temporal and spatial aspects of the mission. To ensure that a complex task is carried out in its entirety, an agent may have to enlist the aid of others, which can then be delegated particular parts of the task. This results in a network of responsibilities between the agents involved and can continue down to the delegation of elementary, indivisible actions. If such a network can be generated in a way that satisfies the associated constraints, the contractor can accept the delegation and is then committed to doing everything in its power to ensure the task is carried out. If not, it must refuse. See Doherty et al. (2013) for a formal characterization of delegation in terms of speech acts as well as a practical delegation protocol that also determines how to allocate tasks to specific agents. Delegation requires a flexible task structure with a clear formal semantics. This role is played by Task Specification Trees (Sect. 37.4.3), whose hierarchical nature also leads to a natural recursive decomposition of tasks where the children of a node are the subtasks that can be re-delegated to other agents. For example, an automated multi-agent planner (Sect. 37.7.2) can generate a TST where elementary action nodes are not assigned to specific agents. Then, the delegation process and its embedded task allocation functionality (Doherty et al. 2013) can recursively determine how the plan can be executed in a way that satisfies associated constraints. The formal semantics of the task being delegated is specified through a close connection to TALF (Sect. 37.7.4). See Doherty et al. (2011, 2013) and Doherty and Meyer (2012) for further details about delegation, its close relation to adjustable autonomy and mixed-initiative interaction, and its integration with automated planning. Legacy Systems. When an agent-based architecture is used together with an existing platform such as an unmanned aircraft, there may already be an existing legacy system providing a variety of lower-level functionalities such as platformspecific realizations of elementary tasks and resources. Existing interfaces to such functionalities can vary widely. The current instantiation of the architecture (Fig. 37.30a) directly supports the use of such legacy functionalities through the use of an agent layer and a gateway. The agent layer (Fig. 37.30b) encapsulates higher-level deliberative functionalities and provides a common interface for multiagent collaboration in complex missions, including support for mission specification languages, delegation, and planning. The gateway must have a platform-specific implementation, but provides a common platform-independent external interface to

932

P. Doherty et al.

a Interface Agent Layer Delegation Capability Resource Reasoning Cap.

Task Exec Capability

Task Planning Capability

Motion Planning Capability

…

Gateway Legacy System Legacy Functionalities

Platform / Ground Station

b

Collaborative Human Robot System Platform 1

Platform n

Operator 1

Operator m

Agent Layer

Agent Layer

Agent Layer

Agent Layer

Legacy System

Legacy System

Legacy System

Legacy System

Fig. 37.30 Agents in a collaborative system. (a) Agentified platform or ground control station. (b) Overview of the collaborative human/robot system

the available legacy functionalities. In essence, this allows newly developed higherlevel functionalities to be seamlessly integrated with existing systems, without the need to modify either the agent layer or the existing system. The agent layer can then be developed independently of the platforms being used.

37 A Distributed Hybrid Deliberative/Reactive Architecture

933

Legacy control stations and user interfaces that human operators use to interact with robotic systems are treated similarly, through the addition of an agent layer. The result is a collaborative human/robot system consisting of a number of human operators and robotic platforms each having an agent layer and possibly a legacy system, as shown in Fig. 37.30b.

37.9

Mission Applications

The research methodology used during the development of the HDRC3 architecture has been very much scenario-based where very challenging scenarios out of reach of current systems are specified and serve as longer-term goals to drive both theoretical and applied research. Most importantly, attempts are always made to close the theory/application loop by implementing and integrating results in AVs and deploying them for empirical testing at an early stage. One then iterates and continually increases the robustness and functionalities of the targeted components. Due to the architecture described in the previous sections, it is relatively easy to build on top of existing functionalities and to add new ones in order to put together sophisticated autonomous missions. Below, two such example mission applications are described.

37.9.1 Emergency Services Assistance The first application focuses again on the ambitious emergency services scenario discussed in the introduction. An emergency relief scenario in this context can be divided into two separate legs or parts.

37.9.1.1 Mission Leg I: Body Identification In the first leg of the mission, a large region should be cooperatively scanned with one or more AVs to identify injured civilians. The result of this scan is a saliency map pinpointing potential victims, their locations, and associated information such as high resolution photos and thermal images. This information could then be used directly by emergency services or passed on to other AVs as a basis for additional tasks. A multi-platform area coverage path-planning algorithm computes paths for n heterogeneous platforms guaranteeing complete camera coverage, taking into account sensor properties and platform capabilities. Two video sources (thermal and color) are used, allowing for high rate human detection at larger distances than in the case of using the video sources separately with standard techniques. The high processing rate is essential in case of video collected onboard an AV in order not to miss potential victims. A thermal image is first analyzed to find human body sized silhouettes. The corresponding regions in a color image are subjected to a human body classifier which is configured to allow weak classifications. This focus of attention allows

934

P. Doherty et al.

Algorithm 4 Saliency map construction 1: Initialize data structures 2: while scanning not finished do 3: Simultaneously grab two images: i mgcolor, i mgthermal 4: Analyze i mgthermal to find potential human body regions 5: for each region in i mgthermal do 6: Find corresponding region rcolor in i mgcolor 7: Compute geographical location loc of rcolor 8: Execute human body classifier on rcolor 9: if classification positive then 10: if loc is new then 11: add location loc to map, initialize certainty factor pbody .loc/ 12: else 13: update certainty factor pbody .loc/ 14: end if 15: end if 16: end for 17: end while

for maintaining body classification at a rate up to 25 Hz. This high processing rate allows for collecting statistics about classified humans and pruning false classifications of the “weak” human body classifier. Detected human bodies are geolocalized on a map which can be used to plan supply delivery. The technique presented has been tested onboard the UASTL RMAX helicopter platform and is an important component in the lab’s research with autonomous search and rescue missions. Saliency Map Construction. Information obtained from thermal and color video streams must be fused in order to create saliency maps of human bodies. An overview of the method used is presented in Algorithm 4. The execution of this algorithm starts when the host AV arrives at the starting position of the area to scan and is terminated when the scanning flight is finished. Its output is a set of geographical locations loci and certainty factors pbody .loci /. Refer to Doherty and Rudol (2007) and Rudol and Doherty (2008) for additional details. After initializing the necessary data structures (line 1), the algorithm enters the main loop (line 2), which is terminated when the entire area has been scanned. The main loop begins with simultaneously grabbing two video frames. The thermal image is analyzed first (line 4) to find a set of regions of intensities which correspond to human body temperatures (details below). Then (line 6), for each of these subregions, a correspondence in the color frame, as well its geographical location loc, is calculated (details below). The calculated corresponding region in the color frame is analyzed with a human body classifier to verify the hypothesis that the location loc contains a human body (details below). If the classification is positive and the location loc has not been previously identified, then loc is added to the map and its certainty factor initialized (line 11). Otherwise, the certainty factor of that location is updated (line 13, details below).

37 A Distributed Hybrid Deliberative/Reactive Architecture

935

Fig. 37.31 Example input to the image processing algorithm: Thermal images and corresponding color images

Thermal Image Processing. The image processing algorithm takes a pair of images as input and starts by analyzing the thermal image (top row of Fig. 37.31). The image is first thresholded to find regions of certain intensities which correspond to human body temperature. The image intensity corresponding to a certain temperature is usually given by the camera manufacturer or can be calibrated by the user. The shapes of the thresholded regions are analyzed, and those which do not resemble a human body, due to the wrong ratio between minor and major axes of the fitted ellipse or due to incorrect sizes, are rejected. Once human body candidates are found in the thermal image, corresponding regions in the color image are calculated. Image Correspondences and Geolocation. Finding corresponding regions using image registration or feature matching techniques is infeasible because of the different appearance of features in color and thermal images. Therefore, a closed form solution, which takes into account information about the cameras’ pose in the world, is preferred. In short, a geographical location corresponding to pixel coordinates for one of the cameras is calculated. It takes into account the camera’s extrinsic and intrinsic parameters and assumes a flat world. The obtained geographical location is then projected back into the other camera image. The method can be extended to relax the flat world assumption given the elevation model. A geographical location of a target can be found by performing ray tracing along the line going from the camera center through a pixel to find the intersection with the ground elevation map. The accuracy of the correspondence calculation is influenced by several factors. All inaccuracies of parameters involved in calculating the position and attitude of cameras in the world contribute to the overall precision of the solution. The evaluation of the accuracy involves investigating the AV state estimation errors, pantilt unit position accuracy, camera calibration errors, etc. In the case of the UASTL

936

P. Doherty et al.

Thermal hypotheses sub-windows

All sub-windows or

Ture Classifier 1 False Sub-window rejected

Classifier 2

Ture

False Sub-window rejected

Classifier n

Ture

Positive classification

False Sub-window rejected

Fig. 37.32 Schematic description of a cascade classifier

RMAX, during the experimental validation, the corresponding image coordinates were within a subregion of 20 % of the video frame size (see the marked subregions in the color images in Fig. 37.31). Human Body Classification. After calculating the coordinates of the pixel in the color image, a region with the Pc as center (the black rectangles in the bottom row images of Fig. 37.31) is analyzed by an object classifier. The classifier used was first suggested in Viola and Jones (2001). It uses a cascade of classifiers for object detection and includes a novel image representation, the integral image, for quick detection of features. Classifiers in the initial stages remove a large number of negative examples. Successive classifiers process a smaller number of sub-windows. The initial set of sub-windows can include all possible sub-windows or can be the result of a previous classification in a thermal image additionally improving processing rate. The method was also improved, for example, in Lienhart and Maydt (2002) by extending the original feature set. The classifier requires training with positive and negative examples. During the learning process, the structure of a classifier is built using boosting. The use of a cascade of classifiers allows for dramatic speed up of computations by skipping negative instances and only computing features with high probability for positive classification. The speed up comes from the fact that the classifier, as it slides a window at all scales, works in stages and is applied to a region of interest until at some stage the candidate is rejected or all the stages are passed (see Fig. 37.32). This way, the classifier quickly rejects subregions that most probably do not include the features needed for positive classification (i.e., background processing is quickly terminated). The implementation of the classifier used in this work is a part of the Open Source Computer Vision Library (http://opencv.org/), and the trained classifier for upper, lower, and full human body is a result of Kruppa et al. (2003). The classifier is best suited for pedestrian detection in frontal and backside views which is exactly the type of views an AV has when flying above the bodies lying on the ground.

37 A Distributed Hybrid Deliberative/Reactive Architecture

937

The classifier parameters have been adjusted to minimize false negatives: In case of rescue operations, it is better to find more false positives than to miss a potential victim. The number of neighboring rectangles needed for successful identification has been set to 1 which makes the classifier accept very weak classifications. The search window is scaled up by 20 % between subsequent scans. Map Building. Since the body classifier is configured to be “relaxed,” it delivers sporadic false-positive classifications. The results are pruned as follows. Every salient point in the map has two parameters which are used to calculate the certainty of a location being a human body: Tframe which describes the amount of time a certain location was in the camera view and Tbody which describes the amount of time a certain location was classified as a human body. The certainty factor is Tbody . A location is considered a body if pbody .loci / calculated by pbody .loci / D Tframe is larger than a certain threshold (0.5 was used during the flight tests) and Tframe is larger than a desired minimal observation time. Locations are considered equal if geographical distance between them is smaller than a certain threshold (depending on the geolocation accuracy) and the final value of a geolocated position is an average of the observations. Experimental Validation. The presented technique for geolocation and saliency map building has been integrated as a perception functionality of the Control Kernel (Sect. 37.3.2) and tested in flight. A mission has then been constructed in the shape of a Task Specification Tree that indirectly makes use of the takeoff, hovering, path-following, and landing flight control modes (through FCL commands, Sect. 37.4.2.2), as well as the pan-tilt unit (through PPCL commands, Sect. 37.4.2.3). Test flights were then performed at the Swedish Rescue Services Agency Test Area (Fig. 37.33a). Since this is a city-like closed area with road structures, buildings, etc., the video streams included different types of textures such as grass, asphalt, gravel, water, and building rooftops. An example complete push button mission setup was as follows: • Two UASTL RMAX helicopters were used starting from H1 and H2 in Fig. 37.33b. • An operator selected a rectangular area on a map where the saliency map was to be built (Fig. 37.33b). • Onboard systems calculated the mission plan taking into account properties of the onboard sensors (such as the field of view) of both AVs. The plan consisted of two separate flight plans for the two AVs. • The mission started with simultaneous takeoffs and flying to starting positions S1 and S2 in Fig. 37.33b. After arriving at the starting positions, the execution of the scanning paths autonomously began, and the saliency map building was initiated. • Upon finishing the scanning paths at positions E1 and E2 , the AVs flew to the takeoff positions and performed autonomous landings.

938

a

P. Doherty et al.

b

H1

H2

S1

S2

E1

E2

Fig. 37.33 (a) Map of the Swedish Rescue Services Agency Test Area in Revinge, (b) A closeup view of the area where the saliency map was built. Approximate flight paths are marked with solid lines

Experimental Results. All 11 human bodies placed in the area were found and geolocated. Corresponding color and thermal images for the identified objects are displayed vertically as pairs in Fig. 37.34. Images 7, 9, and 14 present three falsely identified objects, caused by configuring the human body classifier to accept weak classifications. A more restrictive setup could instead result in missing potential victims. The images could additionally be judged by a human operator to filter out the false-positive classifications. Both human bodies and dummies were detected despite the lower temperature of the latter. Figure 37.35 presents the generated saliency map. The scanning pattern segments for the platform starting in position H2 is marked with a solid line, and its final position is marked with a cross icon. The fields of view of the color and thermal cameras are depicted with light-gray and black rectangles, respectively. These differ slightly as the two cameras have different properties as identified during calibration procedures. The circles indicate the identified body positions. The lighter the shade of the color, the more certain the classification. As can be seen, the most certain body positions are objects number 2 and 11 (in Fig. 37.34). It is due to the fact that these body images are clearly distinguishable from the homogeneous background. Nevertheless, even body images with more cluttered backgrounds were identified. The accuracy of the body geolocation calculation was estimated by measuring GPS positions (without differential correction) of bodies after an experimental flight. The measurement has a bias of approximately 2 m in both east and north directions. It is the sum of errors in GPS measurement, accuracy of the camera platform mounting, PTU measurement, and camera calibration inaccuracies. The spread of measurement samples of approximately 2.5 m in both east and north directions is a sum of errors of the AV’s attitude measurement, the system of springs in the camera platform, and time differences between the AV state estimate, PTU angle measurement, and image processing result acquisition.

37 A Distributed Hybrid Deliberative/Reactive Architecture

939

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Fig. 37.34 Images of classified bodies. Corresponding thermal images are placed under color images

940

P. Doherty et al.

Fig. 37.35 The resulting map with salient points marked as circles. The lighter the shade of the color, the higher the detection certainty

The presented algorithm requires only a single pair of images for human body classification. In practice, however, the more pairs available, the more certain the result of a classification can become. Additionally, thanks to using the results of the thermal image analysis to focus the classification in the color image subregions, a high rate of processing is achieved (above 20 Hz for the presented results).

37.9.1.2 Mission Leg II: Package Delivery After successful completion of leg I of the mission scenario, one can assume that a saliency map has been generated with geolocated positions of the injured civilians. In the next phase of the mission, the goal is to deliver configurations of medical, food, and water supplies to the injured. In order to achieve this leg of the mission, one would require a task planner to plan for logistics, a motion planner to get one or more AVs to supply and delivery points, and an execution monitor to monitor the execution of highly complex plan operators. Each of these functionalities would also have to be tightly integrated in the system. For these logistics missions, the use of one or more AVs with diverse roles and capabilities is assumed. Initially, there are n injured body locations, several supply depots, and several supply carrier depots (see Fig. 37.36). The logistics

37 A Distributed Hybrid Deliberative/Reactive Architecture

941

Fig. 37.36 A supply depot (left) and a carrier depot (right)

Fig. 37.37 The AV logistics simulator

mission is comprised of one or more AVs transporting boxes containing food and medical supplies between different locations (Fig. 37.37). Plans are generated in the millisecond to seconds range using TALplanner (see Sect. 37.7.2), and empirical testing shows that this approach is promising in terms of integrating high-level deliberative capability with lower-level reactive and control functionality.

942

P. Doherty et al.

Fig. 37.38 Two frames of video presenting the prototype mechanism for carrying and releasing packages using an electromagnet. An arrow points to a package being carried. The top left picture presents the onboard camera view

Achieving the goals of such a logistics mission with full autonomy requires the ability to pick up and deliver boxes without human assistance. Thus, each AV has a device for attaching to boxes and carrying them beneath the AV. The action of picking up a box involves hovering above the box, lowering the device, attaching to the box, and raising the device, after which the box can be transported to its destination. There can also be a number of carriers, each of which is able to carry several boxes. By loading boxes onto such a carrier and then attaching to the carrier, the transportation capacity of an AV increases manifold over longer distances. The ultimate mission for the AVs is to transport the food and medical supplies to their destinations as efficiently as possible using the carriers and boxes at their disposal. A physical prototype of a mechanism for carrying and releasing packages has been developed and tested. Figure 37.38 presents two images of the prototype system. The logistics scenario has also been tested in a simulated AV environment with hardware in-the-loop, where TALplanner generates a detailed mission plan which is then sent to a simulated execution system using the same helicopter flight control software as the physical AV. The execution monitor system has been tested in this simulation as well, with a large variety of deviations tested through fault injection in the simulation system. The simulator makes use of the Open Dynamics Engine (http://www.ode.org), a library for simulating rigid body dynamics, in order to realistically emulate the physics of boxes and carriers. This leads to effects such as boxes bouncing and rolling away from the impact point should they accidentally be dropped, which is also an excellent source of unexpected situations that can be used for validating both the domain model and the execution monitoring system.

37 A Distributed Hybrid Deliberative/Reactive Architecture

943

37.9.2 Map Building Using a Laser Range Finder The second application presented here deals with the construction of an elevation map using a laser range finder allowing AVs to navigate safely in unknown outdoor environments. As described in Sect. 37.5.1, the path-planning algorithms used here are based on a geometrical description of the environment to generate collision-free paths. The safety of a UAS operation therefore depends on having a sufficiently accurate 3D model of the environment. Predefined maps may become inaccurate or outdated over time because of the environment changes, for example, due to new building structures and vegetation growth. Therefore, adequate sensors and techniques for updating or acquiring new 3D models of the environment are necessary in many cases. Among the many sensors available for providing 3D information about an operational environment, laser range finders provide high-accuracy data at a reasonable weight and power consumption. One of the reasons for the innovation in this particular sensor technology is its wide use in many industries, but laser range finders have also received a great deal of interest from the robotics community, where their main usage is in navigation and mapping tasks for ground robotic systems, such as localization (Burgard et al. 1997), 2D Simultaneous Localization and Mapping (SLAM (Montemerlo and Thrun 2007)), 3D SLAM (includes 3D position (Cole and Newman 2006)), and 6D SLAM (includes 3D position and attitude (N¨uchter et al. 2004)). The device integrated with the UASTL RMAX system is the popular LMS-291 from SICK AG (http://www.sick.com). This unit does not require any reflectors or markers on the targets nor scene illumination to provide real-time measurements. It performs very well both in indoor and outdoor environments. The system is equipped with a rotating mirror which allows for obtaining distance information in one plane in front of the sensor with a selectable field of view of 100 or 180ı (Fig. 37.39). It gives a resolution of 1 cm with a maximum range of 80 m, or a resolution of 1 mm with a range of 8 m; an angular resolution of 0.25, 0.5, or 1.0ı ; and a corresponding response time of 53, 26, or 13 ms. The laser unit has been modified to reduce its weight from 4.5 to 1.8 kg. It has then been mounted on an in-house developed rotation mechanism supporting continuous rotation of the sensor around the middle laser beam (solid line in Fig. 37.39), which allows for obtaining half-sphere 3D point clouds even when the vehicle is stationary. A similar approach to the integration of the laser range finder with an UASTL RMAX platform is used by Whalley et al. (2008). System Integration. The 3D map building algorithm based on the data provided by LRF sensor has been integrated as a perception functionality of the Control Kernel (Sect. 37.3.2) and tested in flight. A Task Specification Tree node has been specified to achieve the required mission goals (Sect. 37.4.3). Similar to the previous application example, the takeoff, hovering, path-following, and landing flight control modes were used during the mission through Flight Control Language

944

P. Doherty et al.

Fig. 37.39 Top view of the LMS-291 scanning field and the axis of rotation using the rotation mechanism 40°

140°

0°

Fig. 37.40 Overview of the reconstructed elevation map of the Revinge flight test area based on the laser range finder data (left) and a photo of corresponding building structures (right)

commands (Sect. 37.4.2.2). Additionally, the Payload and Perception Control Language (Sect. 37.4.2.3) are used to set a constant angular speed of the LRF rotational mechanism as well as the distance and angular resolutions of the sensor. The 3D maps acquired are saved in the GIS Service database and available to the deliberative services (such as path planners, discussed in Sect. 37.5). Experimental Results. Several missions were performed during which both the LRF data and AV state estimates were collected and integrated. Figure 37.40 presents a reconstructed elevation map of the Revinge flight test area (the left side) focusing on two building structures. A photo of corresponding buildings is presented on the right side of the figure. The elevation map is built by sampling the LRF data with 1 m resolution and constructing a set of triangles in order to represent the elevation. In order to assess the fidelity of the newly generated model, an overlay with the existing model was generated. The result is presented in Fig. 37.41a. The new map

37 A Distributed Hybrid Deliberative/Reactive Architecture

945

Fig. 37.41 Results of the map building procedure. (a) Overlay of the new elevation map with the existing Revinge flight test area model. (b) Example of path planner use in the reconstructed map of the Revinge area

includes some changes in the environment, including a metal container on the left in the figure (A.) and new vegetation (B.) that was not present in the existing model. The new models stored by the GIS Service can be used by the AV platform for path planning in order to generate collision-free paths (see Sect. 37.5.1). Since generated models are based on noisy measurements, a safety margin during the planning is used. For the UASTL RMAX, a safety margin of 6 m is used. An example of path generated by the path planner using the new model is presented in Fig. 37.41b. The accuracy of models built with the raw LRF point clouds (without applying the scan matching algorithms) is sufficient for navigation purposes if the necessary safety margins are used. The inaccuracies introduced by the measurement errors and the uncertainty of the AV state estimate might result in narrowing down the operational environment of the AV. For example, in Fig. 37.41a a narrow passage (C.) can be excluded from the collision-free space although the corridor between the two buildings is wide enough to fly through. Further investigation of methods for improving the model quality is ongoing at the time of writing this chapter.

946

P. Doherty et al.

37.10 Conclusions The goal of this research is the development of autonomous intelligent systems for unmanned aircraft. This chapter has described a hybrid deliberative/reactive architecture which has been developed through the years and successfully deployed on a number of unmanned aircraft. The focus was on an instantiation of this architecture with the UASTL RMAX. The architecture was described in a manner that should be useful to the unmanned aircraft research community, because it isolates generic functionality and architectural solutions that can be transferred and used on existing unmanned aircraft and also those envisioned for the future. The experimental development of such systems has wider impact since the HDRC3 architecture and its instantiation on the UASTL RMAX are an example of a new type of software system commonly known as a software intensive system or cyber physical system. Such systems are characterized by a number of features: the complex combination of both hardware and software, the high degree of concurrency used, the distributed nature of the software, the network-centric nature of the environment in which they reside, the open nature of these systems, and finally, the requirement that they are both scalable and adaptive since specification of such systems is necessarily open due to their interaction with other components not part of the system. The latter features are particularly important in the context of collaborative missions with other heterogeneous systems. The HDRC3 architecture encapsulates many of these features. Emphasis has been placed on the importance of clean transitional interfaces between the different layers of the architecture which are inhabited by computational processes with diverse requirements in terms of temporal latency in the decision cycles and the degree to which the processes use internal models. The HDRC3 architecture is also highly modular and extensible. The use of HCSMs separates specification of the behavior of the system from the atomic conditions and activities which are implemented procedurally. This permits the use of a real-time interpreter for HCSMs and alleviates the requirement for recompilation when new HCSMs are added to the system. The approach is also amenable to distribution where federations of HCSM interpreters can intercommunicate with each other on a single platform as is the case with the UASTL RMAX system, but also across platforms if required. The Platform Server provides a clean declarative interface to both the suite of flight control modes and perception control modes through the use of two languages, FCL and PPCL, respectively. This approach provides a primitive language of building blocks that can be structured and used by other functionalities in the system to define higher-level task specifications. Extending the suite of basic actions in the UASTL RMAX system is done by defining new HCSMs which interface to new or existing control modes and then extending the language interface in the Platform Server. The task specifications themselves have a well-defined and extensible language, Task Specification Trees, which are used by both the reactive and deliberative layers of the architecture as a declarative specification language for tasks. Primitive actions

37 A Distributed Hybrid Deliberative/Reactive Architecture

947

defined through the Platform Server interface can be used with TSTs, or if required, new actions can be defined by specifying new node types. The execution and procedural implementation of these basic action and new node types is separated from their declarative specification through the use of node executors and a node execution facility in the reactive layer of the HDRC3 architecture. This clean separation has a number of advantages besides extensibility and modularity. A declarative specification of tasks permits the sharing of tasks implemented in different ways across platforms. This is especially important in the context of collaborative robotics and heterogeneous robotic systems. It also permits the translation of other task languages into a common representation. The translation of the output of automated task planners is a case in point. There is a direct mapping from the declarative output of a planner to grounded executable components in the architecture. The DyKnow system is another extensible and modular functionality. Any source in an architecture can become a stream-based data source. One naturally thinks of sensors as specific sources of data, but at another extreme, one can think of another platform as a source of streamed data or a data base or the Internet. The DyKnow system provides means for specifying, constructing, generating, using, and managing these diverse sources of data contextually at many different levels of abstraction. This is a unique component in the HDRC3 architecture that has had widespread use by other functionalities such as the execution monitoring system and the task and motion planners. Traditionally, verification and validation of autonomous systems has been a central research issue. In the context of unmanned aircraft, a great deal of research has focused on the Control Kernel and to some extent the lower parts of the reactive layer. In terms of the reactive layer itself and in particular the highly nondeterministic functionality associated with the deliberative layer, very little effort has been put into verification and validation because tools simply do not exist yet. A great deal of research effort has gone into improving this absence of effort during the development of key functionalities in the HDRC3 architecture. Many of the functionalities there, such as the task planner, the task specification language, the mission specification language, and the execution monitoring system, each have a formal semantics based on the use of Temporal Action Logics. In fact the use of logic is highly integrated with many of the processes evoked by deliberative functionality. The execution monitoring system is essentially a real-time dynamic model-checking system where temporal models are generated dynamically by DyKnow during the course of operation, and various constraints are checked relative to those models by evaluating formulas through progression and checking for their satisfiability. The output of TALplanner is a formal narrative structure in Temporal Action Logic. Consequently both the process of generating a plan and the resulting output can be reasoned about using inference mechanisms associated with TALplanner. In summary, this chapter has described an empirically tested unmanned aircraft architecture that combines many years of both engineering and scientific insight acquired through the successful deployment of the UASTL RMAX system in highly

948

P. Doherty et al.

complex autonomous missions. To the extent possible, these insights have been described in a generic manner in the hope that many of the functionalities described can serve as a basis for use by others in the continued and exciting development of such systems. Acknowledgments This work is partially supported by the EU FP7 project SHERPA (grant agreement 600958); the Swedish Foundation for Strategic Research (SSF) Collaborative Unmanned Aircraft Systems (CUAS) project; the Swedish Research Council (VR) Linnaeus Center for Control, Autonomy, and Decision-making in Complex Systems (CADICS); and the ELLIIT network organization for Information and Communication Technology.

References J.S. Albus, F.G. Proctor, A reference model architecture for intelligent hybrid control systems, in IFAC 13th Triennial World Congress, San Francisco (IFAC, 1996), pp. 483–488 R.C. Arkin, Behavior-Based Robotics (MIT, Cambridge, 1998) F. Bacchus, F. Kabanza, Planning for temporally extended goals. Ann. Math. Artif. Intell. 22, 5–27 (1998) K. Ben Lamine, F. Kabanza, Reasoning about robot actions: a model checking approach, in Revised Papers from the International Seminar on Advances in Plan-Based Control of Robotic Agents, Dagstuhl Castle (Springer, 2002), pp. 123–139 G. Berry, The Esterel synchronous programming language: design, semantics, implementation. Sci. Comput. Program. 19, 87–152 (1992). doi:10.1016/0167-6423(92)90005-V G. Booch, J. Rumbaugh, I. Jacobson, The Unified Modeling Language User Guide (AddisonWesley, Upper Saddle River, 2005). ISBN:978-0321267979 R. Brooks, A robot that walks: emergent behaviors from a carefully evolved network, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Scottsdale, 1989, pp. 692–696 A. Brooks, T. Kaupp, A. Makarenko, S. Williams, A. Oreb¨ack, Towards component-based robotics, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Edmonton, 2005 W. Burgard, D. Fox, S. Thrun, Active mobile robot localization, in Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), Nagoya (Morgan Kaufmann, 1997) D. Cole, P. Newman, Using laser range data for 3D SLAM in outdoor environments, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Orlando, 2006 G. Conte, Navigation functionalities for an autonomous UAV helicopter. Licentiate thesis 1307, Department of Computer and Information Science, Link¨oping University, 2007 G. Conte, Vision-based localization and guidance for unmanned aerial vehicles. Ph.D. thesis, Department of Computer and Information Science, Link¨oping University, Link¨oping, 2009 G. Conte, P. Doherty, Vision-based unmanned aerial vehicle navigation using geo-referenced information. EURASIP J. Adv. Signal Process. 2009, 1–18 (2009) G. Conte, S. Duranti, T. Merz, Dynamic 3D path following for an autonomous helicopter, in Proceedings of the IFAC Symposium on Intelligent Autonomous Vehicles, Lisbon, 2004 J. Cremer, J. Kearney, Y. Papelis, HCSM: a framework for behavior and scenario control in virtual environments. ACM Trans. Model. Comput. Simul. 5, 242–267 (1995) G. De Giacomo, R. Reiter, M. Soutchanski, Execution monitoring of high-level robot programs, in Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning (KR), Trento (Morgan Kaufmann, 1998), pp. 453–465 P. Doherty, J. Kvarnstr¨om, TALplanner: an empirical investigation of a temporal logic-based forward chaining planner, in 6th International Workshop on Temporal Representation and Reasoning (TIME-99), Orlando (IEEE Computer Society, 1999), pp. 47. ISBN:0-7695-0173-7

37 A Distributed Hybrid Deliberative/Reactive Architecture

949

P. Doherty, J. Kvarnstr¨om, TALplanner: a temporal logic-based planner. Artif. Intell. Mag. 22(3), 95–102 (2001) P. Doherty, J. Kvarnstr¨om, Temporal action logics, in The Handbook of Knowledge Representation, chapter 18, ed. by F. Van Harmelen, V. Lifschitz, B. Porter (Elsevier, Amsterdam/Boston, 2008), pp. 709–757 P. Doherty, J.-J.C. Meyer, On the logic of delegation: relating theory and practice, in The Goals of Cognition. Essays in Honor of Cristiano Castelfranchi, ed. by F. Paglieri, L. Tummolini, R. Falcone, M. Miceli (College Publications, London, 2012) P. Doherty, P. Rudol, A UAV search and rescue scenario with human body detection and geolocalization, in AI’07: Proceedings of the Australian Joint Conference on Advances in Artificial Intelligence, Gold Coast (Springer, 2007), pp. 1–13. ISBN:3-540-76926-9, 978-3540-76926-2 P. Doherty, J. Gustafsson, L. Karlsson, J. Kvarnstr¨om, (TAL) temporal action logics: language specification and tutorial. Electron. Trans. Artif. Intell. 2(3–4), 273–306 (1998). ISSN 1403–3534 P. Doherty, P. Haslum, F. Heintz, T. Merz, P. Nyblom, T. Persson, B. Wingman, A distributed architecture for intelligent unmanned aerial vehicle experimentation, in Proceedings of the International Symposium on Distributed Autonomous Robotic Systems, Toulouse, 2004 P. Doherty, J. Kvarnstr¨om, F. Heintz, A temporal logic-based planning and execution monitoring framework for unmanned aircraft systems. J. Auton. Agents Multi-Agent Syst. 19(3), 332–377 (2009) P. Doherty, F. Heintz, D. Land´en. A distributed task specification language for mixed-initiative delegation, in Principles and Practice of Multi-Agent Systems: 13th International Conference, PRIMA 2010, Kolkata, India, November 12–15, 2010, Revised Selected Papers, ed. by N. Desai, A. Liu, M. Winikoff. Lecture Notes in Computer Science, vol. 7057 (Springer, Berlin/Heidelberg, 2010), pp. 42–57. ISBN: 978-3-642-25919-7, e-978-3-64225920-3. DOI: 10.1007/978-3-642-25920-3 4 P. Doherty, F. Heintz, D. Land´en, A delegation-based architecture for collaborative robotics, in Agent-Oriented Software Engineering XI: Revised Selected Papers, Toronto. Lecture Notes in Computer Science, vol. 6788 (Springer, 2011), pp. 205–247 P. Doherty, J. Kvarnstr¨om, A. Szalas, Temporal composite actions with constraints, in Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning (KR), Rome, 2012, pp. 478–488 P. Doherty, F. Heintz, J. Kvarnstr¨om, High-level mission specification and planning for collaborative unmanned aircraft systems using delegation. Unmanned Syst. 1(1):75–119 (2013). ISSN:2301-3850, EISSN:2301-3869 S. Duranti, G. Conte, In-flight identification of the augmented flight dynamics of the RMAX unmanned helicopter, in Proceedings of the IFAC Symposium on Automatic Control in Aerospace, Toulouse, 2007 M. Egerstedt, X. Hu, A. Stotsky, Control of mobile platforms using a virtual vehicle approach. IEEE Trans. Autom. Control 46(11), 1777–1782 (2001) J. Eker, J. Janneck, E. Lee, J. Liu, X. Liu, J. Ludvig, S. Neuendorffer, S. Sachs, Y. Xiong, Taming heterogeneity – the Ptolemy approach. Proc. IEEE 91, 127–144 (2003) E.A. Emerson, Temporal and modal logic, in Handbook of Theoretical Computer Science, Volume B: Formal Models and Semantics, ed. by J. van Leeuwen (Elsevier/MIT, Amsterdam/Cambridge, 1990), pp. 997–1072 J.L. Fern´andez, R.G. Simmons, Robust execution monitoring for navigation plans, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Victoria, 1998 M. Fichtner, A. Grossmann, M. Thielscher, Intelligent execution monitoring in dynamic environments. Fundam. Inform. 57(2–4), 371–392 (2003). ISSN:0169-2968 R.J. Firby, An investigation into reactive planning in complex domains, in Proceedings of the National Conference on Artificial Intelligence, Seattle, 1987, vol. 1, pp. 202–206

950

P. Doherty et al.

Foundation for Intelligent Physical Agents, FIPA communicative act library specification (2002), http://www.fipa.org E. Gat, On three-layer architectures, in Artificial Intelligence and Mobile Robots, ed. by D. Kortenkamp, R.P. Bonasso, R. Murphy (AAAI/MIT, Menlo Park, 1997) E. Gat, M.G. Slack, D.P. Miller, R.J. Firby, Path planning and execution monitoring for a planetary rover, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Cincinnati (IEEE Computer Society, 1990), pp. 20–25 M. Ghallab, On chronicles: representation, on-line recognition and learning, in Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning (KR), Cambridge, 1996, pp. 597–607. ISBN 1-55860-421-9 S. Gottschalk, M.C. Lin, D. Manocha, OBBTree: a hierarchical structure for rapid interference detection, in Proceedings of the Annual Conference on Computer Graphics and Interactive Techniques, New Orleans, 1996 D. Harel, Statecharts: a visual formalism for complex systems. Sci. Comput. Program. 8(3), 231–274 (1987) F. Heintz, DyKnow: a stream-based knowledge processing middleware framework. Ph.D. thesis, Department of Computer and Information Science, Link¨oping University, 2009 F. Heintz, P. Doherty, DyKnow: an approach to middleware for knowledge processing. J. Intell. Fuzzy Syst. 15(1), 3–13 (2004) F. Heintz, P. Doherty, DyKnow: a knowledge processing middleware framework and its relation to the JDL data fusion model. J. Intell. Fuzzy Syst. 17(4), 335–351 (2006) F. Heintz, Z. Dragisic, Semantic information integration for stream reasoning, in Proceedings of the International Conference on Information Fusion, Singapore, 2012 F. Heintz, P. Rudol, P. Doherty, From images to traffic behavior – a UAV tracking and monitoring application, in Proceedings of the International Conference on Information Fusion, Qu´ebec, 2007, pp. 1–8 F. Heintz, J. Kvarnstr¨om, P. Doherty, Bridging the sense-reasoning gap: DyKnow – stream-based middleware for knowledge processing. Adv. Eng. Inform. 24(1), 14–26 (2010) F. Heintz, J. Kvarnstr¨om, P. Doherty, Stream-based hierarchical anchoring. K¨unstl. Intell. 27, 119–128 (2013). doi:10.1007/s13218-013-0239-2 I. Horrocks, Ontologies and the semantic web. Commun. ACM 51(12), (2008). ISSN:00010782, doi:10.1145/1409360.1409377 ˘ L.E. Kavraki, P. Svestka, J. Latombe, M.H. Overmars, Probabilistic roadmaps for path planning in high dimensional configuration spaces. IEEE Trans. Robot. Autom. 12(4), 566–580 (1996) M. Kleinehagenbrock, J. Fritsch, G. Sagerer, Supporting advanced interaction capabilities on a mobile robot with a flexible control system, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Sendai, 2004 K. Konolige, Colbert: a language for reactive control in Sapphira, in KI-97: Advances in Artificial Intelligence, Freiburg. Lecture Notes in Computer Science, vol. 1303 (Springer, 1997), pp. 31–52 K. Konolige, K. Myers, E. Ruspini, A. Saffiotti, The Saphira architecture: a design for autonomy. J. Exp. Theor. Artif. Intell. 9(2–3), 215–235 (1997) T. Koo, F. Hoffmann, F. Mann, H. Shim, Hybrid control of an autonomous helicopter, in IFAC Workshop on Motion Control, Grenoble, 1998 R. Koymans, Specifying real-time properties with metric temporal logic. Real-Time Syst. 2(4), 255–299 (1990) H. Kruppa, M. Castrillon-Santana, B. Schiele, Fast and robust face finding via local context, in Joint IEEE International Workshop on Visual Surveillance and Performance Evaluation of Tracking and Surveillance, Nice, France, 2003 J.J. Kuffner, S.M. LaValle, RRT-connect: an efficient approach to single-query path planning, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), San Francisco, 2000, pp. 995–1001 J. Kvarnstr¨om, TALplanner and other extensions to temporal action logic. Ph.D. thesis, Department of Computer and Information Science, Link¨oping University, 2005

37 A Distributed Hybrid Deliberative/Reactive Architecture

951

J. Kvarnstr¨om, Planning for loosely coupled agents using partial order forward-chaining, in Proceedings of the International Conference on Automated Planning and Scheduling (ICAPS), Freiburg (AAAI, 2011), pp. 138–145. ISBN:978-1-57735-503-8, 978-1-57735-504-5 J. Kvarnstr¨om, P. Doherty, TALplanner: a temporal logic based forward chaining planner. Ann. Math. Artif. Intell. 30, 119–169 (2000) D. Land´en, F. Heintz, P. Doherty, Complex task allocation in mixed-initiative delegation: A UAV Case Study, in Principles and Practice of Multi-Agent Systems: 13th International Conference, PRIMA 2010, Kolkata, India, November 12–15, 2010, Revised Selected Papers, ed. by N. Desai, A. Liu, M. Winikoff. Lecture Notes in Computer Science vol. 7057 (Springer, Berlin/Heidelberg, 2010), pp. 288–V303. ISBN: 978-3-642-25919-7, e-978-3-642-25920-3. DOI: 10.1007/978-3-642-25920-3 20 S.M. LaValle, Planning Algorithms (Cambridge University Press, Cambridge, 2004) R. Lienhart, J. Maydt, An extended set of Haar-like features for rapid object detection, in Proceedings of International Conference on Image Processing, Rochester, 2002, pp. 900–903 P. Mantegazza, E.L. Dozio, S. Papacharalambous, RTAI: real time application interface. Linux J. 2000, (72es) (2000) G.H. Mealy, A method for synthesizing sequential circuits. Bell Syst. Tech. J. 34(5), 1045–1079 (1955) T. Merz, Building a system for autonomous aerial robotics research, in Proceedings of the IFAC Symposium on Intelligent Autonomous Vehicles (IAV), Lisbon, 2004 T. Merz, S. Duranti, G. Conte, Autonomous landing of an unmanned aerial helicopter based on vision and inertial sensing, in Proceedings of the International Symposium on Experimental Robotics, Singapore, 2004 T. Merz, P. Rudol, M. Wzorek, Control system framework for autonomous robots based on extended state machines, in International Conference on Autonomic and Autonomous Systems, Silicon Valley, 2006, 0:14. doi:http://doi.ieeecomputersociety.org/10.1109/ICAS.2006.19 M. Montemerlo, S. Thrun, FastSLAM: A Scalable Method for the Simultaneous Localization and Mapping Problem in Robotics. Springer Tracts in Advanced Robotics, vol. 27 (Springer, Berlin, 2007). ISBN:978-3-540-46402-0 E.F. Moore, Gedanken-experiments on sequential machines, in Automata Studies, ed. by C.E. Shannon, J. McCarthy, W.R. Ashby (Princeton University Press, Princeton, 1956), pp. 129–153 A. N¨uchter, H. Surmann, K. Lingermann, J. Hertzberg, S. Thrun, 6D SLAM with an application in autonomous mine mapping, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), New Orleans, 2004, pp. 1998–2003 P.-O. Pettersson, Using randomized algorithms for helicopter path planning. Licentiate thesis, Department of Computer and Information Science, Link¨oping University, 2006 P.-O. Pettersson, P. Doherty, Probabilistic roadmap based path planning for an autonomous unmanned helicopter. J. Intell. Fuzzy Syst. 17(4), 395–405 (2006) M. Quigley, B. Gerkey, K. Conley, J. Faust, T. Foote, J. Leibs, E. Berger, R. Wheeler, A. Ng, ROS: an open-source robot operating system, in ICRA Workshop on Open Source Software, Kobe, 2009 P. Rudol, P. Doherty, Human body detection and geolocalization for UAV search and rescue missions using color and thermal imagery, in Proceedings of the IEEE Aerospace Conference, Big Sky, 2008 S. Sekhavat, P. Svestka, J.-P. Laumond, M.H. Overmars, Multi-level path planning for nonholonomic robots using semi-holonomic subsystems. Int. J. Robot. Res. 17, 840–857 (1996) M.K. Smith, C. Welty, D.L. McGuinness, OWL Web Ontology Language Guide, 2004. See http:// www.w3.org/TR/owl-guide/. P. Viola, M.J.-C. Jones, Rapid object detection using a boosted cascade of simple features, in Proceedings of the Conference on Computer Vision and Pattern Recognition, Kauai, 2001 M. Whalley, G. Schulein, C. Theodore, Design and flight test results for a hemispherical LADAR developed to support unmanned rotorcraft urban operations research, in American Helicopter Society 64th Annual Forum, Montr´eal, 2008

952

P. Doherty et al.

F. White, A model for data fusion, in Proceedings of the National Symposium for Sensor Fusion, Orlando, 1988 M. Wzorek, Selected aspects of navigation and path planning in unmanned aircraft systems. Licentiate thesis 1509, Department of Computer and Information Science, Link¨oping University, 2011 M. Wzorek, G. Conte, P. Rudol, T. Merz, S. Duranti, P. Doherty, From motion planning to control – a navigation framework for an unmanned aerial vehicle, in Proceedings of the Bristol International Conference on UAV Systems, Bristol, 2006a M. Wzorek, D. Landen, P. Doherty, GSM technology as a communication media for an autonomous unmanned aerial vehicle, in Proceedings of the Bristol International Conference on UAV Systems, Bristol, 2006b

Classification of Multi-UAV Architectures

38

Ivan Maza, An´ıbal Ollero, Enrique Casado, and David Scarlatti

Contents 38.1 38.2

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . General Concepts and Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.2.1 Coordination and Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.2.2 Communication and Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.2.3 Classification of Multi-UAV Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.2.4 General Model for Each UAV in the Team . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.3 Physical Coupling: Joint Load Transportation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.4 Vehicle Formations and Coordinated Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.5 Swarms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.6 Intentional Cooperation Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38.7 UAVs Networked with Sensors and Actuators in the Environment . . . . . . . . . . . . . . . . . . . . . 38.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

954 955 955 956 957 959 961 964 966 967 970 972 972

Abstract

This chapter presents a classification of different schemes for the cooperation of multiple UAVs, taking into account the coupling between the vehicles and the type of cooperation. Then, the research and development activities in load

I. Maza () Robotics, Vision and Control Group, University of Seville, Seville, Spain e-mail: [email protected] A. Ollero Robotics, Vision and Control Group, University of Seville, Seville, Spain Center for Advanced Aerospace Technologies (CATEC), Parque Tecnologico y Aeron autico de Andaluca, La Rinconada, Spain e-mail: [email protected] E. Casado • D. Scarlatti Boeing Research & Technology Europe, Madrid, Spain e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 119, © Springer Science+Business Media Dordrecht 2015

953

954

I. Maza et al.

transportation, formation control, swarm approaches, and intentional cooperation architectures are revised. The chapter also considers UAVs networked with other elements in the environment to support their navigation and, in general, their operation. The chapter refers theoretical work but also emphasizes practical field outdoor demonstrations involving aerial vehicles.

38.1

Introduction

This chapter considers the cooperation of multiple UAVs performing jointly missions such as search and rescue, reconnaissance, surveying, detection and monitoring in dangerous scenarios, exploration and mapping, and hazardous material handling. The coordination of a team of autonomous vehicles allows to accomplish missions that no individual autonomous vehicles can accomplish on its own. Team members can exchange sensor information, collaborate to track and identify targets, perform detection and monitoring activities (Ollero and Maza 2007), or even actuate cooperatively in tasks such as the transportation of loads. The advantages of using multiple UAVs when comparing to a single powerful one can be categorized as follows: • Multiple simultaneous interventions. A single autonomous vehicle is limited at any one time to sense or actuate in a single point. However, the components of a team can simultaneously collect information from multiple locations and exploit the information derived from multiple disparate points to build models that can be used to take decisions. Moreover, multiple UAVs can apply simultaneously forces at different locations to perform actions that could be very difficult for a single UAV. • Greater efficiency. The execution time of missions such as exploration and searching for targets can be decreased when using simultaneously multiple vehicles. • Complementarities of team members. Having a team with multiple heterogeneous vehicles offers additional advantages due to the possibility of exploiting their complementarities. Thus, for example, ground and/or aerial vehicles with quite different characteristics and onboard sensors can be integrated in the same platform. For instance, the aerial vehicles could be used to collect information from locations that cannot be reached by the ground vehicles, while these ground members of the team could be equipped with heavy actuators. Then, the aerial and ground vehicles could be specialized in different roles. But even considering the UAVs themselves complementarities can be found: the fixed-wing airplanes typically have longer flight range and time of flight, whereas helicopters have vertical take-off and landing capability, better maneuverability, and therefore can hover to obtain detailed observations of a given target. • Reliability. The multi-UAV approach leads to redundant solutions offering greater fault tolerance and flexibility including reconfigurability in case of failures of individual vehicles. • Technology evolution. The development of small, relatively low-cost UAVs is fuelled by the progress of embedded systems together with the developments on

38 Classification of Multi-UAV Architectures

955

technologies for integration and miniaturization. Furthermore, the progress on communication technologies experienced in the last decade plays an important role in multiple vehicle systems. • Cost. A single vehicle with the performance required to execute some tasks could be an expensive solution when comparing to several low-cost vehicles performing the same task. This is clear for UAVs and particularly in smallsize, light, and low-cost versions, where constraints such as power consumption, weight, and size play an important role. Section 38.2 of this chapter will deal with the general concepts and contains a rough classification of systems with multiple autonomous UAVs. Then, the joint load transportation, formation control, swarm approaches, and teams with intentional cooperation are examined in more detail along Sects. 38.3–38.6. The networking of UAVs with other sensors and actuators in the environment is considered in Sect. 38.7. Finally, Sect. 38.8 concludes the chapter.

38.2

General Concepts and Classification

In the first part of this section, the concepts of coordination and cooperation are briefly presented due to their relevance in any system with multiple autonomous vehicles. Then, the important role of communications in these systems is summarized. Finally, a classification based on the coupling between the vehicles is outlined.

38.2.1 Coordination and Cooperation In platforms involving multiple vehicles, the concepts of coordination and cooperation play an important role. In general, the coordination deals with the sharing of resources, and both temporal and spatial coordination should be considered. The temporal coordination relies on synchronization among the different vehicles, and it is required in a wide spectrum of applications. For instance, for objects monitoring, several synchronized perceptions of the objects could be required. In addition, spatial coordination of UAVs deals with the sharing of the space among them to ensure that each UAV will be able to perform safely and coherently regarding the plans of the other UAVs and the potential dynamic and/or static obstacles. Some formulations are based on the extension of robotics path planning concepts. In this context, the classical planning algorithms for a single robot with multiple bodies (Latombe 1990; LaValle 2006) may be applied without adaptation for centralized planning (assuming that the state information from all the UAVs is available). The main concern, however, is that the dimension of the state space grows linearly in the number of UAVs. Complete algorithms require time that is at least exponential in dimension, which makes them unlikely candidates for such problems. Samplingbased algorithms are more likely to scale well in practice when there are many UAVs, but the resulting dimension might still be too high. For such cases, there

956

I. Maza et al.

are also decoupled path planning approaches such as the prioritized planning that considers one vehicle at a time according to a global priority. Cooperation can be defined as a “joint collaborative behavior that is directed toward some goal in which there is a common interest or reward” (Barnes and Gray 1991). According to Cao et al. (1997), given some task specified by a designer, a multiple-robot system displays cooperative behavior if, due to some underlying mechanism (i.e., the “mechanism of cooperation”), there is an increase in the total utility of the system. The cooperation of heterogeneous vehicles requires the integration of sensing, control, and planning in an appropriated decisional architecture. These architectures can be either centralized or decentralized depending on the assumptions on the knowledge’s scope and accessibility of the individual vehicles, their computational power, and the required scalability. A centralized approach will be relevant if the computational capabilities are compatible with the amount of information to process and the exchange of data meets both the requirements of speed (up-to-date data) and expressivity (quality of information enabling wellinformed decision-taking). On the other hand, a distributed approach will be possible if the available knowledge within each distributed vehicle is sufficient to perform “coherent” decisions, and this required amount of knowledge does not endow the distributed components with the inconveniences of a centralized system (in terms of computation power and communication bandwidth requirements). One way to ensure that a minimal global coherence will be satisfied within the whole system is to enable communication between the vehicles of the system, up to a level that will warranty that the decision is globally coherent. One of the main advantages of the distributed approach relies on its superior suitability to deal with the scalability of the system.

38.2.2 Communication and Networking It should be noticed that communication and networking also play an important role in the implementation of these schemes for multiple unmanned vehicles. Single vehicle communication systems usually have an unshared link between the vehicle and the control station. The natural evolution of this communication technique toward multi-vehicle configurations is the star-shaped network configuration. While this simple approach to vehicles intercommunication may work well with small teams, it could not be practical or cost-effective as the number of vehicles grows. Thus, for example, in multi-UAV systems, there are some approaches of a wireless heterogeneous network with radio nodes mounted at fixed sites, on ground vehicles, and in UAVs. The routing techniques allow any two nodes to communicate either directly or through an arbitrary number of other nodes which act as relays. When autonomous teams of UAVs should operate in remote regions with little/no infrastructure, using a mesh of ground stations to support communication between the mobile nodes is not possible. Then, networks could be formed in an ad hoc fashion, and the information exchanges occur only via the wireless networking equipment carried by the individual UAVs. Some autonomous configurations

38 Classification of Multi-UAV Architectures

957

(such as close formation flying) result in relatively stable topologies. However, in others, rapid fluctuations in the network topology may occur when individual vehicles suddenly veer away from one another or when wireless transmissions are blocked by terrain features, atmospheric conditions, signal jamming, etc. In spite of such dynamically changing conditions, vehicles in an autonomous team should maintain close communications with others in order to avoid collisions and facilitate collaborative team mission execution. In order to reach these goals, two different approaches have been adopted. One, closer to the classical networks architecture, establishes a hierarchical structure and routes data in the classical down-up-down traversing as many levels of the hierarchy as needed to reach destination. The other prospective direction to assist routing in such an environment is to use location information provided by positioning devices, such as global positioning systems (GPS), thus using what it is called location aware protocols. These two techniques are compatible and can be mixed. For example, some of the levels in a hierarchical approach could be implemented using location aware methods.

38.2.3 Classification of Multi-UAV Architectures Multi-UAV systems can be classified from different points of view. One possible classification is based on the coupling between the UAVs (see Fig. 38.1): 1. Physical coupling. In this case, the UAVs are connected by physical links and then their motions are constrained by forces that depend on the motion of other UAVs. The lifting and transportation of loads by several UAVs lies in this category and will be addressed in Sect. 38.3 of this chapter. The main problem is the motion-coordinated control, taking into account the forces constraints. From the point of view of motion planning and collision avoidance, all the members of the team and the load can be considered as a whole. As the number of vehicles is usually low, both centralized and decentralized control architectures can be applied. 2. Formations. The vehicles are not physically coupled, but their relative motions are strongly constrained to keep the formation. Then, the motion planning problem can be also formulated considering the formation as a whole. Regarding the collision avoidance problem within the team, it is possible to embed it in the formation control strategy. Scalability properties to deal with formations of many individuals are relevant, and then, decentralized control architectures are usually preferred. Section 38.4 of the chapter will deal with the formations and will also show how the same techniques can be applied to control coordinated motions of vehicles even if they are not in formation. 3. Swarms. They are homogeneous teams of many vehicles which interactions generate emerging collective behaviors. The resulting motion of the vehicles does not lead necessarily to formations. Scalability is the main issue due to the large number of vehicles involved, and then pure decentralized control architectures are mandatory. Section 38.5 of the chapter will be devoted to the swarm approaches.

958

I. Maza et al.

a

b

c

d

Fig. 38.1 Graphical illustration of a possible classification for multiple UAVs systems: (a) physical coupling (3 UAVs transporting one object), (b) formations, (c) swarms, and (d) team executing tasks represented by crosses following an intentional cooperation approach. The UAVs are represented by gray arrows

4. Intentional cooperation. The UAVs of the team move according to trajectories defined by individual tasks that should be allocated to perform a global mission (Parker 1998). These UAV trajectories typically are not geometrically related as in the case of the formations. This cooperation will be considered in Sect. 38.6 of this chapter. In this case, problems such as multi-UAV task allocation, high-level planning, plan decomposition, and conflict resolution should be solved, taking into account the global mission to be executed and the different UAVs involved. In this case, both centralized and decentralized decisional architectures can be applied. In the rest of sections of this chapter, each type of multi-vehicle system is discussed in further detail. But before to proceed with each one, a general model for each UAV of the team is presented. This model can be particularized to fit any of the types of the above classification, as it will be shown later.

38 Classification of Multi-UAV Architectures

959

38.2.4 General Model for Each UAV in the Team Let us consider a team of UAVs that plan their actions according to a set of coordination and cooperation rules R. In particular, it is assumed that the set R includes k possible tasks  D f1 ; 2 ; : : : ; k g with n logical conditions requiring a change of task in the current plan. Let E D fe1 ; e2 ; : : : ; en g be a set of discrete events associated with n logical conditions requiring a change of task during the execution. Each task has a set of m parameters … D f 1 ;  2 ; : : : ;  m g defining its particular characteristics. Systems composed of a physical plant and a decisional and control engine implementing such kind of cooperation rules R can be modeled as hybrid systems (Fierro et al. 2002; Chaimowicz et al. 2004; Fagiolini et al. 2007; Li et al. 2008). Figure 38.2 shows a simplified hybrid model that summarizes the different interactions that can be found in each member of the classification presented above. The i -th UAV’s current task has a discrete dynamics ı W   E ! , i.e., i C D ı.i ; ei /;

(38.1)

where ei 2 E is an event (internal or external) requiring a change of task from i to i C , both from the set of tasks . It should be noticed that each task can have a different control algorithm or a different set of parameters for the same controller. The control reconfiguration is triggered in the transition between tasks associated to different events. Event activation is generated by ei D ˆ.qi ; "i ; Xi ; N i /;

(38.2)

where "i represents the internal events (such as changes in the execution states of the tasks) and N i is a vector N i D .i1 ; i2 ; : : : ; iNm / containing the messages coming from Nm UAVs cooperating with the i -th UAV. Those messages are used, for example, in the negotiation processes involved in the intentional cooperation mechanisms and are generated onboard each UAV by a decisional module  (see Fig. 38.2). This module encompasses high-level reasoning and planning, synchronization among different UAVs, negotiation protocols for task allocation and conflict resolution purposes, task management and supervision, complex task decomposition, etc. Regarding the perception of the environment, a database ED with “a priori” knowledge about the environment, including static obstacles, objects of interest, and threats can be available and updated with the new information gathered during the mission. On the other hand, object detection and localization (Merino 2007) is usually required in many applications. The state x of the object to be tracked obviously includes its position p, and for moving objects, it is also convenient to add the velocity pP into the kinematic part of the state to be estimated. But further information is needed in general. An important objective in some missions is to

960

I. Maza et al.

Environment Database

z¯i

ED Xi = ° (zi;z¯i;ED)

Onboard Sensors

zi

zi Xi

ei

ei = F(qi;ei;Xi;m-i)

m-i

Δ

q¯i

ui = g(qi;q¯i;ti;Xi)

ui

qi

q˙i = f (qi;ui;gi)

gi

qi cq¯i

mi

qi

ti

g˙i = h(gi;qi;cq¯i)

Fig. 38.2 General blocks and interactions considered in the hybrid model for each UAV described in Sect. 38.2

confirm that an object belongs to a certain class within a set „ (for instance, in the case of fire alarms detection, this set will include as classes of fire alarms and false alarms). Therefore, the state will include information regarding the classification of the object. Also, in certain applications, some appearance information could be needed to characterize an object, which also can help in the task of data association between different UAVs with different cameras. Additionally, this information could even include the 3D volume of the object that can be added to the obstacles database. In general, the appearance information is static and will be represented by . The complete dynamic state to be estimated is composed by the status of all the objects, No , and the number of objects can vary with the time. The state estimated by the i -UAV at time t is then represented by the vector xi D ŒxiT1 ; : : : ; xiTNo T . Each potential object m is defined by xim D Œpim pPim im T . The information about the objects will be inferred from all the measurements zi from the sensors onboard the UAVs and zNi gathered by the fleet of Ns UAVs that can communicate with the i -th UAV fNzj ; j D 1; : : : ; Ns g. The latter vector can be completed with the measurements from sensors located around the environment, such as static

38 Classification of Multi-UAV Architectures

961

surveillance cameras, or nodes from wireless sensor networks (WSNs) deployed in the area of interest. Notice that zi also contains the forces/torques derived from the interaction with the environment that are measured with the sensors onboard. On the other hand, let qi 2 ‚ be a vector describing the state of the i -th UAV taking values in the configuration space ‚, and let i 2  be the task  that the i -th UAV is currently executing. This UAV’s configuration qi has a continuous dynamics qPi D f .qi ; ui ; i /;

(38.3)

where ui 2 U is a control input and i 2 models the dynamics associated to the possible physical interaction with the environment and/or other UAVs Pi D h. i ; qi ; c qN i /;

(38.4)

with vector c qN i D .qi1 ; qi2 ; : : : ; qiNc / containing the configurations of the Nc neighbors physically connected to the i -th UAV. Then i ¤ 0 only if there is physical interaction. Regarding ui , it is a feedback law generated by a low-level controller g W ‚  N    ‡ ! U , i.e., ‚ ui D g.qi ; qNi ; i ; Xi /; (38.5) so that the UAV’s trajectory qi .t/ corresponds to the desired current task i , taking into account the configurations of the N neighbors qN i D .qi1 ; qi2 ; : : : ; qiN / with influence in the control of the i -th UAV. This influence can be found, for example, in the control problem of swarms and formations. On the other hand, Eq. (38.5) also includes the vector Xi 2 ‡, taking values in the environment model space ‡, that encompasses estimations about forces/torques derived from the interaction with the environment, targets to the tracked, obstacles detected during the mission and/or known “a priori,” threats to be avoided, etc. In conclusion, the hybrid dynamics H of the i -th UAV shown in Fig. 38.2 has zNi ; N i ; qNi , and c qN i as inputs and zi ; i , and qi as outputs. This diagram is not intended to be exhaustive or to cover all the possible architectures and existing systems. Instead, it is aimed at providing a general overview of all the possible interactions in order to put into context the approaches presented in the next sections of the chapter.

38.3

Physical Coupling: Joint Load Transportation

The transportation of a single object by multiple autonomous vehicles is a natural extension of the moving by several persons of a large and heavy object that cannot be handled by a single person. The coordinated control of the motion of each vehicle should consider the involved forces induced by the other vehicles and the load itself. Thus, in the scheme depicted in Fig. 38.2, there is a term i ¤ 0 modelling those forces, which is taken into account in the design of the controller in Eq. 38.5.

962

I. Maza et al.

It should be also mentioned that i can be measured using onboard sensors. For instance, in the case of several UAVs transporting a load using ropes, a force sensor in the rope can provide a measurement of the influence of the other UAVs and the load being transported. Each UAV could be controlled around a common compliance center attached to the transported object. Under the assumption that each UAV holds the object firmly with rigid links, the real trajectories of all of the UAVs are equal to the real trajectory of the object. However, in some transportation problems, this assumption cannot be applied, and the transported object moves with a dynamic behavior that can be expressed by means of Eq. 38.4. A suitable approach for the required coordinated control is the leader-follower scheme that will be more detailed in the next section. In this scheme, the desired trajectory is the trajectory of the leader. The followers estimate the motion of the leader by themselves through the motion of the transported object. Several examples of this approach can be found in the robotics community. The leaderfollower scheme extended to multiple followers and to robots with non-holonomic constraints (Kosuge and Sato 1999) has been implemented in an experimental system with three tracked mobile robots with a force sensor. In Sugar and Kumar (2002), the decentralized control of cooperating mobile manipulators is studied with a designated lead robot being responsible for task planning. The control of each robot is decomposed (mechanically decoupled) into the control of the gross trajectory and the control of the grasp. The excessive forces due to robot positioning errors and odometry errors are accommodated by the compliant arms. In Huntsberger et al. (2004), distributed coordinated control of two rovers carrying a 2.5 m long mockup of a photovoltaic tent is presented and demonstrated as an example of the CAMPOUT behavior-based control architecture. Borenstein (2000) details the OmniMate system, which uses a compliant linkage platform between two differential drive mobile robots (Labmate) that provide a loading deck for up to 114 kg of payload. Lifting and transportation of loads by using multiple helicopters has been also a research topic for many years motivated by the payload constraints of these vehicles and the high cost of helicopters with significant payload. In addition, the use of multiple manned helicopters is also problematic and only simple operations, like load transportation with two helicopters, can be performed by extremely skillful and experienced pilots. The level of stress is usually very high, and practical applications are therefore rarely possible. Load transportation and deployment by one and several helicopters is very useful for many applications including the delivery of first-aid packages to isolated victims in disasters (floods, earthquakes, fires, industrial disasters, and many others) and is also a basic technology for other future applications: the building of platforms for evacuation of people in rescue operations and the installation of platforms in uneven terrains for landing of manned and unmanned VTOL aircrafts. This later application would first require the installation of the supporting units defining the horizontal surface and later the installation of the surface itself.

38 Classification of Multi-UAV Architectures

963

Fig. 38.3 Three autonomous helicopters from the Technical University of Berlin (TUB-H model) transporting a wireless camera to the top floor of a building with a height of 12 m in May 2009. A device onboard each helicopter is equipped with a force sensor to estimate the influence of the other helicopters and the load itself – term i in Eq. 38.4. The images show the mission during the actual load transportation (left) and shortly before the load deployment (right). A fourth helicopter which was used to acquire airborne video footage of the mission is visible on the right

The autonomous lifting and transportation by two helicopters (twin lift) has been studied since the beginning of the nineties by means of nonlinear adaptive control (Mittal et al. 1991) and H1 control (Reynolds and Rodriguez 1992). In Lim et al. (1999) an interactive modeling, simulation, animation, and real-time control (MoSART) tool to study the twin lift helicopter system is presented. However, only simulation experiments have been found until December 2007, when lifting and transportation of a load by means of three autonomous helicopters (Bernard and Kondak 2009) was demonstrated experimentally in the framework of the AWARE project. After that first successful test, the load transportation system was used again in 2009 to deploy a camera on the roof of a building with a height of 12 m (see Fig. 38.3) in the framework of the same project (Bernard et al. 2011). Notice that in this case, the physical coupling between UAS are involved through direct interactions of each unmanned aerial vehicle with the joint load. Small-size single or multiple autonomous quadrotors are also considered for load transportation and deployment in Michael et al. (2011), Palunko et al. (2012), and Sreenath et al. (2013). Dynamically coupled quadrotors should cooperate safely to transport load, in contrast to the existing results on formation control of decoupled multi-UAV systems that are addressed in the next section.

964

38.4

I. Maza et al.

Vehicle Formations and Coordinated Control

In the formations, the members of the group of vehicles must keep user-defined distances with the other group members. The control problem consists of maintaining these user-defined distances, and consequently the configurations of the N neighbors qNi D .qi1 ; qi2 ; : : : ; qiN / in the formation should be taken into account in the control law (see Eq. 38.5). Those configurations can be either received via intervehicle communication or estimated using the sensors onboard. Anyway, formation control involves the design of distributed control laws with limited and disrupted communication, uncertainty, and imperfect or partial measurements. Vehicle formation is a basic strategy to perform multi-vehicle missions including searching and surveying, exploration and mapping, active reconfigurable sensing systems, and space-based interferometry. An added advantage of the formation paradigm is that new members can be introduced to expand or upgrade the formation or to replace a failed member. The stability of the formation has been studied by many researchers that have proposed robust controllers to provide insensitivity to possibly large uncertainties in the motion of nearby agents, transmission delays in the feedback path, and the consideration of the effect of quantized information. The close formation flight control of homogeneous teams of fixed-wing UAVs airplanes received attention in the last 10 years. The large group formation of small UAVs also offers benefits in terms of drag reduction and then increased payoffs in the ability to maintain persistent coverage of a large area. Both linear (Giulietti et al. 2000) and nonlinear control laws (Schumacher and Singh 2000) have been proposed and tested in simulation. However, practical implementations are still very scarce. In How et al. (2004) a demonstration of two fixed-wing UAVs simultaneous flying, the same flight plan (tracking waypoints in open-loop formation) is reported. In the same paper, two UAVs were linked to the same receding horizon trajectory planner, and independent timing control was performed about the designed plans. The leader-follower approach mentioned in the previous section has been also used to control general formations where the desired positions of followers are defined relative to the actual state of a leader. It should be noted that every formation can be further divided into simplest leader-follower schemes. Then, in this approach, some vehicles are designated as leaders and track predefined trajectories, while the followers track transformed versions of these trajectories according to given schemes. In the leader-follower approach, path planning only needs to be performed in the leader workspace. The leader-follower pattern is adopted in Yun et al. (2010) to maintain a fixed geometrical formation of unmanned helicopters while navigating following certain trajectories. The leader is commanded to fly on some predefined trajectories, and each follower is controlled to maintain its position in formation using the measurement of its inertial position and the information of the leader position and velocity, obtained through a wireless modem. In Gu et al. (2006) two-aircraft formation flights confirmed the performance of a formation

38 Classification of Multi-UAV Architectures

965

controller designed to have an inner- and outerloop structure, where the outerloop guidance control laws minimized the forward, lateral, and vertical distance error by controlling the engine propulsion and generating the desired pitch and roll angles to be tracked by the innerloop controller. In the formation flight configuration, a radio control pilot maintained ground control of the leader aircraft, while the autonomous follower aircraft maintained a predefined position and orientation with respect to the leader aircraft. The leader-follower approach is applied in Galzi and Shtessel (2006) in the design of robust and continuous controllers to achieve collision-free path-tracking formation in the presence of unknown bounded disturbances acting on each UAV. In Bayraktar et al. (2004) an experiment with two fixed-wing UAVs is presented. The leader UAV was given a predetermined flight plan, and the trajectory of the UAV was updated once per second in real time through the ground station to keep the follower at a fixed distance offset from the leader. Finally, the vehicles platooning can be considered as a particular case consisting of a leader followed by vehicles in a single row. Both lateral and longitudinal control to keep the safe headway, and lateral distance should be considered. The simplest approach relies on individual vehicle control from the data received from the single immediate front vehicle (Bom et al. 2005). Other methods are based on a virtual leader, a moving reference point whose purpose is to direct, herd, and/or manipulate the vehicle group behavior. The lack of a physical leader among the vehicles implies that any vehicle is interchangeable with any other in the formation. A solution based on a virtual leader approach combined with an extended local potential field is presented in Paul et al. (2008) for formation flight and formation reconfiguration of small-scale autonomous helicopters. And for fixed-wing models, experimental results with YF-22 research aircrafts can be found in Campa et al. (2007), validating the performance of a formation control law using also a virtual leader configuration. Practical applications of formation control should include a strategy for obstacle avoidance and reconfiguration of the formation. The avoidance of big obstacles could be performed by changing the trajectory of the whole formation to go around the obstacle or to pass through a narrow tunnel (Desai et al. 2001). If the obstacles are smaller than the size of the formation, the vehicles should be able to compromise the formation until the obstacle is passed. In order to do so, the obstacle avoidance behavior should be integrated in the control strategy of the individual members of the formation to avoid/bypass obstacles. Hybrid control techniques have been applied to avoid obstacles and solve the formation reconfiguration (Zelinski et al. 2003). Formation is not the only cooperative scheme for UAVs in applications such as exploration and mapping. The cooperation of multiple UAVs can be also examined from the point of view of the intentionality to achieve a given mission. Then, according to Parker (1998), it is possible to distinguish between intentional cooperation and swarm-type cooperation. Those approaches are considered in the following two sections:

966

38.5

I. Maza et al.

Swarms

The key concept in the swarms is that complex collective global behaviors can arise from simple interactions between large numbers of relatively unintelligent agents. This swarm cooperation is based on concepts from biology (Sharkey 2006) and typically involves a large number of homogeneous individuals, with relatively simple sensing and actuation, and local communication and control that collectively achieve a goal. This can be considered as a bottom-up cooperation approach. It usually involves numerous repetitions of the same activity over a relatively large area. The agents execute the same program and interact only with other nearby agents by measuring distances and exchanging messages. Thus, according to Fig. 38.2 the configurations of the N neighbors qN i D .qi1 ; qi2 ; : : : ; qiN / should be considered as well as the messages N i D .i1 ; i2 ; : : : ; iNm / coming from Nm UAVs cooperating with the i -th UAV. Nevertheless, it should be mentioned that depending on the particular communication and sensing capabilities of the UAVs in the swarm, simplified mechanisms based on partial or imperfect information could be required. For example, the estimation of the full vector qNi is not possible in many swarm-based systems, and partial information such as the distances with the neighbors is the only measurement available. The same is applicable to the messages interchanged that can range from data packets sent through wireless links to simple visual signals based on lights of different colors. The concept of operations for a micro-UAV system is adopted from nature from the appearance of flocking birds, movement of a school of fish, and swarming bees among others. This “emergent behavior” is the aggregate result of many simple interactions occurring within the flock, school, or swarm. Exploration of this emergent behavior in a swarm is accomplished through a highperformance computing parallel discrete event simulation in Corner and Lamont (2004). In Kube and Zhang (1993) different mechanisms that allow populations of behavior-based robots to perform collectively tasks without centralized control or use of explicit communication are presented. Matari´c (1992) provides the results of implementing group behaviors such as dispersion, aggregation, and flocking on a team of robots. In Kovacina et al. (2002) a rule-based, decentralized control algorithm that relies on constrained randomized behavior and respects UAV restrictions on sensors, computation, and flight envelope is presented and evaluated in a simulation of an air vehicle swarm searching for and mapping a chemical cloud within a patrolled region. Another behaviorbased decentralized control strategy for UAV swarming by using artificial potential functions and sliding-mode control technique is presented in Han et al. (2008). Individual interactions for swarming behavior are modeled using the artificial potential functions. For tracking the reference trajectory of the swarming of UAVs, a swarming center is considered as the object of control. The sliding-mode control technique is adopted to make the proposed swarm control strategy robust with respect to the system uncertainties and varying mission environment.

38 Classification of Multi-UAV Architectures

967

The bio-inspired motivation of swarms can be found, for example, in Zhang et al. (2007), which describes an adaptive task assignment method for a team of fully distributed vehicles with initially identical functionalities in unknown task environments. The authors employ a simple self-reinforcement learning model inspired by the behavior of social insects to differentiate the initially identical vehicles into “specialists” of different task types, resulting in stable and flexible division of labor; on the other hand, in dealing with the cooperation problem of the vehicles engaged in the same type of task, the so-called ant system algorithm was adopted to organize low-level task assignment. Dasgupta (2008) presents a multiagent-based prototype system that uses swarming techniques inspired from insect colonies to perform automatic target recognition using UAVs in a distributed manner within simulated scenarios. In Altshuler et al. (2008) a swarm of UAVs is used for searching one or more evading targets, which are moving in a predefined area while trying to avoid a detection by the swarm (Cooperative Hunters problem). By arranging themselves into efficient geometric flight configurations, the UAVs optimize their integrated sensing capabilities, enabling the search of a maximal territory. In general, the above approaches deal with homogeneous teams without explicit consideration of tasks decomposition and allocation, performance measures, and individual efficiency constraints of the members of the team. Those aspects are considered in the intentional cooperation schemes described in the next section.

38.6

Intentional Cooperation Schemes

In the intentional cooperation approaches, each individual executes a set of tasks (subgoals that are necessary for achieving the overall goal of the system and that can be achieved independently of other subgoals) explicitly allocated to perform a given mission in an optimal manner according to planning strategies (Parker 1998). The UAVs cooperate explicitly and with purpose but also has the limitation of independent subgoals: If the order of task completion is mandatory, additional explicit knowledge has to be provided to state ordering dependencies in the preconditions. It is also possible to follow a design based on “collective” interaction, in which entities are not aware of other entities in the team, yet they do share goals, and their actions are beneficial to their teammates (Parker 2008). Key issues in these systems include determining which UAV should perform each task (task allocation problem) so as to maximize the efficiency of the team and ensuring the proper coordination among team members to allow them to successfully complete their mission. In order to solve the multi-robot task allocation problem, some metrics to assess the relevance of assigning given tasks to particular robots are required. In Gerkey and Matari´c (2004) a domain-independent taxonomy for the multiagent task allocation problem is presented. In the last years, a popular approach to solve this problem in a distributed way is the application of marketbased negotiation rules. An usual implementation of those distributed negotiation rules (Botelho and Alami 1999; Dias and Stenz 2002; Gerkey and Matari´c 2002) is

968

I. Maza et al.

based on the Contract Net Protocol (Smith 1980). In those approaches, the messages N i D .i1 ; i2 ; : : : ; iNm / coming from Nm UAVs cooperating with the i -th UAV are those involved in the negotiation process: announce a task, bid for a task, allocate a task, ask for the negotiation token, etc. Once the tasks have been allocated, it is necessary to coordinate the motions of the vehicles, which can be done by means of suitable multi-vehicle path/velocity planning strategies, as mentioned in Sect. 38.2. The main purpose is to avoid potential conflicts among the different trajectories when sharing the same working space. It should be mentioned that even if the vehicles are explicitly cooperating through messages, a key element in many motion coordination approaches is the updated information about the configurations of the N neighbors qN i D .qi1 ; qi2 ; : : : ; qiN /. Formal approaches to the collision avoidance problem and different approaches that can be applied to solve it can be found in LaValle (2006) and Latombe (1990). On the other hand, teams composed by heterogeneous members involve challenging aspects, even for the intentional cooperation approach. In Ollero and Maza (2007) the current state of the technology, existing problems, and potentialities of platforms with multiple UAVs (with emphasis on systems composed by heterogeneous UAVs) are studied. This heterogeneity is twofold: firstly in the UAV platforms looking to exploit the complementarities of the aerial vehicles, such as helicopters and airships, and secondly in the information-processing capabilities onboard, ranging from pure remotely teleoperated vehicles to fully autonomous aerial robots. The multi-UAV coordination and control architecture developed in the European COMETS project (Gancet et al. 2005) was demonstrated for the autonomous detection and monitoring of fires (Ollero and Maza 2007) by using two helicopters and one airship (see Fig. 38.4). Regarding teams involving aerial and ground vehicles, the CROMAT architecture also implemented cooperative perception and task allocation techniques (Viguria et al. 2010) that have been demonstrated in fire detection, monitoring, and extinguishing. Multiagent (combined ground and air) tasking and cooperative target localization have been also demonstrated recently (Hsieh et al. 2007) as well as multi-target tracking (ground vehicles) with a microUAV (He et al. 2010). In Maza et al. (2011) a distributed architecture for the autonomous coordination and cooperation of multiple UAVs for civil applications is presented. The architecture is endowed with different modules that solve the usual problems that arise during the execution of multipurpose missions, such as task allocation, conflict resolution, and complex task decomposition. One of the main objectives in the design of the architecture was to impose few requirements to the execution capabilities of the autonomous vehicles to be integrated in the platform. Basically, those vehicles should be able to move to a given location and activate their payload when required. Thus, heterogeneous autonomous vehicles from different manufacturers and research groups can be integrated in the architecture developed, making it easily usable in many multi-UAV applications. The software implementation of the architecture was tested in simulation and finally validated in field experiments with four autonomous helicopters. The validation process included several multi-UAV missions for civil applications in a simulated urban setting: surveillance applying

38 Classification of Multi-UAV Architectures

969

Fig. 38.4 Coordinated flights in the COMETS project involving an airship and two autonomous helicopters

the strategies for multi-UAV cooperative searching presented in Maza and Ollero (2007); fire confirmation, monitoring, and extinguishing; load transportation and deployment with single and multiple UAVs; and people tracking. Finally, cooperative perception can be considered as an important tool in many applications based on intentional cooperation schemes. It can be defined as the task of creating and maintaining a consistent view of a world containing dynamic objects by a group of agents each equipped with one or more sensors. Thus, a team of vehicles can simultaneously collect information from multiple locations and exploit the information derived from multiple disparate points to build models that can be used to take decisions. In particular, cooperative perception based on artificial vision has become a relevant topic in the multi-robot domain, mainly in structured environments (Thrun 2001; Schmitt et al. 2002). In Merino et al. (2006) cooperation perception methods for multi-UAV system are proposed. Each UAV extracts knowledge, by applying individual perception techniques, and the overall cooperative perception is performed by merging the individual results. This approach requires knowing the relative position and orientation of the UAVs. In many outdoor applications, it is assumed that the position of all the UAVs can be obtained by means of GPS and broadcasted through the communication system. However, if this is not the case, the UAVs should be capable of identifying and of localizing each other (Konolige et al. 2003) which could be difficult with the onboard sensors. Another approach consists of identifying common objects in the scene. Then, under certain assumptions, the relative pose displacement between the vehicles can be computed from these correspondences. In Merino et al. (2006) this strategy has been demonstrated with heterogeneous UAVs. In the ANSER project (see, e.g., Sukkarieh et al. 2003), decentralized sensor data fusion using multiple aerial vehicles is also researched and experimented with fixed-wing UAVs with navigation and terrain sensors.

970

38.7

I. Maza et al.

UAVs Networked with Sensors and Actuators in the Environment

The development of wireless communication technologies in the last 10 years makes possible the integration of autonomous vehicles with the environment infrastructure. Particularly, the integration with wireless sensor and actuator networks is very promising. The benefit of this integration can be seen from two different points of view: • The use of UAVs to complement the information collected by the wireless sensor network (WSN), to perform as mobile “data mules,” to act as communication relays, to improve the connectivity of the network, and to repair it in case of malfunctioning nodes. • The use of WSNs as an extension of the sensorial capabilities of the UAVs. In this case, the information about the objects in the environment will be inferred from all the measurements zi from the sensors onboard and zNi gathered by the fleet of Ns UAVs and nodes that can communicate with the i -th UAV fNzj ; j D 1; : : : ; Ns g. Static wireless sensor networks have important limitations as far as the required coverage and the short communication range in the nodes are concerned. The use of mobile nodes could provide significant improvements. Thus, they can provide the ability to dynamically adapt the network to environmental events and to improve the network connectivity in case of static nodes failure. Node mobility for ad hoc and sensor networks has been studied by many researchers (Grossglauser and Tse 2002; Venkitasubramaniam et al. 2004). Moreover, mobile nodes with singlehop communication and the ability to recharge batteries, or refueling, have been proposed as data mules of the network, gathering data while they are near of fixed nodes and saving energy in static node communications (Jain et al. 2006). The coordinated motion of a small number of nodes in the network to achieve efficient communication between any pair of other mobile nodes has been also proposed. An important problem is the localization of the nodes of a WSN. This is an open problem because GPS-based solutions in all the nodes are usually not viable due to the cost, the energy consumption, and the satellite visibility from each node. In Caballero et al. (2008) a probabilistic framework for the localization of an entire WSN based on a vehicle is presented. The approach takes advantage of the good localization capabilities of the vehicle and its mobility to compute estimation of the static nodes positions by using the signal strength of the messages interchanged with the network. However, in many scenarios, the motion of the mobile nodes installed on ground vehicles or carried by persons is very constrained, due to the characteristics of the terrain or the dangerous conditions involved, such as in civil security and disaster scenarios. The cooperation of aerial vehicles with the ground wireless sensor network offers many potentialities. The use of aircrafts as data sinks when they fly over the fixed sensor networks following a predictable pattern in order to

38 Classification of Multi-UAV Architectures

971

Fig. 38.5 Sensor deployment from an autonomous helicopter in the AWARE project experiments carried out in 2009

gather data from them has been proposed by several authors in the WSN community. In Corke et al. (2003) an algorithm for path computation and following is proposed and applied to guide the motion of an autonomous helicopter flying very close to the sensor nodes deployed on the ground. It should be noticed that flight endurance and range of the currently available low-cost UAVs is very constrained (Ollero and Merino 2004). Moreover, reliability and fault tolerance is a main issue in the cooperation of the aerial vehicles. Furthermore, these autonomous vehicles need communication infrastructure to cooperate or to be teleoperated by humans in emergency conditions. Usually this infrastructure is not available, or the required communication range is too large for the existing technology. Then, the deployment of this communication infrastructure is a main issue. In the same way, in most wireless sensor networks projects, it is assumed that the wireless sensor network has been previously fully deployed without addressing the problems to be solved when the deployment is difficult. Moreover, in the operation of the network, the infrastructure could be damaged or simply the deployment is not efficient enough. Then, the problem is the repairing of the coverage or the connectivity of the network by adding suitable sensor and communication elements. In Corke et al. (2004) the application of an autonomous helicopter for the deployment and repairing of a wireless sensor network is proposed. This approach has been also followed in the AWARE project (Maza et al. 2010), whose platform has self-deployment and self-configuration features for the operation in sites without sensing and communication infrastructure. The deployment includes not only wireless sensors (see Fig. 38.5) but also heavier loads such as communication equipment that require the transportation by using several helicopters (see Fig. 38.3).

972

38.8

I. Maza et al.

Conclusions

The concepts of coordinated and cooperative control of multiple UAVs deserved significant attention in the last years in the control, robotics, artificial intelligence, and communication communities. The implementation of these concepts involve integrated research in the control, decision, and communication areas. For instance, the communication and networking technologies play an important role in the practical implementation of any multi-vehicle system. Thus, the integrated consideration of communication and control problems is a relevant research and development topic. This chapter has first reviewed the existing work on the transportation of a single load by different autonomous vehicles. In order to solve this problem, control theory based on models of the vehicles and their force interactions have been applied. The chapter also studied formation control. In this problem, the application of control theory based on models of the vehicles is dominant. However, behaviorbased approaches that do not use these models have been also demonstrated. The work on swarms has been also reviewed. Approaches inspired in biology and multiagent systems are common. The problems are typically formulated for large number of individuals, but up to now, the practical demonstrations involve few physical UAVs. The intentional task-oriented cooperation of robotic vehicles, possibly heterogeneous, has been also addressed. The task allocation problem and the path planning techniques play an important role here, as well as the application of cooperative perception methods. Finally, the chapter has explored the integration and networking of one or many UAVs with sensors and actuators in the environment pointing out the benefits of this integration. The self-deployment of the network and the motion planning to maintain quality of service are promising approaches that have been preliminarily studied but still require significant attention.

References Y. Altshuler, V. Yanovsky, I. Wagner, A. Bruckstein, Efficient cooperative search of smart targets using UAV swarms. Robotica 26(4), 551–557 (2008) D. Barnes, J. Gray, Behaviour synthesis for co-operant mobile robot control, in International Conference on Control, Edinburgh, 1991, vol. 2, pp. 1135–1140 S. Bayraktar, G.E. Fainekos, G.J. Pappas, Experimental cooperative control of fixed-wing unmanned aerial vehicles, in Proceedings of the IEEE Conference on Decision and Control, Atlantis, Paradise Island, the Bahamas, 2004 M. Bernard, K. Kondak, Generic slung load transportation system using small size helicopters, in Proceedings of the International Conference on Robotics and Automation, Kobe, Japan (IEEE, 2009), pp. 3258–3264 M. Bernard, K. Kondak, I. Maza, A. Ollero, Autonomous transportation and deployment with aerial robots for search and rescue missions. J. Field Robot. 28(6), 914–931 (2011) J. Bom, B. Thuilot, F. Marmoiton, P. Martinet, Nonlinear control for urban vehicles platooning, relying upon a unique kinematic GPS, in Proceedings of the IEEE International Conference on Robotics and Automation, Barcelona, Spain, 2005, pp. 4138–4143 J. Borenstein, The OmniMate: a guidewire- and beacon-free AGV for highly reconfigurable applications. Int. J. Prod. Res. 38(9), 1993–2010 (2000)

38 Classification of Multi-UAV Architectures

973

S.C. Botelho, R. Alami, M+: a scheme for multi-robot cooperation through negotiated task allocation and achievement, in Proceedings of the IEEE International Conference on Robotics and Automation, Detroit, 1999, vol. 2, pp. 1234–1239 F. Caballero, L. Merino, P. Gil, I. Maza, A. Ollero, A probabilistic framework for entire WSN localization using a mobile robot. Robot. Auton. Syst. 56(10), 798–806 (2008) G. Campa, Y. Gu, B. Seanor, M. Napolitano, L. Pollini, M. Fravolini, Design and flight-testing of non-linear formation control laws. Control Eng. Pract. 15(9), 1077–1092 (2007) Y.U. Cao, A.S. Fukunaga, A. Kahng, Cooperative mobile robotics: antecedents and directions. Auton. Robots 4(1), 7–27 (1997) L. Chaimowicz, V. Kumar, M.F.M. Campos, A paradigm for dynamic coordination of multiple robots. Auton. Robots 17(1), 7–21 (2004) P. Corke, R. Peterson, D. Rus, Networked robots: flying robot navigation using a sensor net, in Proceedings of the International Symposium of Robotic Research, Siena, 2003 P. Corke, S. Hrabar, R. Peterson, D. Rus, S. Saripalli, G. Sukhatme, Autonomous deployment and repair of a sensor network using an unmanned aerial vehicle, in Proceedings of the IEEE International Conference on Robotics and Automation, New Orleans, LA, USA, 2004, pp. 3602–3608 J.J. Corner, G.B. Lamont, Parallel simulation of UAV swarm scenarios, in Proceedings of the 36th Conference on Winter Simulation, WSC’04, Washington, DC, USA, 2004, pp. 355–363 P. Dasgupta, A multiagent swarming system for distributed automatic target recognition using unmanned aerial vehicles. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 38(3), 549–563 (2008) J.P. Desai, J.P. Ostrowski, V. Kumar, Modeling and control of formations of nonholonomic mobile robots. IEEE Trans. Robot. Autom. 17(6), 905–908 (2001) M.B. Dias, A. Stenz, Opportunistic optimization for market-based multirobot control, in Proceedings IEEE/RSJ International Conference on Intelligent Robots and Systems, Lausanne, 2002, pp. 2714–2720 A. Fagiolini, G. Valenti, L. Pallottino, G. Dini, A. Bicchi, Decentralized intrusion detection for secure cooperative multi-agent systems, in Proceedings of the IEEE International Conference on Decision and Control, New Orleans, LA, USA, 2007, pp. 1553–1558 R. Fierro, A. Das, J. Spletzer, J. Esposito, V. Kumar, J.P. Ostrowski, G. Pappas, C.J. Taylor, Y. Hur, R. Alur, I. Lee, G. Grudic, B. Southall, A framework and architecture for multi-robot coordination. Int. J. Robot. Res. 21(10–11), 977–995 (2002) D. Galzi, Y. Shtessel, UAV formations control using high order sliding modes, in Proceedings of the American Control Conference, Minneapolis, MN, USA, 2006, vol. 2006, pp. 4249–4254 J. Gancet, G. Hattenberger, R. Alami, S. Lacroix, Task planning and control for a multi-UAV system: architecture and algorithms, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, Edmonton, Canada, 2005, pp. 1017–1022 B. Gerkey, M. Matari´c, Sold!: auction methods for multi-robot coordination. IEEE Trans. Robot. Autom. 18(5), 758–768 (2002) B. Gerkey, M. Matari´c, A formal analysis and taxonomy of task allocation in multi-robot systems. Int. J. Robot. Res. 23(9), 939–954 (2004) F. Giulietti, L. Pollini, M. Innocenti, Autonomous formation flight. IEEE Control Syst. Mag. 20(6), 34–44 (2000) M. Grossglauser, D.N.C. Tse, Mobility increases the capacity of ad hoc wireless networks. IEEE/ACM Trans. Netw. 10(4), 477–486 (2002) Y. Gu, B. Seanor, G. Campa, M. Napolitano, L. Rowe, S. Gururajan, S. Wan, Design and flight testing evaluation of formation control laws. IEEE Trans. Control Syst. Technol. 14(6), 1105–1112 (2006) K. Han, J. Lee, Y. Kim, Unmanned aerial vehicle swarm control using potential functions and sliding mode control. Proc. Inst. Mech. Eng., Part G: J. Aerosp. Eng. 222(6), 721–730 (2008)

974

I. Maza et al.

R. He, A. Bachrach, N. Roy, Efficient planning under uncertainty for a target-tracking micro-aerial vehicle, in IEEE International Conference on Robotics and Automation, 2010, Kobe, Japan, 2010 J. How, E. King, Y. Kuwata, Flight demonstrations of cooperative control for UAV teams, in Proceedings of the AIAA 3rd Unmanned-Unlimited Technical Conference, Workshop, and Exhibit, Chicago, 2004, vol. 1, pp. 505–513 M.A. Hsieh, L. Chaimowicz, A. Cowley, B. Grocholsky, J.F. Keller, V. Kumar, C.J. Taylor, Y. Endo, R.C. Arkin, B. Jung, D.F. Wolf, G. Sukhatme, D.C. MacKenzie, Adaptive teams of autonomous aerial and ground robots for situational awareness. J. Field Robot. 24(11), 991–1014 (2007) T.L. Huntsberger, A. Trebi-Ollennu, H. Aghazarian, P.S. Schenker, P. Pirjanian, H.D. Nayar, Distributed control of multi-robot systems engaged in tightly coupled tasks. Auton. Robots 17(1), 79–92 (2004) S. Jain, R.C. Shah, W. Brunette, G. Borriello, S. Roy, Exploiting mobility for energy efficient data collection in wireless sensor networks. Mobile Netw. Appl. 11(3), 327–339 (2006) K. Konolige, D. Fox, B. Limketkai, J. Ko, B. Stewart, Map merging for distributed robot navigation, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, Las Vegas, 2003, pp. 212–217 K. Kosuge, M. Sato, Transportation of a single object by multiple decentralized-controlled nonholonomic mobile robots, in Proceedings of the IEEE International Conference on Intelligent Robots and Systems, Kyongju, Korea, 1999, vol. 3, pp. 1681–1686 M. Kovacina, D. Palmer, G. Yang, R. Vaidyanathan, Multi-agent control algorithms for chemical cloud detection and mapping using unmanned air vehicles, in Proceedings of the IEEE International Conference on Intelligent Robots and Systems, Lausanne, Switzerland, 2002, vol. 3, pp. 2782–2788 C.R. Kube, H. Zhang, Collective robotics: from social insects to robots. Adapt. Behav. 2(2), 189–218 (1993) J.C. Latombe, Robot Motion Planning (Kluwer, 1990) J.C. Latombe, Robot Motion Planning (Kluwer Academic Publishers, Boston, MA, 1991) S.M. LaValle, Planning Algorithms (Cambridge University Press, Cambridge, 2006) Available at http://planning.cs.uiuc.edu/ H. Li, F. Karray, O. Basir, I. Song, A framework for coordinated control of multiagent systems and its applications. IEEE Trans. Syst. Man Cybern., Part A: Syst. Hum. 38(3), 534–548 (2008) C. Lim, R.P. Metzger, A. Rodriguez, Interactive modeling, simulation, animation and real-time control (MoSART) twin lift helicopter system environment, in Proceedings of the American Control Conference, San Diego, CA, USA, 1999, vol. 4, pp. 2747–2751 M.J. Matari´c, Designing emergent behaviors: from local interactions to collective intelligence, in From Animals to Animats 2, 2nd International Conference on Simulation of Adaptive Behavior (SAB-92), ed. by J.-A. Meyer, H. Roitblat, S. Wilson (MIT, Cambridge, Honolulu, Hawaii, USA, 1992), pp. 432–441 I. Maza, A. Ollero, Multiple UAV cooperative searching operation using polygon area decomposition and efficient coverage algorithms, in Distributed Autonomous Robotic Systems 6, ed. by R. Alami, R. Chatila, H. Asama. Volume 6 of Distributed Autonomous Robotic Systems (Springer, Tokyo/New York, 2007), pp. 221–230 I. Maza, K. Kondak, M. Bernard, A. Ollero, Multi-UAV cooperation and control for load transportation and deployment. J. Intell. Robot. Syst. 57(1–4), 417–449 (2010) I. Maza, F. Caballero, J. Capitan, J.M. de Dios, A. Ollero, A distributed architecture for a robotic platform with aerial sensor transportation and self-deployment capabilities. J. Field Robot. 28(3), 303–328 (2011) L. Merino, Cooperative perception techniques for multiple unmanned aerial vehicles: applications to the cooperative detection, localization and monitoring of forest fires. PhD thesis, Dpto. Ingenieria de Sistemas y Automatica – University of Seville

38 Classification of Multi-UAV Architectures

975

L. Merino, F. Caballero, J.M. de Dios, J. Ferruz, A. Ollero, A cooperative perception system for multiple UAVs: application to automatic detection of forest fires. J. Field Robot. 23(3–4), 165–184 (2006) N. Michael, J. Fink, V. Kumar, Cooperative manipulation and transportation with aerial robots. Auton. Robots 30(1), 73–86 (2011) M. Mittal, J.V.R. Prasad, D.P. Schrage, Nonlinear adaptive control of a twin lift helicopter system. IEEE Control Syst. Mag. 11(3), 39–45 (1991) A. Ollero, I. Maza (eds.), Multiple Heterogeneous Unmanned Aerial Vehicles. Springer Tracts on Advanced Robotics (Springer, Berlin/New York, 2007) A. Ollero, L. Merino, Control and perception techniques for aerial robotics. Annu. Rev. Control 28(2), 167–178 (2004) I. Palunko, P. Cruz, R. Fierro, Agile load transportation: safe and efficient load manipulation with aerial robots. IEEE Robot. Autom. Mag. 19(3), 69–79 (2012) L. Parker, ALLIANCE: an architecture for fault-tolerant multi-robot cooperation. IEEE Trans. Robot. Autom. 14(2), 220–240 (1998) L. Parker, Distributed intelligence: overview of the field and its application in multi-robot systems. J. Phys. Agents 2(1), 5–14 (2008) T. Paul, T. Krogstad, J. Gravdahl, Modelling of UAV formation flight using 3D potential field. Simul. Model. Pract. Theory 16(9), 1453–1462 (2008) H.K. Reynolds, A.A. Rodriguez, H1 control of a twin lift helicopter system, in Proceedings of the 31st IEEE Conference on Decision and Control, Tucson, AZ, USA, 1992, pp. 2442–2447 T. Schmitt, R. Hanek, M. Beetz, S. Buck, B. Radig, Cooperative probabilistic state estimation for vision-based autonomous mobile robots. IEEE Trans. Robot. Autom. 18, 670–684 (2002) C. Schumacher, S. Singh, Nonlinear control of multiple UAVs in close-coupled formation flight, in Proceedings of the AIAA Guidance, Navigation, and Control Conference, Denver, CO, USA, 2000, pp. 14–17 A.J.C. Sharkey, Robots, insects and swarm intelligence. Artif. Intell. Rev. 26(4), 255–268 (2006) G. Smith, The Contract Net Protocol: high-level communication and control in a distributed problem solver. IEEE Trans. Comput. 29(12), 1104–1113 (1980) K. Sreenath, N. Michael, V. Kumar, Trajectory generation and control of a quadrotor with a cablesuspended load – a differentially-flat hybrid system, in Proceedings of the IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, 2013 T.G. Sugar, V. Kumar, Control of cooperating mobile manipulators. IEEE Trans. Robot. Autom. 18(1), 94–103 (2002) S. Sukkarieh, E. Nettleton, J.-H. Kim, M. Ridley, A. Goktogan, H. Durrant-Whyte, The ANSER project: data fusion across multiple uninhabited air vehicles. Int. J. Robot. Res. 22(7–8), 505–539 (2003) S. Thrun, A probabilistic online mapping algorithm for teams of mobile robots. Int. J. Robot. Res. 20(5), 335–363 (2001) P. Venkitasubramaniam, S. Adireddy, L. Tong, Sensor networks with mobile agents: optimal random access and coding. IEEE J. Sel. Areas Commun. (Special issue on Sens. Netw.) 22(6), 1058–1068 (2004) A. Viguria, I. Maza, A. Ollero, Distributed service-based cooperation in aerial/ground robot teams applied to fire detection and extinguishing missions. Adv. Robot. 24(1–2), 1–23 (2010) B. Yun, B. Chen, K. Lum, T. Lee, Design and implementation of a leader-follower cooperative control system for unmanned helicopters. J. Control Theory Appl. 8(1), 61–68 (2010) S. Zelinski, T.J. Koo, S. Sastry, Hybrid system design for formations of autonomous vehicles, in Proceedings of the IEEE Conference on Decision and Control, Maui, HI, USA, 2003, vol. 1, pp. 1–6 D. Zhang, G. Xie, J. Yu, L. Wang, Adaptive task assignment for multiple mobile robots via swarm intelligence approach. Robot. Auton. Syst. 55(7), 572–588 (2007)

Operator Interaction with Centralized Versus Decentralized UAV Architectures

39

M. L. Cummings

Contents 39.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39.2 Operator Interaction in Centralized UAV Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39.3 Operator Capacity in Centralized UAV Architectures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39.4 Operator Interaction in Decentralized UAV Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

978 978 980 986 990 991

Abstract

There has been significant recent research activity attempting to streamline Unmanned Aerial Vehicle (UAV) operations and reduce staffing in order to invert the current many-to-one ratio of operators to vehicles. Centralized multiple UAV architectures have been proposed where a single operator interacts with and oversees every UAV in the network. However, a centralized network requires significant operator cognitive resources. Decentralized multiple UAV networks are another, more complex possible architecture where an operator interacts with an automated mission and payload manager, which coordinates a set of tasks for a group of highly autonomous vehicles. While a single operator can maintain effective control of a relatively small network of centralized UAVs, decentralized architectures are more scalable, particularly in terms of operator workload, and more robust to single points of failure. However, in terms of operator workload, the ultimate success of either a centralized or decentralized UAV architecture is not how many vehicles are in the network per se but rather how many tasks the group of vehicles generates for the operator and how much autonomy is onboard these vehicles. Task-based control of UAV architectures with higher degrees

M.L. Cummings Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, MA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 117, © Springer Science+Business Media Dordrecht 2015

977

978

M.L. Cummings

of autonomy (i.e., decentralized networks) can mitigate cognitive overload and reduce workload. Mutually exclusive boundaries for humans and computers in multiple UAV systems should not be the goal of designers for either centralized or decentralized architectures, but rather more effort needs to be spent in defining mutually supportive roles such that humans and computers complement one another.

39.1

Introduction

The use of unmanned aerial vehicles (UAVs), often referred to as drones, has recently revolutionized military operations worldwide and holds similar promise for commercial settings. The U.S. Air Force now has more UAVs than manned aircraft, small UAVs are now used worldwide by various first response and police units, and they can be found fighting forest fires, monitoring wildlife and possible poachers, in cargo missions and even in entertainment. UAVs require human guidance to varying degrees and often through several operators. Most military and government UAVs require a crew of two to be fully operational. Conventional stick-and-rudder skills have been replaced by point-andclick control so that traditional pilots are no longer needed to control such systems. Onboard automation currently determines the most efficient control response, which is true in many commercial aircraft. While one operator supervises the actual flight activity of the UAV (the “pilot”), the other operator typically monitors the UAV’s sensors, such as a camera, and coordinates with the “pilot” so that he or she can maneuver the UAV for the best system response. There has been significant recent research activity attempting to streamline UAV operations and reduce staffing in order to invert the current many-to-one ratio of operators to vehicles. This is important not just for military operations but also for future commercial operations where air traffic controllers will direct both manned and unmanned aircraft. This chapter will discuss the implications of this staffing inversion for multiple UAV control, particularly in terms of two UAV control architectures, centralized and decentralized. It should be noted that these two architectures are not mutually exclusive and that there really exists a continuum of architectures in between these two bookends. Moreover, while there are many aspects of control architectures that are critical to consider, this chapter focuses on the human implications of such control architectures.

39.2

Operator Interaction in Centralized UAV Architectures

The shift from stick-and-rudder to point-and-click control in UAVs represents a shift in the role of humans from the need for highly rehearsed skill sets to more knowledge-based reasoning inputs. For UAVs and for fly-by-wire military and commercial aircraft, pilots are less in direct manual control of systems but more involved in the higher levels of planning and decision making, particularly for

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

Actuators

Controls Human Operator (Supervisor)

979

Task

Computer Displays

Sensors

Fig. 39.1 Human supervisory control (Sheridan and Verplank 1978)

Fig. 39.2 Hierarchical control loops for a single UAV

remote operations. This shift in control from lower-level skill-based behaviors to higher-level knowledge-based behaviors is known as Human Supervisory Control (HSC). HSC is the process by which a human operator intermittently interacts with a computer, receiving feedback from and providing commands to a controlled process or task environment, which is connected to that computer (Sheridan and Verplank 1978) (Fig. 39.1). In a centralized UAV control architecture, human supervisory control in UAV operation is hierarchical, as represented in Fig. 39.2. The innermost loop of Fig. 39.2 represents the basic guidance and motion control loop, which is the most critical loop that must obey physical laws of nature such as aerodynamic constraints for UAVs. In this loop, the autopilot optimizes local control (keeping the aircraft in stable flight), and while UAV pilots could theoretically take control in this loop, with inherent time latencies that can cause pilot-induced instabilities, this loop is generally left to the automation. The second loop, the navigation loop, represents the actions that some agent, whether human or computer, must execute to meet mission constraints such as routes to waypoints, time on targets, and avoidance of threat areas and no-fly zones. In most current systems, humans enter GPS coordinates as waypoints, and then the system automatically flies to these waypoints. Only now are more advanced automated path planners that generate entire missions instead of serial waypoints starting to appear in operationally deployed UAV systems. The outermost loop of Fig. 39.2 represents the highest levels of control, that of mission and payload management. In this loop, sensors must be monitored and decisions made based on the incoming information to meet overall mission requirements.

980

M.L. Cummings

In the mission management loop, human operators provide the greatest benefit since their decisions require knowledge-based reasoning that includes judgment, experience, and abstract reasoning that in general cannot be performed by automation. Finally, the system health and status monitoring loop on the top of Fig. 39.2 represents the continual supervision that must occur, either by a human or automation or both, to ensure all systems are operating within normal limits. This control loop is a highly intermittent loop in terms of the human, i.e., if the human is engaged in another task, with the highest priority given to the innermost loop, health and status monitoring becomes a distant, secondary task. From the human-in-the-loop perspective, if the inner loops fail, then the higher (outer) loops will also fail. The dependency of higher loop control on the successful control of the lower loops drives human limitations in control of a single and, especially so, for multiple UAVs. If humans must interact in the guidance and motion control loop (e.g., manually fly a UAV), the cost is high because this effort requires significant cognitive resources. What little spare mental capacity is available must be divided between the navigation and mission management control loops. Violations of the priority scheme represented in Fig. 39.2 have led to numerous crashes (Williams 2004). When operators become cognitively saturated or do not correctly allocate their cognitive resources to the appropriate control loops in the correct priorities, they violate the control loops constraints, potentially causing catastrophic failure.

39.3

Operator Capacity in Centralized UAV Architectures

In centralized UAV systems supervised by a human, the primary consideration for system design is how many vehicles a single controller can effectively supervise. Since this supervisor has to interact with each vehicle individually, just how many vehicles can be effectively and safely controlled will primarily be driven by the amount of autonomy onboard the aircraft, which is subsumed across the four loops as shown in Fig. 39.2. By increasing UAV autonomy, operator workload will theoretically be reduced as it could reduce the number of tasks for the operator, and it should reduce the level of interaction even at the highest levels of control in Fig. 39.2. For example, those UAVs that are flown in an autopilot mode relieve the operator from the manual flying tasks that require significant cognitive resources. This frees the operator to perform other critical tasks like mission planning and imagery analysis. While there have been many studies that have attempted to experimentally derive the number of UAVs a single operator can control in a given setting (e.g., Ruff et al. 2002; Dixon et al. 2003; Cummings and Guerlain 2007), model-based approaches are generally more useful in determining not only an upper bound, but also provide insight into how much autonomy will be needed if a certain number of UAVs in a system is desired.

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

Interaction Time

Neglect Time

Wait Time

Interaction Time

981

Neglect Time B

A

Performance Threshold C

Fig. 39.3 The relationship between interaction, neglect, and wait times (Cummings and Mitchell 2008)

One such approach is the fan-out approach (Olsen and Wood 2004), which predicts the upper bound of the number of vehicles a single operator can control given the amount of autonomy in a vehicle as represented through its neglect time (NT) and the required human-computer interaction called interaction time (IT). The fan-out model was later modified to account for wait times that would inevitably be experienced by the system due to human inefficiencies. This adjustment provides a more realistic and lowered upper bound (Cummings and Mitchell 2008). For example, one UAV can operate in a period of NT, and during that period of NT, the operator can attend to other vehicles. However, if an operator fails to notice that a UAV needs assistance (such as needing a new goal once a waypoint has been achieved), or becomes engrossed in a mission plan for a UAV with a system malfunction, one or more UAVs can wait for the operator’s attention, causing a delay for one or more vehicles. Figure 39.3 represents how NT, IT, and wait times (WT) interrelate. Point A represents a discrete event that occurs after a period of neglect time, which causes the vehicle to require immediate operator assistance such as an engine loss. NTs may not be so clearly observable, as exemplified by Point B, which represents performance degradation causing vehicle performance to drop below the NT performance threshold, e.g., a slow degradation of an inertial navigation system. In both NT cases in centralized UAV architectures, once performance has dropped below an acceptable level requiring human interaction, the UAV must wait until the operator recognizes and solves the problem and so that the UAV can move to another NT state. Point C illustrates the system time delay if the problem is not addressed at the appropriate time. Equation 39.1 represents the fan-out mathematical relationship where NT and IT are as defined above. However, it should be noted that IT should account for not just the time an operator inputs commands to a UAV, but also it should include delays where an operator has entered a command and waits for a system’s response.

982

M.L. Cummings

Wait times that add delays to the necessary ITs are accounted for in a separate term as defined by Eq. 39.2. FO D

WT D

NT C1 IT C WT X X i D1

WTQi C

Y X

(39.1)

WTSAj

(39.2)

j D1

WT: Wait time WTQ: Queuing wait time WTSA: Wait time caused by a loss of situation awareness X = Number of human-automation interaction queues that build Y = Number of time periods in which a loss of SA causes wait time In Eq. 39.2, WTQ, or wait time due to queue, results when multiple UAVs require attention, but the operator can only serially attend to them, effectively causing a queue to form for the operator’s attention. For example, if an operator is controlling two UAVs on a search mission and both require the operator to insert waypoints near-simultaneously, the second UAV may have to loiter in place while the operator attends to the first. Assuming the operator can switch attention quickly once the first UAV is back to NT, the time the second UAV waits in the queue (WTQ) is effectively the IT for the first vehicle. WTSA, or wait time due to a loss of situation awareness, is perhaps the most difficult wait time component to model because it represents how effectively an operator can manage his or her attention. Situation awareness (SA) is generally defined as having three levels, which are (1) the perception of the elements in the environment, (2) the comprehension of the current situation, and (3) the projection of future status (Endsley 1995). While SA can decrease under high workload due to competition for attentional resources (Andre and Wickens 1995), it can also decrease under low workload due to boredom and complacency (Rodgers et al. 2000). If an operator does not realize a UAV needs attention (and thus experiences a loss of SA), the time from the initial onset of the need for IT to actual operator recognition of the problem could range from seconds to minutes. Thus, Eq. 39.2 categorizes system wait times as the summation of wait times that result from queues due to near-simultaneous arrival of events that require human intervention plus the wait times due to the operator loss of SA. Wait times increase overall IT and reduce the number of vehicles a single operator can supervise. In terms of operator capacity, Eq. 39.2 demonstrates that as a UAV’s degree of autonomy increases (expressed as an increase in NT), holding all other parameters equal means that a single operator could control more vehicles. Consequently, for a fixed NT (or degree of autonomy), operator capacity could be increased by making the ground control station interactions more streamlined (i.e., lower IT) or ensuring all the correct alarms are in place so that operators do not miss critical points of interventions (i.e., lower WTSA).

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures Table 39.1 Levels of autonomy

LOA I II III IV

V

VI

Table 39.2 Multiple UAV Study Comparison

1 2 3 4 5 6 7 8 9

983

Automation description The computer offers no assistance: human must take all decision and actions The computer offers a complete set of decision/action alternatives The computer offers a selection of decisions/actions The computer suggests a plan and executes that suggestion if the human approves (management by consent) The computer suggests a plan and allows the human a restricted time to veto before automatic execution (management by exception) The human is not involved in the decision-making process; the computer decides and executes autonomously

Experiment Dixon et al. (2005) (baseline) Dixon et al. (2005) (autopilot) Dixon et al. (2005) (auto-alert) Ruff et al. (2002) Dunlap (2006) Cummings et al. (2008) Lewis et al. (2006) Cummings and Guerlain (2007) Hilburn et al. (1997) (ATC)

LOA I I IV IV IV III–IV IV–V IV N/A

Max UV# 1 2 2 4 4 5 8 12 11

A meta-analysis of several previous studies looking at various levels or degrees of automation/autonomy in centralized single operator control of multiple UAVs, particularly at the mission management level (the outermost loop in Fig. 39.2), demonstrates that as NT increases (meaning the degree of autonomy increases), the maximum number of unmanned vehicles a single operator can control increases (Cummings and Mitchell 2008) (Tables 39.1 and 39.2 and Fig. 39.4). The levels of autonomy (LOAs) in Table 39.1 are loosely modeled on Sheridan’s levels of automation framework (Sheridan and Verplank 1978). While there have been numerous other levels of automation and autonomy proposed for UAVs, as recently highlighted by a 2012 Department of Defense report (Defense Science Board 2012), such frameworks pose many problems. These LOAs used here only illustrate increasing degrees of autonomy and thus NT and are not meant to be normative. As shown in Fig. 39.4, research has previously demonstrated that with very low levels of mission management automation, a single operator can supervise at best only two UAVs. However, given high neglect times enabled by higher degrees of autonomy (i.e., UAVs plan their own routes and obstacle avoidance, only seeking high-level mission plan approval), experimentally operators have

984

M.L. Cummings

14

Max Number of UVs

12

8 TLAM

10 7

8

WASM

6 6 UAV UAV 4 5 UCAV

4 2

2 UAV 1 UAV

3 UAV

0 I

II

III

IV

V

VI

Level of Automation

Fig. 39.4 Meta-analysis of previous experiments demonstrating increased operator capacity for increasing neglect time (as expressed by levels of automation)

successfully controlled up to 12 Tomahawk Land Attack Missiles (TLAM), which are highly automated missiles that navigate on their own. These vehicles are effectively one-way UAVs, controlled in much the same way as UAVs through GPS commands. WASMs, or wide area search munitions, are also similar weapons that are launched from another aircraft and then fly themselves until they find their assigned target. Curiously the number of WASMs and TLAMs, which are highly automated and operate in centralized control systems, a single operator can simultaneously control (8–12) is very similar to the number of airplanes a single air traffic controller can handle (11, Hilburn et al. 1997). Arguably, manned aircraft have similar degrees of autonomy since the pilot onboard is expected to obey all the high-level commands of the controller. One of the limitations common across the studies in Table 39.2 is the lack of measurable system-level performance metrics. In general for the studies in Table 39.2, the performance of the operators was deemed acceptable as a function of expert observation, which is a valid method for performance assessment (Endsley and Garland 2000) but is not generalizable across domains and only useful as a descriptive and not predictive metric. Thus, a system-level performance metric should capture both aspects of human and automation performance, which indicates an objective level of goodness and/or satisficing (Simon et al. 1986) (i.e., a “good enough” solution as opposed to optimal). Such system-level metrics are often referenced as key performance parameters (KPPs) (Joint Chiefs of Staff 2007). Towards this end of developing more comprehensive KPPs for multiple UAV systems, a recent study demonstrated that the number of UAVs that a single operator can control in a centralized architecture is not just a function of the level of decision

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

2

985

4 x 10

1.8 1.6

Mission primarily constrained by human limitations

1.4

Cost

1.2 1 0.8 Mission primarily constrained by operational demands

0.6 0.4 0.2 0 1

2

3

4

5

6

7

8

9

10

Number of UAVs

Fig. 39.5 Operator capacity as a function of mission constraints

support automation but is inextricably tied to both mission complexity and overall system performance (Cummings et al. 2007). Using human experimentation in a multiple UAV simulation test bed and a simulated annealing (SA) technique for heuristic-based optimization, operator performance was predicted to be significantly degraded beyond approximately five UAVs with approximately levels 3–4 of autonomy as defined in Table 39.1. The optimal range was predicted to be between 2 and 4 vehicles (Fig. 39.5). Interestingly, in a different single operator-multiple UAV study with an entirely different test bed but similar levels of autonomy and centralized architecture, the optimal number was experimentally determined to be 4 UAVs (Cummings and Mitchell 2008). The KPP in Fig. 39.5 is cost, which takes into account not just operational costs such as fuel but also the cost of missed targets and cost in terms of mission delays introduced by inefficient human interactions. The solid curve in Fig. 39.5 represents a theoretically perfect human operator, and the dotted line represents more realistic human performance that accounts for delays due to inefficient decision making, communication problems, cognitive load, etc. Thus, the performance of the system (the automation and the operator) can vary both as a function of the operator but also can vary due to the operational constraints such as number of targets and operational costs. This variation is why it is important to explicitly link system performance to operator capacity.

986

39.4

M.L. Cummings

Operator Interaction in Decentralized UAV Architectures

While Fig. 39.2 demonstrates supervisory control at the single vehicle level, which for centralized multiple UAV control is simply replicated for each vehicle under control, Fig. 39.6 represents a notional system architecture that will be required for single operator control of multiple decentralized UAVs. In order to achieve this futuristic system, operators will need to interact with an overall automated mission and payload manager, which coordinates a set of tasks for a group of vehicles, instead of individually tasking each vehicle. This effectively represents a decentralized architecture, where operators convey high-level goals to an automated mission manager (such as requesting that an area be searched), which then allows the UAVs to coordinate across the group to determine how to assign particular tasks, which may be dynamic. In a decentralized architecture, navigation and motion control tasks are necessarily subsumed by automation. The decentralized architecture provides a substantial benefit in that the operator and his or her ground control station do not become a single point of failure, i.e., if the operator has intermittent or loss of communications with the vehicles, the system can still function. For example, because the network of vehicles communicates with one another, if one vehicle breaks down, another can take its place. Another advantage is that the system is robust to lapses in operator situation awareness and delays since vehicles do not necessarily have to wait for commands. However, emergent UAV behavior in such systems can be complex and confusing for an operator, and if the system operates in a suboptimal fashion, it could be difficult for operators to correct problems unless they have the ability to understand and then execute the necessary commands to correct the system. For operators supervising a decentralized system, the fan-out approach as depicted in Eq. 39.1 cannot be used to estimate operator capacity, since for centralized systems, the assumption is that the vehicles have their own independent NTs and ITs, which drives the overall number of vehicles that can be controlled. Since operators only provide high-level goals at the mission and payload management level, they do not have an IT for each vehicle but rather an IT for high-level interaction with the team. Similarly for NT, there is not distinct per vehicle NT, since they work together. In control of a decentralized UAV network, the question of operator capacity is driven by how many tasks an operator can handle instead of how many vehicles. Under task-based, decentralized control, a human operator provides high-level control by approving which tasks should be completed by the team of vehicles without directly tasking a particular vehicle. Then the decentralized network of vehicles chooses how to allocate the approved tasks among themselves and can make tactical-level changes on their own, such as switching tasks. In controlling a network of collaborative, decentralized UAVs, the operator could control, for example, 2, 20, 200, or even 2,000 UAVs, as long as the tasks generated by the group of UAVs were manageable by a single operator. Determining the task load manageable by a single operator can roughly be thought of as the number of

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

987

Fig. 39.6 Decentralized control for multiple UAVs

tasks that can be successfully accomplished over the course of the mission. While this number will, of course, vary widely across different missions and with different mixes of vehicles and people, there is one proxy metric that can be used across any decentralized UAV system for workload comparison, which is the concept of utilization. Utilization refers to the “percent busy time” of an operator, i.e., given a time period, the percentage of time that person is busy. In supervisory control settings, this is generally meant as the time an operator is directed by external events to complete a task (e.g., replanning the path of a UAV because of an emergent target) or attending to internal tasks like responding to text messages. What is not included in this measurement is the time spent monitoring the system, i.e., just watching the displays and/or waiting for something to happen. The concept of utilization as a mental workload measure has been used in numerous studies examining supervisory controller performance (Schmidt 1978; Rouse 1983; Cummings and Guerlain 2007; Donmez et al. 2010). These studies generally support that when tasked beyond 70 % utilization, operators’ performances decline. In terms of the previously discussed IT and NT terms, utilization can generally be thought of as IT/(NT C IT) on the aggregate task level. It can be used to describe an operator’s response to task load in centralized UAV architectures as well as decentralized, but it is especially useful for decentralized system analysis given the operator’s interaction at the meta-level instead of the individual vehicle level, which reflects the architecture of Fig. 39.6. In order to determine whether decentralized systems provide any utility in terms of reduced workload and improved performance as compared to centralized UAV architectures, a set of studies was compared that span increasing degrees of

988

M.L. Cummings

autonomy and increasing task loads, summarized in Table 39.1. In the first experiment at the lowest degree of autonomy (Nehme et al. 2008), operators controlled 2–8 mostly centralized unmanned ground vehicles (UGVs), which resulted in four different experiment levels. For this experiment set, the individual vehicles relied on the human for goal setting but had some ability to share local navigation information with one another. In the second experiment with three experimental levels (Nehme 2009), a single operator controlled various mixes of five multiple unmanned aerial and underwater vehicles, with slightly more autonomy in the sense that vehicles would not only path plan themselves, but if the operator did not assign an ultimate goal within a prespecified time, the vehicles would assign themselves to the nearest target. However, the operator could override any individual vehicle and redirect not only its path but its ultimate goal as well. The last experiment with four increasing task load levels represented the highest degree of decentralization, in that operators could only specify a task list to a group of five unmanned aerial and unmanned surface ships. The vehicles negotiated among themselves through a consensus-based bundled algorithm which vehicle would be assigned to which task, and each vehicle determined its own route. The operator could only insert and reprioritize tasks but never direct an individual vehicle. Two different studies were included that used this test bed which focused on medium (Cummings et al. 2010) and high (Clare and Cummings 2011) levels of task loading. In order to directly compare these different studies, the average tasks per minute were determined, which shows how many tasks in each experiment the operator was expected to complete over an average 1-min time interval. Performance scoring was aggregated into low, medium, and high categories since the performance metrics used in each of the sets of experiments could not be directly compared. Low means that for the specific study, that condition resulted in the worse performance and, respectively, for the high performance ranking. Lastly, the average utilization, which is the percentage of time the operator was busy performing tasks required by the system, was also listed. Figure 39.7 illustrates that for each of the three sets of experiments, utilization increased with an increasing task load, which is expected and also an internal validity check. In addition, Fig. 39.7 also illustrates that as the degree of decentralization increases, (i.e., more autonomy across a network of vehicles and less direct control by a human operator), utilization decreases. Interestingly all three experiments had an experimental level of 8 tasks/min and the most decentralized architecture allowed the average operator to work less by 11 % as compared to the more centralized architecture (and 3 % less than the somewhat centralized architecture (Nehme 2009)). And in no case did the centralized architectures produce lower utilizations for similar task loads. In terms of performance, each of the observed data points in Table 39.3 are color coded in Fig. 39.7 to reflect the relative performance score, with red indicating the worst performance; yellow demonstrates moderate performance, and green represents the best performance. Recall that each of these scores is a relative ranking so caution is advised in interpretation. In general the lower utilizations

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

989

0.9 Study [25] 0.8

Study [26] Study [27,28]

Utilization

0.7 0.6 0.5 0.4 0.3 1

3

5 Tasks/Minute

7

9

Fig. 39.7 Utilization for the average tasks/minute for the experiments in Table 39.3 Table 39.3 Task load, workload, and performance for three multiple UAV architectures with increasing autonomy Experiment #1: Unmanned ground vehicles (Nehme et al. 2008) #2: Unmanned aerial and underwater vehicles (Nehme 2009)

#3–4: Unmanned ground and aerial vehicles, medium (Cummings et al. 2010) and high (Clare and Cummings 2011) workload

Architecture Mostly centralized, with some navigation sharing Somewhat centralized with some autonomous path planning and goal selection, 5 vehicles Decentralized, 5 vehicles

Primary task load 2 vehicles 4 vehicles 6 vehicles 8 vehicles

Average tasks/min 1.70 4.02 5.34 8.09

Performance Low Medium Medium Medium

Average utilization 0.59 0.68 0.70 0.78

Same-type vehicles 2 different types All different vehicles

4.90

Medium

0.52

5.59 8.00

High Low

0.64 0.70

23.3 tasks 30 tasks 43.3 tasks 50 tasks

3.96 4.05 7.30 7.90

High High Medium Medium

0.44 0.48 0.65 0.67

produced the best performances, with the caveat that little work has been done in terms of the possible negative impact of low task load on performance (in one exception, see Cummings et al. 2013). For all studies, performance suffered when

990

M.L. Cummings

the 70 % utilization threshold was exceeded, with only the somewhat centralized study participants exhibited markedly worse performance. For the somewhat decentralized (Nehme 2009) and more decentralized experiments (Cummings et al. 2010; Clare and Cummings 2011), there appears to be a nonlinear relationship in the 6 tasks/min and below regions, which is not evident in the centralized experiment (Nehme et al. 2008). More work is needed to determine why such nonlinear relationships exist, the nature of the critical point for the sharp rise, and how a system could be better designed to reduce the sharp increases in workload.

39.5

Conclusion

Taken together, the results from both the centralized and decentralized sets of experiments demonstrate that decentralized, task-based control of UAV architectures with higher degrees of autonomy can mitigate cognitive overload and reduce workload. While a single operator can maintain effective control of a relatively small network of centralized UAVs, decentralized architectures are more scalable, since adding additional agents also adds computational capability (assuming the tasks generated by the system do not linearly increase). Moreover, the decentralized UAV framework is robust to a single point of failure, since no single agent is globally planning for the fleet. In terms of workload for a supervising operator, the ultimate success of either a centralized or decentralized UAV architecture is not how many vehicles are in the network per se but rather how many tasks the group of vehicles generates for the operator and how much autonomy is onboard these vehicles so that neglect time can be increased. And while this increasing autonomy in a decentralized network of UAVs can mean reduced workload for the operator, it also adds significant more complexity to the system, which not only means increased developmental costs than a centralized network but also one that is much harder to certify as safe. Another caveat to the use of increased autonomy to mitigate workload across a UAV network is that such increased autonomy and increased neglect times can exacerbate a loss of operator situation awareness, as well as promote complacency and skill degradation (Parasuraman et al. 2000). Management-by-exception architectures, which occur when automation takes action based on some set of predetermined criteria and only gives operators a chance to veto the automation’s decision, have been shown to improve operator performance (Cummings and Mitchell 2006). However, in such control schemes, operators are also more likely to exhibit automation bias, a decision bias that occurs when operators become overreliant on the automation and do not check to ensure automated recommendations are correct (Mosier and Skitka 1996). Automation bias is a significant concern for command and control systems, so it will be critical to ensure that when higher levels of automation are used, especially at the management-by-exception level, this effect is minimized.

39 Operator Interaction with Centralized Versus Decentralized UAV Architectures

991

Lastly, while it is critical to consider the mental workload of a supervisor of multiple UAVs, the ability of the human to add value to the performance of a team of UAVs cannot be overlooked. In one study that examined whether human supervisors added value in a search and track task for a decentralized, highly autonomous network of UAVs, results in a controlled study show that a 30–50 % increase in overall system performance, particularly in the search task, could be achieved by letting humans coach the automation (Cummings et al. 2012). Thus, instead of attempting to dictate mutually exclusive boundaries for human and computers in multiple UAV systems, either centralized or decentralized, more effort needs to be spent in trying to define mutually supportive roles such that humans and computers complement one another.

References A. Andre, C. Wickens, When users want what’s not best for them. Ergon. Des. 3, 10–14 (1995) A.S. Clare, M.L. Cummings, Task-based interfaces for decentralized multiple unmanned vehicle control, in AUVSI Unmanned Systems North America, Washington, D.C., 2011 M.L. Cummings, S. Guerlain, Developing operator capacity estimates for supervisory control of autonomous vehicles. Hum. Factors 49(1), 1–15 (2007) M.L. Cummings, P.J. Mitchell, Automated scheduling decision support for supervisory control of multiple UAVs. AIAA J. Aerosp. Comput. Inf. Commun. 3(6), 294–308 (2006) M.L. Cummings, P.J. Mitchell, Predicting controller capacity in supervisory control of multiple UAVs. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 38(2), 451–460 (2008) M.L. Cummings, C.E. Nehme, J. Crandall, Predicting operator capacity for supervisory control of multiple UAVs, in Innovations in Intelligent Machines, vol. 70, ed. by J.S. Chahl, L.C. Jain, A. Mizutani, M. Sato-Ilic (Springer, Berlin/New York, 2007) M.L. Cummings, A. Clare, C. Hart, The role of human-automation consensus in multiple unmanned vehicle scheduling. Hum. Factors 52(1), 17–27 (2010) M.L. Cummings, J. How, A. Whitten, O. Toupet, The impact of human-automation collaboration in decentralized multiple unmanned vehicle control. Proc. IEEE 100(3), 660–671 (2012) M.L. Cummings, C. Mastracchio, K.M. Thornburg, A. Mkrtchyan, Boredom and distraction in multiple unmanned vehicle supervisory control. Interact. Comput. 25(1), 34–47 (2013) Defense Science Board, The role of autonomy in DoD systems. Department of Defense, 2012 S.R. Dixon, C.D. Wickens, D. Chang, Comparing quantitative model predictions to experimental data in multiple-UAV flight control, in Human Factors and Ergonomics Society 47th Annual Meeting, Denver, 2003 S. Dixon, C. Wickens, D. Chang, Mission control of multiple unmanned aerial vehicles: a workload analysis. Hum. Factors 47, 479–487 (2005) B. Donmez, C. Nehme, M.L. Cummings, Modeling workload impact in multiple unmanned vehicle supervisory control. IEEE Syst. Man Cybern. Part A Syst. Hum. 99, 1–11 (2010) R.D. Dunlap, The evolution of a distributed command and control architecture for semiautonomous air vehicle operations, in Moving Autonomy Forward Conference, Grantham (Muretex, 2006) M.R. Endsley, Toward a theory of situation awareness in dynamic systems. Hum. Factors 37(1), 32–64 (1995) M.R. Endsley, D.J. Garland, Situation Awareness Analysis and Measurement (Lawrence Erlbaum, Mahwah, 2000) B. Hilburn, P.G. Jorna, E.A. Byrne, R. Parasuraman, The effect of adaptive air traffic control (ATC) decision aiding on controller mental workload, in Human-Automation Interaction:

992

M.L. Cummings

Research and Practice, ed. by M. Mouloua, J.M. Koonce (Lawrence Erlbaum, Mahwah, 1997), pp. 84–91 Joint Chiefs of Staff, Chairman of the Joint Chiefs of Staff instruction 6212.01D. DoD, 2007 M. Lewis, J. Polvichai, K. Sycara, P. Scerri, Scaling-up human control for large UAV teams, in Human Factors of Remotely Operated Vehicles, ed. by N. Cooke, H. Pringle, H. Pedersen, O. Connor (Elsevier, New York, 2006), pp. 237–250 K.L. Mosier, L.J. Skitka, Human decision makers and automated decision aids: made for each other? in Automation and Human Performance: Theory and Applications, Human Factors in Transportation, ed. by R. Parasuraman, M. Mouloua (Lawrence Erlbaum, Mahwah, 1996), pp. 201–220 C.E. Nehme, Modeling human supervisory control in heterogeneous unmanned vehicle systems. Doctor of philosophy, Massachusetts Institute of Technology, 2009 C.E. Nehme, J. Crandall, M.L. Cummings, Using discrete-event simulation to model situational awareness of unmanned-vehicle operators, in 2008 Capstone Conference, Norfolk, 2008 D.R. Olsen, S.B. Wood, Fan-out: measuring human control of multiple robots, in SIGCHI conference on Human factors in Computing Systems, Vienna, 2004 R. Parasuraman, T.B. Sheridan, C.D. Wickens, A model for types and levels of human interaction with automation. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 30(3), 286–297 (2000) M.D. Rodgers, R.H. Mogford, B. Strauch, Post hoc assessment of situation awareness in air traffic control incidents and major aircraft accidents, in Situation Awareness Analysis and Measurement, ed. by M. Endsley, D.J. Garland (Lawrence Erlbaum, Mahwah, 2000), pp. 73–112 W.B. Rouse, Systems Engineering Models of Human-Machine Interaction (North Holland, New York, 1983) H. Ruff, S. Narayanan, M.H. Draper, Human interaction with levels of automation and decision-aid fidelity in the supervisory control of multiple simulated unmanned air vehicles. Presence 11(4), 335–351 (2002) D.K. Schmidt, A queuing analysis of the air traffic controller’s workload. IEEE Trans. Syst. Man Cybern. 8(6), 492–498 (1978) T.B. Sheridan, W. Verplank, Human and computer control of undersea teleoperators. Man-Machine Systems Laboratory, Department of Mechanical Engineering, MIT, Cambridge, 1978 H.A. Simon, R. Hogarth, C.R. Piott, H. Raiffa, K.A. Schelling, R. Thaier, A. Tversky, S. Winter, Decision making and problem solving, in Research Briefings 1986: Report of the Research Briefing Panel on Decision Making and Problem Solving (National Academy Press, Washington D.C., 1986) K.W. Williams, A summary of unmanned aircraft accident/incident data: human factors implications. Federal Aviation Administration, Civil Aerospace Medical Institute, Oklahoma City, 2004

Section IX UAV Health Management Issues Kai Goebel and Michael J. Roemer

UAV Health Management Issues: Introduction

40

Kimon P. Valavanis and George J. Vachtsevanos

Reliability, availability, maintainability, and safety of UAVs, and more generally of other unmanned systems, are a critical concern for the OEM and user communities. The development and application of emerging Prognostics and Health Management and Condition-Based Maintenance (CBM) technologies to all UAV classes is recognized as a key enabler for their effective utility in multiple application domains. UAV Health Management addresses fundamental technologies and their application to UAVs that relate to monitoring and sensing strategies for health management, data acquisition and processing/analysis onboard and off-board the vehicle, fault diagnosis and failure prognosis algorithms, as well as fault-tolerant control routines that enhance the UAV’s useful life and mitigate potential catastrophic events. CBM methods and tools are also covered that take advantage of the prevailing paradigm shift from scheduled or breakdown maintenance towards maintenance practices that are based on the current condition of these assets. Integrated Vehicle Health and Fault Contingency Management for UAVs by Roemer and Tang presents various concepts for integrating real-time vehicle health assessment and fault contingency management technologies for UAVs. The presented integrated vehicle health management (IVHM) and automated contingency management (ACM) system architecture are shown to support real-time, onboard health state assessment and fault management so that UAVs can enjoy greater autonomy and survivability during anomalous operating conditions. Selected

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 140, © Springer Science+Business Media Dordrecht 2015

995

996

K.P. Valavanis and G.J. Vachtsevanos

real-time system identification and automated health assessment algorithms are presented that can readily identify the dynamics and performance limitations of degraded UAV systems. Additionally, a high-level mission adaptation approach is presented to estimate the safe flight operating envelope after the occurrence of faults. Reconfigurable fault-tolerant control techniques that directly utilize the identified UAV subsystem dynamic models have been developed and tested in simulation. Finally, proof-of-concept demonstrations are presented using NASA engine and aircraft dynamic models with simulated engine and actuator faults. Automated Failure Effect Analysis for PHM of UAV by Snooke describes how model-based simulation can be employed to automatically generate the systemlevel effects for comprehensive sets of component failures on systems within the aircraft. The results of the simulation can be used in several ways. They can be used to produce a system-level failure modes and effects analysis (FMEA) for aircraft systems. They can be used to identify the sensors necessary to discriminate remotely between different failures on the aircraft. Once a set of sensors have been chosen for placement on the vehicle, the simulation results can also be used to generate diagnostic and prognostic software for deployment on the vehicle. Using automated FMEA safety analysis software is more efficient than doing the same work without software and also provides a guaranteed level of performance. Using the results of this analysis can provide sensor selection and diagnostic capability while retaining some of the benefits of rule-based diagnostic systems. Alternative model-based techniques have been widely used to create diagnostic systems in a variety of domains, and these approaches are compared with the diagnostic capability provided by a failure effect-oriented technique from the perspective of the UAV application. Prognostics Applied to Electric Propulsion UAV by Goebel and Saha explores the technical underpinnings of how to perform prognostics and shows an implementation on the propulsion of an electric UAV. An accurate run-time battery life prediction algorithm is of critical importance to ensure the safe operation of the vehicle if one wants to maximize in-air time. Current reliability-based techniques turn out to be insufficient to manage the use of such batteries where loads vary frequently in uncertain environments. A particle filter is shown as the method of choice in performing state assessment and predicting future degradation. The method is then applied to the batteries that provide power to the propeller motors. Actuator Fault Detection in UAVs by Ducard is dedicated to actuator fault detection systems for UAVs, with two main requirements: real-time capability and modularity. After defining the terminology employed in this field, it first reviews some commonly used techniques in FDI systems, followed by presenting briefly the mathematical model of a UAV that serves as a basis for the design of two actuator FDI systems. The first method presents and illustrates the multiple-model approach, whereas the second method presents an FDI system, which is based on a single model. Both methods are enhanced by a mechanism that actively tests actuators in order to efficiently detect and isolate actuator faults and failures. Advantages and drawbacks of each method are stated issues of robustness and are discussed against model uncertainties and external perturbation. In addition, aspects of computational

40 UAV Health Management Issues: Introduction

997

load are addressed. The FDI systems applied to a realistic model of an unmanned aircraft and the performance of the methods are shown in simulation. Experimental Validation of Fault Detection and Diagnosis for Unmanned Aerial Vehicles by Chamseddine, Hadi Amoozgar, and Zhang investigates the problems of fault detection, diagnosis, and fault-tolerant control for UAVs. It first presents a detailed overview of existing experimental research focusing on fixedwing as well as rotorcraft UAVs including single-rotor and multi-rotor helicopters. It then discusses three Kalman filters employed for actuator fault detection and diagnosis, namely, the unscented Kalman filter, the two-stage Kalman filter, and the adaptive two-stage Kalman filter. The three filters are experimentally applied to a quadrotor helicopter UAV test bed, and results are shown, compared, and discussed. Fault Detection and Diagnosis for NASA GTM UAV with Dual Unscented Kalman Filter by Zhang investigates a simultaneous state and parameter estimationbased fault detection and diagnosis (FDD) scheme applied to a realistic nonlinear six degree-of-freedom fixed-wing unmanned aerial vehicle (UAV) model, the NASA GTM (generic transport model). By introducing partial loss faults in actuators into the NASA GTM, a Dual Unscented Kalman Filter (DUKF) algorithm is applied for the purpose of online estimation of both flight states and fault parameters. A Bayesian rule is then used for detection and isolation decision-making. The developed FDD scheme is implemented in both the nonlinear GTM and the linear parameter-varying (LPV) representation of the nonlinear GTM. Simulation results show satisfactory results for detecting and diagnosing the control effectors faults. Fault Diagnosis of Skew-Configured Inertial Sensor System for Unmanned Aerial Vehicles by Yoon, Kim, Bae, Kim, and Kim presents fault detection and isolation scheme to handle three successive faults in the skew-configured inertial sensors of an unmanned aerial vehicle. The skew-configured inertial sensors are composed of the primary inertial measurement unit and the redundant secondary inertial measurement unit. Since small unmanned aerial vehicles are restricted by cost and payload space, the secondary small and low-cost inertial measurement unit is installed with a skewed angle in addition to the primary inertial measurement unit. In the hybrid fault detection and isolation scheme, a parity space method and an in-lane monitoring method are combined to increase system tolerance to the occurrence of multiple successive faults during flight. The first and second faults are detected and isolated by the parity space method. The third fault is detected by the parity space method and isolated by the in-lane monitoring method based on the discrete wavelet transform. Hardware-in-the loop tests and flight experiments with a fixed-wing unmanned aerial vehicle are performed to verify the performance of the proposed fault diagnosis scheme.

Integrated Vehicle Health and Fault Contingency Management for UAVs

41

Michael J. Roemer and Liang Tang

Contents 41.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1000 41.2 Integrated IVHM and ACM Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1001 41.3 Real-Time System Health Identification and System Performance Estimation . . . . . . . . . . 1002 41.3.1 Real-Time Strong Tracking System ID Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1003 41.3.2 Self-Tuning Kalman Filter (STKF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004 41.3.3 Engine Health Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005 41.3.4 Flight Control Actuator Health Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1007 41.3.5 Dynamic Flight Envelope Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1010 41.4 ACM System Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012 41.4.1 Model Predictive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014 41.5 Flight Envelope Assessment and Fault-Tolerant Control Simulation Results . . . . . . . . . . . . 1015 41.5.1 Dynamic Flight Envelope Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1015 41.5.2 Detection of Control Effectiveness Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1016 41.5.3 Model Predictive Control for Stuck Elevator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1018 41.5.4 Prognostics-Enhanced ACM System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1019 41.6 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1023 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024

Abstract

This chapter presents various concepts for integrating real-time vehicle health assessment and fault contingency management technologies for unmanned air vehicles. The presented integrated vehicle health management (IVHM) and automated contingency management (ACM) system architecture is shown to support real-time, onboard health state assessment and fault management so that UAVs can enjoy greater autonomy and survivability during anomalous operating conditions. Selected real-time system identification and automated

M.J. Roemer () • L. Tang Impact Technologies, LLC, Rochester, NY, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 46, © Springer Science+Business Media Dordrecht 2015

999

1000

M.J. Roemer and L. Tang

health assessment algorithms are presented that can readily identify the dynamics and performance limitations of degraded UAV systems. Additionally, a highlevel mission adaptation approach is presented to estimate the safe flight operating envelope after the occurrence of faults. Reconfigurable fault-tolerant control techniques that directly utilize the identified UAV subsystem dynamic models have been developed and tested in simulation. Finally, proof-of-concept demonstrations are presented using NASA engine and aircraft dynamic models with simulated engine and actuator faults. Simulation results and remarks on future work are also presented.

41.1

Introduction

At the forefront of technology in today’s military, unmanned systems are highly desired by combatant commanders (COCOMs) for their versatility and persistence. Through their expanding roles in areas such as surveillance; signals intelligence (SIGINT); precision target designation; mine detection; and chemical, biological, radiological, and nuclear (CBRN) reconnaissance, unmanned systems have made key contributions to the Global War on Terror. Unmanned system technology has gone from a prolonged period of limited acceptance and utilization spanning a couple of decades starting with the early drones, to proving and expanding their value in Desert Storm and Kosovo, to a multifaceted, highly desired resource with the recent conflicts in Iraq and Afghanistan. As of October 2008, coalition unmanned aircraft systems (UAS) (exclusive of hand-launched systems) have flown almost 500,000 flight hours in support of Operations Enduring Freedom and Iraqi Freedom; unmanned ground vehicles (UGVs) have conducted over 30,000 missions, detecting and/or neutralizing over 15,000 improvised explosive devices (IEDs); and unmanned maritime systems (UMSs) have provided security to ports [Office of the Secretary of Defense Unmanned Systems Roadmap (2009–2034)]. UAVs and their supporting systems and personnel are seeing a dramatic increase in the operational tempo they support compared to what was originally planned or designed. In addition, modern unmanned technology is still very young, measuring the maturity of any particular platform in years and hundreds of thousands of flight hours for the most mature and heavily utilized, rather than decades and tens of millions of flight hours for many manned platforms. Add to this the fact that many platforms in high demand entered low-rate production while still under advanced concept technology demonstration (ACTD) programs and had not been fully vetted though the traditional process of going to full production. With the military under continued strain to do “more with less” (people, money, etc.), one begins to understand why early data from a 2003 OSD UAV reliability study reports that the proportions of human error versus mechanical, system, and environmentally induced mishaps are nearly reversed between UAVs and the aggregate of manned aircraft, i.e., human error is the primary cause of roughly 85 % of manned mishaps, but only 17 % of unmanned ones. The report also states the cumulative mishap rate (i.e., class A accidents per 100,000 h of flight) of 32 for predator, 334 for

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1001

pioneer, and 55 for hunter (16 since the major reliability improvements in 1996). In comparison to manned aviation mishap rates, general aviation aircraft suffer about 1 mishap per 100,000 h, regional/commuter airliners about a 10th that rate, and larger airliners about a 100th that rate [Unmanned Aerial Vehicle Reliability Study, Feb, 2003, Office of the Secretary of Defense]. Clearly, there is a significant opportunity for both integrated vehicle health management (IVHM) and automated contingency management (ACM) technologies to be deployed, which will help UAVs reach the reliability and safety statistics attained by manned systems. The combination of these features will also help provide the opportunity to reconfigure the appropriate UAV control systems in the event of failures with the aim of increasing the survivability of the aircraft. Enhanced supervisory control methodologies and associated algorithms have been developed for optimizing the operations and performance of such systems due to changing environmental conditions and system degradation (Vachtsevanos et al. 2005, 2006; Bodson 1995; Byington et al. 2004; Volponi and Wood 2005; Litt et al. 2004; Tang et al. 2005). However, in order to provide a truly comprehensive level of integrated fault isolation and accommodation for potentially hundreds of fault scenarios that could be encountered in flight requires a dynamic, modelbased fault contingency management concept that can make decisions “on the fly,” under fault conditions that were never expected or accounted for previously. Hence, a dynamically reconfigurable control architecture that can support on-line fault contingency management based on faults being detected in real time is the key enabling issue to be addressed. This chapter presents the development of a hierarchical architecture and enabling techniques for a real-time aircraft fault contingency management system with the following features: 1. A hierarchical architecture that integrates real-time system health identification, control reconfiguration, and high-level reasoning. 2. The dynamics and performance limitation of the degraded/damaged system are estimated on-line in real time. 3. Adaptive reconfigurable control will be utilized to stabilize and recover the system using the system dynamic model identified on-line in real time. NASA’s Modular Aero-Propulsion System Simulation (MAPSS) engine model and Generic Transport Model (GTM) have been utilized in proof-of-concept simulation studies, and preliminary results are presented.

41.2

Integrated IVHM and ACM Overview

As shown in Fig. 41.1, the core functionality of the integrated vehicle health management (IVHM) and automated contingency management (ACM) system is to use the real-time health information from detected/isolated faults to evaluate a system’s capability for performing its future operations. The ACM system will perform the contingency analysis including the assessment of controller capabilities and core operational capabilities based on reduced system functionality.

1002

M.J. Roemer and L. Tang Real-time System Health Assessment Vehicle Performance Estimation/Reasoning

Automated Contingency Management (ACM) Performance limit

Structure/Actuator Health Condition Aircraft PHM/ System ldentification

Trajectory System dynamic models

Engine Fault /Degradation Severity Engine PHM

Mission Management

Adaptive Flight Control Set point

Health parameter

Vehicle States

Thrust

Propulsion Control

estimates

Fig. 41.1 A conceptual system overview

The analysis results will form the basis for the new constraints of the system’s operation. If the current mission operation plan is incapable of being satisfied, then operation replanning and optimization becomes necessary under the new constraints in order to respond both intelligently and efficiently to the changes of system’s health status or environment. Within this architecture, UAV subsystem faults are detected by prognosis and health management (PHM) modules in real time. At the core of the real-time system health assessment modules is a real-time on-line system identification technique that can rapidly track time-varying parameters. The fault accommodation functionalities are provided in three levels. At the lowest (subsystem) level, fault-tolerant control strategies are utilized to accommodate the fault in the subsystem. A propulsion example is utilized in Fig. 41.1. However, it can also be a fault-tolerant controller for an EMA flight actuator at the subsystem level. At the flight control level, an adaptive control architecture is utilized to compute a time-varying model of the plant dynamics. Using the identified dynamics, the controller calculates the commands required to achieve the desired system responses as specified by a set of flying-qualities models. At the mission level, results from the system health assessment modules are utilized to estimate achievable system performance. The adaptive mission planner can then adjust/optimize mission commands based on the estimated performance of the degraded system.

41.3

Real-Time System Health Identification and System Performance Estimation

In order to adapt to the changing operational environment, fault conditions, or performance degradation, it is very important that the IVHM system possesses the

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1003

capability to identify/update various models in real time so that the appropriate control can be applied. This chapter will present two system/health identification techniques that have been developed and evaluated in simulation studies using NASA engine and aircraft models. These include (1) a real-time strong tracking system identification (SID) algorithm for adaptive flight control and (2) a self-tuning Kalman filter algorithm for engine health assessment (Tang et al. 2007a). Based on the health assessment and newly identified UAV dynamics, the operational flight envelop in real time can be assessed to ensure safe flight regimes are utilized under compromised conditions.

41.3.1 Real-Time Strong Tracking System ID Algorithm The discrete state-space model or equations of motion for a UAV can often be stated (in discrete format) as y.n/ D .n/T .n/ C v.n/

(41.1)

where y.n/ is a measurable system state or output, .n/ is a vector representing the parameters to be tracked/estimated, .n/ is the regressor vector, v.n/ represents the modeling error, and n is a time step index. For example, if the parameters (C and D) of a linear output equation (y D C x C Du) are being tracked with this algorithm, then .n/ would be a vector considering of a row of the C matrix and a row of the D matrix and .n/ would be a vector of the states and inputs, i.e., (x T I uT )T . The same simple linear formulation can be applied to estimate aerodynamic coefficients of a nonlinear equation of motion when it can be linearized or the nonlinear terms can be wrapped into the regressor vector. The least square algorithm drives .n/ toward its true value,   , by a modified least squares algorithm that seeks to minimize an augmented cost function (Bodson 1995): J..n// D

1 2

n X

nk jjy.k/  .n/T '.k/jj2 C ˛jj.n/  .n  1/jj2 (41.2)

kDnN C1

where  is the time-varying forgetting factor, which will be modified adaptively according to a fault detection module. ˛ is the weighting coefficient that adjusts the influence of the derivation of the current estimate from the previous estimate. The augmented cost function includes one additional term that restricts the movement of the estimate, .n/, in the temporal dimension. Similarly, another term can be added to restrict the movement of .n/ in the spatial dimension, O 2 , where O is an a priori estimation of .n/. This formulation e.g., ˇjj.n/  jj will further help when .n/ does not deviate too much from O and result in a more

1004

M.J. Roemer and L. Tang

robust estimate in the presence of noisy measurements. However, at the same time, extra computational burden is introduced. As an improvement for fast detection and tracking in real time, a change detection scheme and a time-varying forgetting factor can be integrated with the above algorithm. When an abrupt change in prediction residuals is detected, a smaller forgetting factor is utilized to discount “old” data and place more weight on the latest measurements. This mechanism helps to improve system response when an abrupt fault occurs.

41.3.2 Self-Tuning Kalman Filter (STKF) In order to make the onboard subsystem models adaptive to potential performance variations due to performance degradation or anomalous conditions, an STKF can be designed to implement the capability to adjust its performance through the tuning of health parameters which are introduced to the system dynamics as “virtual states.” This technique has been applied to aircraft engine fault diagnosis in recent years (Volponi and Wood 2005). For example, the health parameters for a typical gas turbine engine can be defined as Xc D .Fan wFan ; LPC ; wLPC ; HPC ; wHPC ; LPT ; wLPT ; HPT ; wHPT /

(41.3)

where Xc is the augmented virtual states including the efficiency and flow capacity of fan, low- and high-pressure compressors, and low- and high-pressure turbines. The tuning parameters are embedded in the Kalman filter design. If sensor outputs deviate from nominal condition values due to component degradation and/or faults, the Kalman filter will attribute the cause of sensor output deviations to the tuning parameters, so that the residuals of state/output estimates will remain small. At the same time, the values of the tuning/health parameters will be a good estimate of engine component health condition as long as the underlining engine model is reasonably accurate. The accuracy of the underlining subsystem model significantly affects the performance of the fault detection, isolation, and estimation (FDIE) algorithm presented in the previous section. Discrepancies between model outputs and sensor measurements caused by modeling mismatch can be mistakenly considered as a fault effect leading to a false alarm. Therefore, a technical challenge for this type of model-based approach is building an accurate, real-time subsystem model that is robust to unmodeled system dynamics, engine-to-engine variation, and nominal deterioration over time. To solve this problem, empirical modeling is often introduced to capture the modeling mismatch using neural networks (NNs) and improve the accuracy of the physics-based subsystem model. The resulting hybrid model consisting of both a physics-based baseline model and a neural networkbased empirical model is often utilized in real-time engine health management applications as illustrated in Fig. 41.2.

41 Integrated Vehicle Health and Fault Contingency Management for UAVs Fig. 41.2 Hybrid engine modeling for online real-time engine health management

1005

FDIE u

Diagnosis

y

Online

NNs-Enhanced Engine Model Engine

Data Compression / Collection

Update model

Offline

Data Hidden

Input

NNs Training (onboard or off-board)

Output

Trained NNS

(every number of flights)

HPC Degradation Effects on Parameters

Changes from Nominal Value (%)

7

Fig. 41.3 Effects of high pressure compressor degradation on engine parameters

6 LPT Outlet Temp

5 4

HPT Outlet Temp

3

Fuel Flow

2 1

N2

0 −1 −2

Static Burner Pressure 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

HPC Efficiency & Flow Capacity Loss (%)

41.3.3 Engine Health Management Let us further examine the previous engine model tuning example in terms of modeling the potential degradation and how it can be used within the IVHM system architecture. Engine performance monitoring exploits the availability of monitored gas path parameters, including spool speeds, temperatures, pressures, and flow rates at key points in the engine. Degradations in the engine’s performance will commonly manifest themselves as gradual shifts away from expected values corresponding to the current operating state of the engine, as illustrated in Figs. 41.3 and 41.4 (based on simulation results on C-MPASS engine model running at sea level static maximum power). It is recommended that the engine performance monitoring module implement a model-based approach based upon detecting and classifying these shifts as they occur. This technology assesses signal

1006

LPC Degradation Effects on Parameters 1 Changes from Nominal Value (%)

Fig. 41.4 Effects of low pressure compressor degradation on engine parameters

M.J. Roemer and L. Tang

0.8 LPT Outlet Temp

0.6

HPT Outlet Temp

0.4

Fuel Flow 0.2 N2

0 −0.2 −0.4 −0.6

Static Burner Pressure 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

LPC Efficiency & Flow Capacity Loss (%)

Fig. 41.5 Model based anomaly detection, diagnostics, and trending

health and system performance by employing a combination of signal processing, statistics, and data-driven modeling techniques. The process is illustrated in Fig. 41.5. Measured engine parameter data are input to the model, which provides expected values as output. These expected values are then evaluated relative to the original measured data to determine the residual value. These residual values used together form a pattern which can then compared to the characteristic degradations like those depicted in Figs. 41.3 and 41.4. Subsequent reasoning can then be applied to isolate the source of the performance degradation to a specific section of the engine. The techniques described here have been

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1007

implemented successfully in various gas turbine engine applications. Consideration of mechanical vibrations in the engine’s moving parts augments the engine performance evaluation to form a more complete propulsion monitoring system. Such techniques can be applied to drive train components specific to the engine such as rolling-element bearings. In addition, engine orders (1, 2, and higher harmonics of running speed) can be tracked to provide indications of a variety of faults. Critical component prognostics techniques based on fatigue accumulation models for specific components and engine configurations can provide an estimate of the remaining life before maintenance or overhaul is required. These models are typically a function of specific usage, mission profiles, speeds, and loads and can be adjusted by diagnostic indicators of the existence of faults. If such models are unavailable, then a much simpler approach can be implemented which only tracks total accumulated cycles (TAC) to provide a rudimentary approach to assessing engine life consumption. This approach, applied to gas turbines, monitors very generalized engine speed movements to gauge the severity of thermal transients and assess usage accumulation.

41.3.4 Flight Control Actuator Health Management Linear actuators are used extensively in UAVs for everything from manipulation of flight control surfaces to weapon release systems. One commonly applied type of actuator is the electromechanical actuator (EMA), which uses an electric motor and gear set to produce linear or rotary motion. EMAs have some advantages over equivalent hydraulic actuators; less weight, size, and complexity, characteristics that allow them to gain acceptance for aerospace applications. Prognostics and health management for EMAs typically employ two types of approaches to yield an overall assessment of EMA health. A model-based approach utilizing only command and response information from the EMA’s controller employs a core group of diagnostic features, including local gear stiffness, frictional damping, and torque constant to determine EMA system health. In this implementation, the model-based approach illustrated in Fig. 41.5 utilizes command signals as input and exercises a model of the actuator system varying the values for the three diagnostic features until a suitable match is found, measured by a minimization of the residual between the modeled response and the measured response. The set of the three diagnostic features then serve as the axes of a threedimensional failure space wherein different failure modes propagate through the failure space on different, unique paths. A second technique can utilize vibration data collected from an accelerometer to monitor the characteristics of the rolling elements in the actuator. This type of approach utilizes traditional vibration-based features to determine overall actuator health. Results from the two approaches can then be evaluated by a higher-level reasoning structure as described in the next section.

1008

M.J. Roemer and L. Tang

Critical Vehicle Systems System Remaining Functional Availability

System Health

Subsystems Condition Indicators External Sources

Enhanced Detection & Diagnostics

High-Level

Functionality Roll-Up

Assemblies Subsystem Health

Severity

LRU Health

Severity

Fault

Fault

External Sources

Subsystem Remaining Functional Availability LRU Remaining Functional Availability

Mid-Level

Low-Level

Line Replaceable Units Fig. 41.6 Hierarchical reasoning

41.3.4.1 Vehicle- and Subsystem-Level Reasoning One of the key aspects distinguishing an IVHM approach from a basic health and usage monitoring capability is an enhanced reasoning capability utilized to provide a vehicle-wide assessment of the implications of any detected failure mechanisms on the overall capability of the vehicle. Reasoning, within an IVHM system, processes condition or health indicators obtained from the subsystem-specific health modules to determine the current health state of the vehicle and its constituent subsystems. This goes beyond the basic diagnoses of observed indications by attempting to quantify the remaining functionality, which has been reduced by the indicted failure mode(s). The reasoning engine employs a hierarchical architecture, as illustrated in Fig. 41.6, to not only identify failure modes of line replaceable components (LRC) but also determine the functional impact in terms of remaining functional/operational availability at the subsystem and vehicle levels. When analyses result in an ambiguous outcome, the reasoner works to isolate the root cause or at least reduce the ambiguity group. The reasoner also attempts to provide estimates of the severity of the underlying failure and the remaining useful life of the LRC where prognostic capabilities are available. The reasoned results output by

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1009

the IVHM system provide real-time health and remaining functionality information useful to operations personnel for decision support. At the lowest level, diagnostic reasoning seeks to classify latent failure mode indications from raw sensor data or diagnostic feature data processed by the subsystem-specific modules. The mid-level of the reasoning architecture is employed to determine the overall functional availability of the constituent subsystems, i.e., what are the implications of the detected failure modes on the functional availability of the subsystem? The vehicle- or system-level reasoning, the highest level of onboard reasoning, shares this task of determining and quantifying functional availability, but from a vehicle- or system-wide perspective. At this highest level, the functional availability assessments, or condition indicators, from all underlying subsystems are utilized to determine the capability of the vehicle to continue operations as expected or needed.

41.3.4.2 Prognostics Performing onboard prognostics for a given set of LRUs or subsystems within the UAV IVHM system architecture is desirable but often times a challenge due to the necessary information and models needed to perform the task in real time. One such technique that has been shown to be useful for performing on-line prognostics for critical UAV subsystems is called the particle filter. Particle filtering is an emerging and powerful methodology for sequential signal processing based on the concepts of Bayesian theory and sequential importance sampling (SIS) and is suitable for the embedded applications onboard the IVHM system’s embedded element. Particle filtering can be used for a wide range of applications in science and engineering and is very suitable for nonlinear systems or in the presence of non-Gaussian process/observation noise. For example, the technique has been successfully used in the diagnosis and prognosis of a helicopter gearbox fault, gas turbines, and electrochemical batteries at Impact. The fault diagnosis procedure fuses and utilizes the information present in a feature vector (observations) with the objective of determining the operational condition (state) of a system and the causes for deviations from desired behavioral patterns. From a nonlinear Bayesian state estimation standpoint, this task is accomplished by the use of a particle filterbased module built upon the nonlinear dynamic state model. One particular advantage of this particle filtering approach is the ability to characterize the evolution in time of the above-mentioned nonlinear model through modification of the probability masses associated with each particle, as new feature information is received. Furthermore, the output of the fault diagnosis module, defined as the current expectation of each Boolean state, provides a recursively updated estimation of the probability for each fault condition considered in the analysis. The probability density function (PDF) estimates for the system’s continuousvalued states provide swift transition to failure prognosis algorithms, which is a primary advantage for a particle filter-based diagnosis framework. Since prognosis intends to project the current condition of the indicator in the absence of future measurements, it necessarily entails large-grain uncertainty. These facts suggest a prognosis scheme based on recursive Bayesian estimation techniques, combining

fault dimension

1010

M.J. Roemer and L. Tang

L*

L(tf )

L(t4 ) L(t1 )

fault detection

L(t0 ) t0

t1

t2

t3 t4

begin long term prediction

tÚ mean time to failure

time

Fig. 41.7 Illustration of prognostics

both the information from fault growth models and on-line data obtained from sensors monitoring key fault parameters (observations or features). The particle filtering-based approach for prognosis bases long-term predictions for the failure evolution on both an accurate estimation of the current state and a model describing the fault progression. This procedure tends to reduce the uncertainty associated with long-term predictions by using both the current state PDF estimation and a record of corrections made to previous computed predictions. Thus, the particle filtering algorithm serves the dual purpose of managing uncertainty and providing a means to continuously adapt the fault progression model. The probability of failure at any future time instant is estimated by combining both the weights of predicted trajectories and specifications for the hazard zone, as shown in Fig. 41.7: 6. The resulting Remaining Useful Life PDF provides the basis for the generation of confidence intervals and expectations for a prognosis.

41.3.5 Dynamic Flight Envelope Assessment The flight envelope of a UAV or any air vehicle represents the ranges of altitudes and speeds over which it effectively operates. This envelope may be drawn as a series of constraining curves on the altitude-Mach number plane. These constraints may arise owing to aerodynamic or propulsion considerations (e.g., the characteristics of the flight surfaces wings, and ailerons) or the power plant characteristics. The flight envelope (as determined for any particular aircraft) is usually an approximation and may be represented as a “doghouse plot” as shown in Fig. 41.8.

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

Service ceiling

D Maximum Speed

C Stall Speed

Altitude (1000 ft)

20

10

1011

B

A 0 0

0.5

1 Mach Number

1.5

2

Fig. 41.8 Representative flight envelope

The flight envelope in Fig. 41.8 is bounded by four curves. Points which are well within the envelope (such as “A” and “B”) represent “comfortable” operating points, where the aircraft has considerable excess capacity for maneuvers. Points close to the edges of the envelope (“C” and “D”) represent operation of the aircraft close to its limits, and thus, there is limited capacity for maneuvers at these points. Faults in UAV subsystems may affect the flight envelope significantly. Anomalies may occur in a variety of forms such as abrupt changes in aerodynamics, failed actuators, loss of propulsion, and damaged structures or control surfaces. In order to maintain vehicle stability and safe flight under adverse conditions, detection of the aircraft current state and prediction of achievable performance must be accomplished through onboard systems, in real time, and communicated to the flight control system, mission management system, and pilot. This onboard system must allow decision making to reduce risk while operating in the presence of uncertain information from various subsystems. For example, severe engine faults may reduce the available thrust, which affects the maximum speed and the service ceiling curves. Similarly, damages in the control surfaces may cause loss of lift or increased drag. This would lead to higher stall speeds, while also affecting the service ceiling and maximum speed. Thus, the flight envelope would shrink due to engine or flight control surface faults. The maximum speed attainable by the aircraft depends on the maximum engine thrust, as well as the wing area and the minimum attainable drag coefficient. This may be expressed as 1 2 max D  .H / Cdmin Aw Vmax 2s (41.4) 2max Vmax D  .H / Cdmin Aw where  max is the maximum thrust,  is the density of air (as a function of altitude H ), Cdmin is the minimum drag coefficient, Aw is the wing area, and Vmax is the

1012

M.J. Roemer and L. Tang

maximum possible air speed. Thus, loss of thrust due to engine degradation as well as increased drag owing to wing damage may both reduce the maximum speed, which shrinks the flight envelope to the right. Similarly, the stall speed may be calculated as 1  .H / CLmax Aw Vs2 2 s 2W Vs D  .H / CLmax Aw

W D

(41.5)

where W is the weight of the aircraft, CLmax is the maximum lift coefficient, and Vs is the stall speed. Reduced lift (owing to wing faults) could thus cause the stall speed to increase, which shrinks the flight envelope to the left. Due to the experimentation and modeling capabilities available to the author, dynamic flight envelope estimation was initially investigated for severe engine faults. The reduction in the flight envelope due to an engine fault was simulated using the NASA MAPSS engine model (Parker and Melcher 2004). Specific engine faults were seeded into the MAPSS model which causes significant efficiency and flow capability loss in several engine components. The corresponding changes in the flight envelope were then estimated in real time by estimating the maximum thrust that the engine could produce under a particular level of degradation. As illustrated in Fig. 41.9, the maximum thrust estimate was provided by an artificial neural network which was trained off-line using the data from high-fidelity engine model simulations with engine controllers in the loop.

41.4

ACM System Optimization

Most ACM strategies for large-scale dynamical systems, such as an aircraft, rely on heuristic information about a reduced set of severe, frequent, and testable fault modes, a reasonable number of active controllers, and a mapping between the fault modes and the control reconfiguration routines. Such a strategy has the ability to adaptively switch from one controller to another, if control limits are reached, and by this switching action, critical mission objectives can be realized. This chapter presents a new approach in which an optimization problem is dynamically formulated and solved on-line to solve the optimal contingency strategy constrained by the available performance and resource. A typical cost model is expressed in terms of such elements as on-time execution of critical components, time to complete the mission, tracking error, fuel consumption, etc. Costs are computed in real time and may change dynamically. Analytically, the objective of the ACM system is to optimize the utility of the vehicle with impaired capability to accomplish an assigned mission. The ACM system can be formulated as an optimization problem in two levels:

41 Integrated Vehicle Health and Fault Contingency Management for UAVs Off-line

Fault/degradation scenario 1

Maximal Thrust

1013

Flight Envelope

High fidelity Engine Model Simulation scenario 2 High fidelity Engine Model Simulation

…

… scenario n

High fidelity Engine Model Simulation

On-line Fan Efficiency Loss

Sensor Measurement

Health Estimation

…

Engine

Predicted Maximal Flight Envelope Assessment Thrust

HPT Efficiency Loss

Fig. 41.9 Neural network-based dynamic flight envelope estimation for severe engine faults

High (mission)-level planner: J.M / D max U.Pe ; Pr ; M; Mcom / M

(41.6)

Lower (control reconfiguration) level: J.R/ D max Pe .Fm ; Pr ; R; M / R

(41.7)

where U is a cost function that quantifies the usefulness of the vehicle to accomplish its mission. U is a function of the available prognostic information (Pr ), the system’s closed loop performance (Pe ), and the mission objectives (M and Mcom ). Pe is a function of fault mode Fm , future prediction Pr , as well as any restructuring/reconfiguration R applied to the system and current mission objective M . Fm is a vector of indicators (0 or 1) that characterizes the fault modes detected on the aircraft; R is a vector of indicators that characterizes all restructuring applied to the system. Mcom describes the mission assigned to the aircraft. M allows the fault-tolerant control architecture, specifically the mission adaptation and resource management components, to modify the parameters of the assigned mission and redistribute available resources based on the vehicle’s current performance, Pe . At the high level, mission adaptation and resource redistribution (M / allow the control architecture to pursue relaxed mission objectives in order to achieve greater vehicle usefulness U . At the lower level, the objective is to optimize vehicle performance Pe while satisfying the mission constraints, through

1014

M.J. Roemer and L. Tang

restructuring and reconfiguration, R. Practically, the above optimization problems have to be solved while adhering to various constraints including system dynamics and resource limitations. While this generic problem formulation applies to all of the ACM levels (mission management, adaptive flight control, and propulsion control) in Fig. 41.1, the focus of this chapter is adaptive flight control. Fault-tolerant propulsion controls and mission reconfiguration are discussed in the previous publications (Tang et al. 2005, 2007b).

41.4.1 Model Predictive Control Model predictive control (MPC) is a particularly effective method for control reconfiguration due to its ability to handle constraints and changing model dynamics systematically. MPC relies on an internal model of the system, which can be identified on-line in real time using the strong tracking system identification algorithms described in previous section. System failures can be handled naturally in an MPC framework via changes in the input constraints and internal model. For example, actuator limit and rate constraints can be written as: uli  ui .t/  uhi (41.8) uli  uP i .t/  uhi for actuator inputs u1 through um . If actuator i becomes jammed at position uQ i , the MPC controller can be made to accommodate the changes by simply changing the constraints on input i to uQ i  ui .t/  uQ i (41.9) 0  uP i .t/  0 Structural failures can also be handled in a natural fashion by changing the system dynamics model used to make prediction in either an adaptive fashion, a multi-model switching scheme or in this case, by using a real-time system identification module to provide the estimated degraded system dynamics. MPC fault-tolerant control can be cast as a constrained model-following problem as shown in Fig. 41.10. The controller comprises three main components: the block “real-time system ID” which performs identification of the fault’s effects on system dynamics, a “reference model” which uses pilot commands to generate a reference trajectory for the aircraft’s state vector, and the MPC controller whose objective is to track the reference trajectory, using the output of “real-time system ID” to update its internal model, constraints, etc. The pilot gives commands to the reference model, and the goal of the controller is to cause the aircraft to track the resultant trajectory. At each time step, the MPC controller chooses an input sequence which minimizes the difference between the predicted future trajectory, given by the reference model

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1015

MPC Fault Tolerant Controller

Pilot commands

Reference Model

MPC Controller

Aircraft

Real-time System ID

Fig. 41.10 MPC fault tolerant control architecture

under the assumption that the pilot’s inputs are constant over the prediction horizon, and the predicted trajectory of the aircraft.

41.5

Flight Envelope Assessment and Fault-Tolerant Control Simulation Results

The NASA MAPSS engine model (Parker and Melcher 2004) was utilized to simulate engine faults for the evaluation of the fault detection and resulting dynamic flight envelope assessment techniques previously described. The self-tuning Kalman filter was used to identify and estimate the severity of the faults in terms of efficiency loss and/or flow capacity loss. At the flight control level, the NASA GTM aircraft model (Bailey et al. 2005) was utilized, and various actuator faults including control effectiveness loss and stuck actuators were simulated. Some select simulation results are presented below.

41.5.1 Dynamic Flight Envelope Assessment In order to quantify the flight envelope, it was assumed that the thrust produced by the MAPSS model was utilized by an aircraft with certain predetermined characteristics. The characteristics of the aircraft are listed in Table 41.1. Within the simulation, pressure and temperature were varied as shown in Fig. 41.11, with the pressure assumed to drop roughly exponentially with altitude, while the temperature drops linearly with altitude up to a point, beyond which the temperature stays constant.

M.J. Roemer and L. Tang

Table 41.1 Aircraft characteristics

Parameter Aw CLmax Cdmin ˆ

Pressure (kPa)

M TSL

Explanation Surface area of wings Maximum lift coefficient before stall Minimum drag coefficient Equivalence ratio at sea level Total mass of aircraft Sea level thrust

Value 27.88 1.6

Units m2 None

0.01 0.7

None None

10,780 130

Kg kN

150

300

100

200

50

100

0 0

5000

10000 Altitude (m) Pressure

15000

Temperature (K)

1016

0 20000

Temperature

Fig. 41.11 Variation of pressure and temperature with altitude

Figure 41.12 shows a plot of the instantaneous flight envelope for a given level of engine fault. The NASA MAPSS software was modified with a new plot showing the estimated flight envelope. The result in Fig. 41.13 shows how the flight envelope shrinks as a function of the engine fault severity level (in this case, the HPT efficiency loss).

41.5.2 Detection of Control Effectiveness Loss With the strong tracking recursive identification algorithm, the A and B matrix of the linearized aircraft state-space model are identified in real time. Various scenarios including elevator/rudder/aileron/throttle effectiveness loss and aircraft weight loss have been simulated and evaluated with the fault-tolerant controller. Figure 41.14 shows the identified control effectiveness change by the strong tracking recursive identification algorithm and the comparison with the result from classic recursive least square method. The true value is changed from 1.98 to 0.198

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1017

Fig. 41.12 Dynamic flight envelope assessment results added to NASA MAPSS GUI

Fig. 41.13 Flight envelope reduction due to severe engine fault added to NASA MAPSS GUI

at t D 20 s (i.e., a 90 % control effectiveness loss). The result shows that the tracking performance of the strong tracking recursive identification algorithm (blue solid curve) outperforms the classic recursive least square method algorithm (red dashed curve) in tracking a sudden change in control effectiveness of the elevator.

1018

M.J. Roemer and L. Tang

identified B(7,1)by STRI (solid line), identified B(7,1)by RI (dashed line); True B(7,1) (dotted line)

2 1.8 1.6 1.4 1.2 STRI

1

RI

0.8 0.6 True B(7,1) from 1.98 to 0.198 at t=20 sec.

0.4 0.2 0

20

10

0

40

30

50

60

Fig. 41.14 Detection of control effectiveness loss

b

2000 1800 1600 1400 1200 1000 800 600 400 200 0

3 2

Actuator Command

Altitude

a

0

100

200

300

400

500

600

700

1 0 −1

elevator (MPC) throttle (MPC) elevator (nominal) throttle (nominal)

−2 −3

0

100

200

300

Time

400

500

600

700

Time

Fig. 41.15 MPC simulation results with stuck elevator actuator (t D 75). (a) Altitude response and (b) controls

41.5.3 Model Predictive Control for Stuck Elevator An adaptive model predictive controller that utilizes the system dynamic model identified in real time was developed and tested with a stuck elevator fault scenario. A classic altitude hold controller was developed which utilizes throttle to control airspeed and elevator for altitude tracking. As shown in Fig. 41.15, after the occurrence of a stuck elevator fault at time 75, the aircraft would keep climbing up

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1019

without the fault-tolerant controller because the elevator is stuck at a climbing-up position (the black curve in Fig. 41.15a). With the adaptive MPC controller, throttle is used to control longitudinal dynamics, and the aircraft was able to accomplish the mission (the red curve in Fig. 41.15a) with degraded performance (the dashed blue curve in Fig. 41.15a). The elevator and throttle controls and the comparison with nominal controls are shown in Fig. 41.15b.

41.5.4 Prognostics-Enhanced ACM System The safety of the unmanned vehicle involved in long-duration missions is dependent on not only the current health state (diagnostics) but also near-future health states (prognostics) as well. While a diagnostics-driven ACM system can react to and compensate for faults and performance degradation after they are detected, it cannot hope to overcome the fact that it will always be a reactive paradigm. As a result, the control strategies may in fact not be optimal over a longer period of time. By incorporating near-term prognostic information, undesirable future states of the system may be avoided or deferred through a suitable change in the control strategy while achieving required (or relaxed) mission objectives. In case a failure in certain system component is inevitable, a prognostics-enhanced ACM system can possibly minimize the effect or propagation of the failure by appropriate control reconfiguration and isolation strategies. To illustrate the concept, one can consider the case of emergency landing of a distressed aircraft with a rudder failure (floating rudder with total loss of control authority). The loss of control is taken up by the engines operating in a thrustenhanced, differential throttle mode. However, traditional engine time constants are too slow to be used in this situation. A fast-response engine controller is needed to reduce the time constants by pushing the engine operation more aggressively. As a result, engine stall margin and engine life can be significantly reduced which may lead to catastrophic engine failure. Therefore, competing requirements on aircraft performance (demanding aggressive engine operation) and engine safety (demanding mild engine operation) must be considered within the ACM framework. The ACM strategy implemented is a supervisory safety verification controller (SSVC) which is a middle-level controller that determines low-level control configuration (switching between slow-response and fast-response engine controllers) at a certain decision interval. This level of control specifies parameters that the lowlevel controllers must assume in order to reallocate the controls once a fault has been diagnosed. It must be noticed that the low-level controllers (either slow or fast response) are responsible for the stability and performance of the engine while the SSVC optimizes the engine operation to achieve required aircraft performance while avoiding any violations of engine safety limits. Since the SSVC only operates at a near real-time decision interval (typically every couple of seconds), more computationally demanding optimization algorithms can potentially be applied at this level of controls.

1020

M.J. Roemer and L. Tang

FDI ↓ Safety Verification Configuration Specified Track/Heading

Line-of-Sight Guidance (autopilot)

Commands

Damaged Closed-Loop Aircraft

Position, Velocity, Attitude, Attitude rate

Fig. 41.16 Diagram of the safety verification controller using diagnostics and prognostics

Figure 41.16 illustrates the overall structure of a supervisory safety verification controller. The aircraft states including diagnostic and prognostic information are fed into an FDI/safety verification module which, in turn, computes the optimal control configuration given the fault and mission profile. The adopted configuration dictates which low-level controls are used and how the control is allocated (i.e., whether a conservative or aggressive policy should be adopted at any given time). A cross-track guidance system is adopted in lieu of a pilot as a baseline to provide commands to the flight controller (DeCastro et al. 2011). For this emergency landing case study with rudder failure, the UAV’s contingency envelope can be defined by three criteria: 1. Performance criteria: The maximum allowable error from the reference trajectory is 1,500 feet. 2. Operability limits: Engine stall margin is maintained above zero to avoid immediate loss of thrust and potential engine failure. 3. Prognostic (life) limits: The prognostic criteria penalize any trajectory or control policy that results in a large probability of accumulated engine failure by the end of the mission. When considering engine prognostic algorithms, various types of prognostic algorithms may be used to track engine damage, including particle filter, Kalman filter, or data-driven algorithms. In order to use these algorithms in the supervisory control, the essential behavior of the algorithm is abstracted to a finite-state machine (FSM). The automaton discretizes the trajectory of the engine damage index and assigns transition probabilities to each node as a discrete-time Markov chain. Figure 41.17 shows how the probability of crossing the failure threshold is mapped to an FSM. The unsafe set corresponds to the failed state of the engine. It is important to note that the abstraction allows the prognostic model to be captured in the ACM optimization in a simplified way without losing essential information. To demonstrate the overall reconfiguration concept, simulations are executed with the GTM simulation using only fast and slow engine control reconfiguration options. In the following scenarios, a new control decision is made every 10 s.

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1021

Damage

Failed

Transform algorithm into FSM

damage pdf

Unsafe safe Severe

Failure Threshold

Moderate

Light

t0

thoriz

Time

No damage

Fig. 41.17 Transforming a prognostic algorithm into a finite-state machine

Scope

1 input

I AC state eng state v_ref input generator

state 1

Scope1

inputs

state dot

OL/CL

aux var

wind generator

IO to vector

wind

aircraft model

1 state 3 state dot 2 auxvar 4 ctrl pos 5 eng state

Scope2 state v_ref, phi_ref state derivs Los guidance

Fig. 41.18 Simulink diagram of the NASA full-scale GTM with safety verification control capability

The additive uncertainties to simulate effects such as wind gusts were modeled as Gaussian noise on each of the aircraft’s states. A Simulink model developed for this study is shown in Fig. 41.18. In this simulated scenario, the GTM is making a series of coordinated turns to reduce the cross-track error to a specified parallel track to prepare for landing. The cross-track trajectory is shown in Fig. 41.19. For the 200-s simulation, the reconfiguration policy along with the trajectories on the aircraft’s lateral position and the stall margin between engines are shown in Fig. 41.20. The policy switching is dominated initially by the stall margin requirement by starting in the slow configuration. After 80 s, the control attempts to meet the cross-tracking performance

Y (feet)

1022

M.J. Roemer and L. Tang

6000 4000 2000 0 1

0

2

3

4

5

6

7

8 x 104

X (feet)

a

6000 4000

slow

0 −2000

100

50

150

200

0

50

150

200

250 roll pitch yaw

10 0 −10 0

50

250

100

150

200

250

200

250

time (sec)

d 45.5

0.4

aileron rudder

0.2 −0.2 −0.4

0

50

100

150

200

250

2 Δ stall margin

45

0 Damage index

control surface deflections (deg)

c

100

20

−20 0

actual trajectory reference trajectory

2000

attitude (deg)

fast

b cross track (feet)

Fig. 41.19 UAV cross-track trajectory

44.5 44 43.5

1

43

0

42.5

−1

42 0

50

100 150 time (sec)

200

250

0

50

100 150 time (sec)

Fig. 41.20 Optimal configuration and states for SSVC with consideration of prognostics. (a) Optimal control sequence, (b) aircraft states, (c) aircraft control inputs and engine stall margin, and (d) engine damage index

requirement by switching to the fast configuration. The ACM strategy switches at every decision instant in order to improve the remaining useful life of the engine, thereby reducing the probability of failure before the end of the mission. Although the controls are not configured to mitigate bumpless transfer, the switching does not introduce oscillations in the overall response of the aircraft. Both the cross-track performance error and stall margin remain below thresholds throughout the horizon. The figure also shows the role of the prognostic model on the ensuing damage accumulated on the engines during the maneuver. The hypothetical hybrid abstraction of the prognostic algorithm is applied, assuming that the system is

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

b cross track error (feet)

a fast

1023

6000

actual trajectory reference trajectory

4000 2000 0 −2000

0

50

100

150

200

attitude (deg)

20

slow

50

100

150

200

0 −10

d

−0.2 −0.4

0

50

100

150

200

250

1 Δ stall margin

50

100

150

200

250

200

250

time (sec) 48 47

0 Damage index

control surface deflections (deg)

aileron rudder

0.2

0

250

c 0.4

roll pitch yaw

10

−20 0

250

46 45 44

0 43

−1 −2

42 0

50

100 150 time (sec)

200

250

0

50

100 150 time (sec)

Fig. 41.21 Optimal configuration and states for SSVC without consideration of prognostics. (a) Optimal control sequence, (b) aircraft states, (c) aircraft control inputs and engine stall margin, and (d) engine damage index

initially in the moderate damage state. When prognostic information is included in the optimization, the damage index reaches 45.5 at the end of the mission. As a comparison, the UAV cross-track control results without engine prognostics in the loop are shown in Fig. 41.21. Similar to the previous result, the ACM policy switching is dominated initially by the stall margin requirement by starting in the slow configuration. After 80 s, the control attempts to meet the cross-tracking performance requirement by switching to the fast configuration and remains in this control configuration for the duration of this simulation. As a result, the engine damage index has grown to about 48 (as compared to 45.5 in the previous simulation) at the end of the simulation. This is a clear evidence of an engine safety threat when engine life (prognostics) has been ignored in the ACM optimization.

41.6

Conclusions and Future Work

This chapter presented selected technologies for implementing an integrated aircraft health assessment and fault contingency management system. An innovative hierar-

1024

M.J. Roemer and L. Tang

chical fault contingency management architecture that integrates real-time system health management functionality, high-level reasoning, control reconfiguration, and automated contingency management modules has been developed. Real-time system identification and health assessment algorithms have been implemented to identify the dynamics and performance limitation of the degraded system. An intelligent approach has been developed to estimate safe flight envelope after the occurrence of engine faults. MPC-based reconfigurable fault-tolerant control techniques have been developed and tested in simulation to accommodate flight actuator faults using the system dynamic model identified in real time. Simulation studies using the NASA MAPSS engine model and GTM aircraft model have successfully demonstrated the feasibility of the presented concept and architecture. This chapter further explored a more advanced topic on incorporating prognostic information in a proactive ACM system. Simulation studies on a GTM aircraft model have been conducted and demonstrated the benefits of using near-term prognostics in control reconfiguration. This capability upgrades the conventional diagnosticsbased reactive fault-tolerant strategies to a prognostics-enhanced proactive fault accommodation paradigm. To further develop the ACM concept and eventually integrate it into real-world autonomous systems, several important issues need to be addressed. First, the diagnostic and prognostic information fed into the ACM system have to be reliable, accurate, and available in real time. This is certainly a challenging requirement on the design of onboard PHM system. Prognosis uncertainty management is another important issue that affects the performance of the ACM system significantly. In practice, accurate and precise prognostics has proven rather difficult to accomplish; thus, uncertainty management and reduction techniques have to be implemented before the prognostics can be optimally utilized by the ACM system. Furthermore, in order for ACM systems to be used in safety-critical aerospace applications, they must be proven to be highly safe and reliable. Rigorous methods for ACM system verification and validation (V&V) must be developed to ensure that ACM+P system software failures will not occur, to ensure the system functions as required, to eliminate unintended functionality, and to demonstrate that certification requirements can be satisfied.

References R.M. Bailey, R.W. Hostetler, K.N. Barnes, C.M. Belcastro, C.M. Belcastro, Experimental validation: subscale aircraft ground facilities and integrated test capability, in AIAA Guidance, Navigation, and Control Conference, Washington, DC, 2005 M. Bodson, An adaptive algorithm with information-dependent data forgetting, in Proceedings of the American Control Conference, Seattle, WA, 1995, pp. 3485–3489 C.S. Byington, M. Watson, D. Edwards, P. Stoelting, A model-based approach to prognostics and health management for flight control actuators, in Proceedings of the IEEE Aerospace Conference, Big Sky MN, paper 1047, 2004 J.A. DeCastro, L. Tang, B. Zhang, G.J. Vachtsevanos, A safety verification approach to faulttolerant aircraft supervisory control, in AIAA Guidance, Navigation and Control Conference and Exhibit, Portland, OR, 8–11 Aug 2011

41 Integrated Vehicle Health and Fault Contingency Management for UAVs

1025

J.S. Litt, D.L. Simon, S. Garg, et al., A survey of intelligent control and health management technologies for aircraft propulsion systems. J. Aerosp. Comput. Inf. Commun. 1(12), 543– 563 (2004) K.I. Parker, K.J. Melcher, The modular aero-propulsion system simulation (MAPSS) users’ guide, NASA TM-2004-212968, 2004 L. Tang, G.J. Kacprzynski, M.J. Roemer, G. Vachtsevanos, A. Patterson-Hine, Automated contingency management design for advanced propulsion systems, in Infotech@Aerospace, Arlington, VA, 26–29 Sept 2005 L. Tang, M. Roemer, G. Kacprzynski, J. Ge, Dynamic decision support and automated fault accommodation for jet engines, in IEEE Aerospace Conference, Big Sky, MT, 3–10 Mar 2007 (2007a) L. Tang, G. Kacprzynski, K. Goebel, J. Reimann, M. Orchard, A. Saxena, B. Saha, Prognostics in the control loop, in The 2007 AAAI Fall Symposium on Artificial Intelligence for Prognostics, Arlington, VA, 9–11 Nov 2007 (2007b) G. Vachtsevanos, L. Tang, G. Drozeski, L. Gutierrez, From mission planning to flight control Of unmanned aerial vehicles: strategies and implementation tools. Ann. Rev. Control 29, 101–115 (2005) G. Vachtsevanos, F.L. Lewis, M. Roemer, A. Hess, B. Wu, Intelligent Fault Diagnosis and Prognosis for Engineering Systems (Wiley, Hoboken, 2006) A. Volponi, B. Wood, Engine health management for aircraft propulsion systems, in First International Forum on Integrated System Health Engineering and Management in Aerospace, Napa, CA, 7–10 Nov 2005

Automated Failure Effect Analysis for PHM of UAV

42

Neal A. Snooke

Contents 42.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1028 42.1.1 Prognostics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1028 42.2 Model-Based Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029 42.3 Failure Modes and Effects Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1031 42.4 Functional Description. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034 42.4.1 Generating FMEA Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035 42.5 Sensor Selection and Diagnosability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1038 42.5.1 Generating Symptoms from an FMEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1041 42.5.2 Associating Measurement to Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1043 42.5.3 Fault Exoneration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044 42.5.4 FMEA Coverage and Symptom Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044 42.6 Diagnosability Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047 42.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1050 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1050

Abstract

This chapter describes how model-based simulation can be employed to automatically generate the system-level effects for comprehensive sets of component failures on systems within the aircraft. The results of the simulation can be used in several ways. They can be used to produce a system-level failure modes and effects analysis (FMEA) for aircraft systems. They can be used to identify the sensors necessary to discriminate remotely between different failures on the aircraft. Once a set of sensors have been chosen for placement on the vehicle, the simulation results can also be used to generate diagnostic and prognostic software for deployment on the vehicle.

N.A. Snooke () Department of Computer Science, Aberystwyth University, Llandinam Extension, Aberystwyth, Ceredigion, UK e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 40, © Springer Science+Business Media Dordrecht 2015

1027

1028

N.A. Snooke

Using automated FMEA safety analysis software is more efficient than doing the same work without software and also provides a guaranteed level of performance. Using the results of this analysis can provide sensor selection and diagnostic capability while retaining some of the benefits of rule-based diagnostic systems. Alternative model-based techniques have been widely used to create diagnostic systems in a variety of domains, and these approaches are compared with the diagnostic capability provided by a failure effect-oriented technique from the perspective of the UAV application.

42.1

Introduction

Prognostics and health management (PHM) comprises the technology and systems needed to enable UAVs to monitor their own state. Based on that monitoring, the PHM systems need to make assessments of mission readiness that can be utilized by higher-level planning in the decision modeling area. One recent UK Project specifies the goal that for use in unsegregated airspace, PHM systems should be capable of replicating the fault detection, assessment, and decision-making ability of pilots. The current generation of computer systems falls far short of the highlevel reasoning and decision-making capabilities of an experienced pilot; however, this limitation can be mitigated by the use of sophisticated and high-coverage lowerlevel fault detection and reconfiguration capabilities. This chapter will show how model-based reasoning and simulation can be used to automate the process of producing a comprehensive failure modes and effects analysis and how this information can then be used to assist in the assessment of diagnostic capability and sensor selection. The techniques of this chapter were combined and integrated during the first phase of a UK government-funded program: Autonomous Systems Technology Related Airborne Evaluation and Assessment (Astraea 2009). The aim of the ASTRAEA program is to enable the routine use of UAS (Unmanned Aircraft Systems) in all classes of airspace without the need for restrictive or specialized conditions of operation. This is achieved through the coordinated development and demonstration of key technologies and operating procedures required to open up the airspace to UAS.

42.1.1 Prognostics Prognostic capability provides a forecast of some future occurrence that may impact on a system. This general concept is interpreted in different ways dependent on the level of abstraction being used. To component engineers, the aim is usually to predict component failure or wear out before failure or degradation occurs that will negatively impact on system behavior. Competing objectives for reduced maintenance schedules and cost lead to activities such as condition-based maintenance. Prognostics at this level are inherently difficult and require detailed numerical monitoring and usage data. In some domains such as mechanical systems symptoms

42 Automated Failure Effect Analysis for PHM of UAV

1029

such as increased vibration can provide prognostic indicators; however, in other areas such as electrical components, there are often very few if any indications prior to catastrophic component failure. In some systems data such as usage pattern logging can be used in models that predict degradation and assist in condition-based maintenance. The specialist modeling and limited applicability of these systems are outside of the scope of this chapter. At the systems or mission level, prognosis involves the ability to reason about the impact of the loss of low-level functionality on higher-level objectives. For example, the detection of a fuel leak may require a change of mission objective (abort, emergency landing, etc.), and the prognostic task is to determine available scenarios and select appropriate action. The prognostics and health management (PHM) area is responsible for developing the technology and systems needed to enable UAVs to monitor their own state. Based on that monitoring, the PHM systems need to make assessments of mission readiness that can be utilized by higher-level planning in the decision modeling area. These tasks are usually considered as the vehicle health management systems and are addressed in other chapters of this book. The input information to these prognostic tasks is the result of system or subsystem design and diagnosis, and this will be the focus of this chapter.

42.2

Model-Based Reasoning

Model-based systems and qualitative reasoning (MBS&QR) as a visible subfield of artificial intelligence can be traced back some 20 years to the publication of seminal collection of papers in the field (Bobrow 1984; Hamscher et al. 1992), although there was of course earlier work in this area (de Kleer 1977; Brown et al. 1982). Model-based reasoning encapsulates a number of key concepts identified in (Price et al. 2006b; Peischl and Wotawa 2003) as follows: • Separation of system model from problem solving. Model-based systems are based on a separation of the problem-solving algorithm from the model of the domain. Once a library of appropriate component models has been established, only a structural description of the respective device or system (e.g., obtained from design data) is required to automatically generate a system model and, based on it, problem-solving software dedicated to this device or system. • Compositionality. Since systems are assembled from standard components and the behavior (and misbehavior in the case of a fault) of the system emerges from the behavior of these components, establishing a model library is feasible and entails collecting models of (correct and faulty) behavior of such standard components. This is important: this kind of model-based reasoning cannot be performed if the overall behavior of the system cannot be composed from the behaviour of the components and the way in which they are linked. Where there is compositionality of models, a high degree of reuse is possible, as well as prediction of what will happen in unexpected circumstances such as failure situations.

1030

N.A. Snooke

• Existence of different abstract problem solvers. The combination of generic domain independent problem solvers and compositional models of a domain can produce model-based systems that are able to reason efficiently about a product using relatively easy to obtain and objective design data. • Modeling at different levels of abstraction. In diagnosis, for example, modeling different kinds of problems may well involve modeling different phenomena and at different levels of detail. However, while there may be a need for quantitative and or semiquantitative models, qualitative models provide a nice solution for representing phenomena at a higher level of abstraction. Model-based reasoning (MBR) appears in standard AI texts (Luger 2002) and has been widely used in several domains for a variety of tasks as documented in the review by Price et al. (2006a). In particular, the efficiency and coverage provided by qualitative models have many advantages in the early stages of design and in the natural abstraction of explanations provided by such simulations and analysis. The efficiency of MBR is particularly suited to the task of failure modes and effects analysis due to the high level of system state coverage required combined with the large number of failure modes to be considered. An FMEA must present failure modes and associated risk together with the possible causes of the failure mode. The abstraction provided by QR (Kuipers 1986) makes it possible to identify failure modes and effects using language that is intuitive to an engineer. For example, an explanation that a low-pressure failure mode has the effect of reduced engine output power or shutdown and is caused by a pipe blocked fault may be derived directly from a qualitative model. Detailed numerical models are often difficult to obtain, particularly early in the design process when it is still feasible to change the design and sensing requirements. This consideration is of particular importance if the objective is to design diagnosis capability and mitigation into the system rather than retrofitting it to a completed design. The major limitation of qualitative techniques is that in some circumstances the simulation can predict alternative behaviors. For example, a qualitative electrical circuit containing a bridge network may predict current flow in either direction across the bridge resistance because the flow depends on the ratio of the surrounding resistance values. Similarly, for systems that use a qualitative representation of time, there may be alternative behavior predictions dependent upon the ordering of events within the same qualitative time period. This chapter considers a modeling strategy that ensures qualitative ambiguity does not become a limiting factor and that it is related to physical system characteristics that have real behavioral significance. Given these constraints qualitative ambiguity actually provides useful additional information by identification of critical parameters within the system and can be addressed by adding additional constraints to the model (effectively making explicit a design specification) or appealing to selected numerical simulation for particular situations. The following sections will consider how model-based reasoning can be used to automatically generate a comprehensive FMEA. This report contains a great deal of diagnostic information and may also be used to create workshop diagnosis or to automatically generate a diagnostic system.

42 Automated Failure Effect Analysis for PHM of UAV

42.3

1031

Failure Modes and Effects Analysis

Manual FMEA is generally carried out by identification of failure modes and subsequent derivation of system effects and potential root causes, such as component failure, by an engineering team. Automated FMEA generally approaches the problem from the bottom up by performing system simulation for each potential fault and interpreting the resulting behaviors in terms of system functionality (Price 1998). The main concept is that a description of the overall behavior of a system can be constructed by knowing the structure of a system and the behavior of each of its components. In order to determine the behavior of the system when a failure occurs, it is only necessary to replace the component that has failed with a version of the component that reproduces the faulty behavior, and the model of the system will then reproduce the effects of the failure for the whole system. Figure 42.1 shows the different levels of reasoning needed to perform modelbased simulation of a system containing electrical and/or fluid flow components. The system description is usually provided at the component level – the system is described in terms of its components and their states and connections. This can be mapped onto the resistive state of all components, and the state of the overall system can be determined qualitatively (CIRQ). Flow and effort variables are global in nature and cannot be derived by propagating local component input/output behavior. At the qualitative level, a generalized solver (Lee 1999) can be used to determine global current and voltage in the electrical domain or pressure and

'Lamp Illumination' ACHIEVED Functional level reasoning active functions deduced from behaviour

12V

switch

+ −

Component level reasoning

fuse relay

resistive state is passed downwards

0

0/

L 0/

8

0

component states changes passed upwards

8

S1

lamp

Fig. 42.1 Levels of electrical simulation

G1 L

Qualitative circuit based reasoning

1032

N.A. Snooke

flow in a hydraulic or fluid transfer system. Alternative approaches such as bond graphs (Borutzky 2010; Mukherjee and Karmakar 1999) can also be used to provide similar information, although this may require more modeling effort and careful consideration of the causal flow through the circuit. For the purpose of FMEA, and the generation of rule-based diagnostics, and as a natural consequence of qualitative analysis, a state-based modeling approach is often adequate at the component level and used in preference to a complex dynamic numerical model for most cases. For states or subsystems where a qualitative or state-based model cannot provide the required behavioral detail or is unable to answer the required questions, numerical models can be used. Changes to the effort and flow variables may result in changes in component state (e.g., if current is now flowing through a relay, then it may close, changing the state of the relay component, necessitating further qualitative simulation). Eventually, the state of the system will either stabilize or oscillate. Local component behavior such as the filling of a tank or the logic of an electronic control unit (ECU) is modeled at the component level often using finite state machines (FSM) (Harel and Politi 1998). The FSM representation is widely used to represent abstracted system or component behavior, being used to represent systems as diverse as communications protocols (SDL) and software behaviour (UML) (UML 2011). The results of the interchange between the component level reasoning and the qualitative electrical reasoning can be abstracted to the functional level and returned in terms of which functions of the system have occurred. Figure 42.1 illustrates the relationship between the three levels for a very simple system, but these techniques work for electrical/electronic systems with several thousand components and produce useful FMEA results that pinpoint significant potential failures early in the design process. To summarize the information required: • Component models or a component library. Each model contains structural information such as a resistance network between the terminals of the component and also behavioral information such as a state chart that controls changes to the structural elements. • A system schematic that connects individual component instances. • A description of the system functions and which behaviors provide these functions. A choice must be made as to the level of granularity required for the qualitative variables. A minimal but surprisingly powerful scheme for qualitative resistance is f0, load, 1g. In this scheme, an open switch has 1 resistance and a closed switch has zero resistance, and a power-consuming device is represented as a load. From this information the simulation can derive that a fractured wire will cause a motor to stop or that a bad connector (load) will lower the voltage across a motor. This is enough information to derive functional effects of an FMEA and for diagnosability analysis. There are situations where a finer-grained model is useful, for example, when dealing with event durations using a qualitative scale such as fday, hour, second, mS, uSg is useful to avoid ambiguity. For example, if two components have state-changing behavior, the potential event ordering ambiguity between a tank

42 Automated Failure Effect Analysis for PHM of UAV

[coil.i == 0] deactivate / switch.R =

Active

8

Fig. 42.2 Example simple relay component model

1033

Inactive

Behaviour model

[coil.i == ACTIVE] activate / switch.R = 0

resistive element

terminals

L 86

coil

85

Qualitative Structure model

8

87

switch

30

emptying and a relay switching is resolved. The tank model may specify that at nominal flow a partly filled tank will take “hours” to empty whereas the relay will switch in “mS.” Care must be taken with the semantics of such a scheme since qualitatively it is specifying that any number of sequential events in one qualitative timeslot will finish before any events in the next timeslot. Therefore, the qualitative model must not be too fine grained, and each level must represent a qualitatively significant and distinct region of behavior. A major benefit of the qualitative modeling scheme is that the models are highly reusable since none of the numerical parameters are present. The model library will therefore contain a single model to represent a whole class of actual device part numbers. For example, a single model covers many “normally open single pole” relays. As an example Fig. 42.2 shows a structural and behavioral model for a basic relay. Such models are easy to produce and can be used in early stage design before detailed parameterization of the system is possible allowing early failure mode analysis and diagnosability characterization to be carried out. The failure behaviors for each component failure mode are provided by a modification of the nominal component structure or behavior model. For example, a blocked pipe will simply change the resistance of the pipe to “infinite.” These faults are created for each type of component and are stored in the component library as part of the component description. Faults are then automatically inserted for each instance of the component present in a system. This section has described how nominal and failure behavior can be generated; however, an FMEA report is required to summarize the abnormal behavior of each failure in terms of the impact on the system functions. To automate this requires the system functions to be provided in a machine-readable form. The next section describes briefly

1034

N.A. Snooke

one approach to this known as the functional interpretation language (Bell et al. 2007) that was developed specifically for behavioral interpretation and abstraction.

42.4

Functional Description

Functional reasoning is a term that includes a number of knowledge representation techniques aimed at capturing the purpose or intent of an engineered system. Functional modeling has been in use for a number of years, and there are a variety of methods (see Far and Halim Elamy 2005; van Wie et al. 2005 for a comparative summary), both for deriving the behavior of a system from knowledge of its structure and component function and also for interpreting system behavior. Interpretation and verification tasks benefit particularly well from inclusion of relatively small amounts of functional information. Functional modeling is an important aspect of techniques such as MADe (PHM Technology) where a standardized taxonomy of functions is used together with a structured framework for modeling system functional dependencies. MADe is used to model systems in order to identify and report on potential functional, monitoring, and maintenance issues in system design. This enables failure analysis to be conducted concurrently with system design using fuzzy cognitive mapping and bond graph simulations. In the case of FMEA, a simple functional model is able to interpret many hundreds of qualitative variable values into the state of a set of system functions. One approach considers functions as having four states as shown in Table 42.1. The function state is determined from the state presence or absence of the function triggers and expected effects. An example functional model fragment is shown in Fig. 42.3 for an aircraft fuel transfer function. The functions are linked to the simulation via the triggers and effects produced from the simulation. In this example “TVL TR” are valve positions and “OC WT” are fuel tanks used in the trigger and effect expressions, respectively. This functional interpretation language allows hierarchies of functions to be formally specified and relates the function to its purpose within a deployment environment (larger system) and identifies the behavior (triggers and effects) that determines the state of the function. In this example the behavior is linked to a qualitative simulation that returns values such as “FORWARD” and “REVERSE” for flow direction and “FUEL” or “AIR” as the substance present in pipes. This

Table 42.1 Function states Trigger False (absent) False (absent) True (present) True (present)

Effect False (absent) True (present) False (absent) True (present)

Function state Inoperative Spontaneous (unexpected effect) Failed Achieved

42 Automated Failure Effect Analysis for PHM of UAV

1035

FUNCTION wing_transfer_to_left { ACHIEVES transfer_fuel_from_right_to_left_wing BY switch_on_transfer_pump AND set_transfer_valves_R_to_L TRIGGERS transfer_into_left_tank} PURPOSE transfer_fuel_from_right_to_left_wing { DESCRIPTION 'transfer fuel from right wing tank to left wing tank' FAILURE_CONSEQUENCE 'cannot balance aircraft heavy on right' SEVERITY 5 DETECTABILITY 4} TRIGGER TVL_TR_LH.position == 'crossover' AND TVL_TR_RH.position == 'normal' IMPLEMENTS set_transfer_valves_R_to_L TRIGGER CP_TR.Control == 'on' IMPLEMENTS switch_on_transfer_pump EFFECT OC_WT_LH.transfer.FLOWDIRECTION == 'REVERSE' AND OC_WT_LH.transfer.FLOW == 'NORMAL' AND 'FUEL' IN OC_WT_LH.transfer_port.SUBSTANCE IMPLEMENTS transfer_into_left_tank UNEXPECTED_CONSEQUENCE 'aircraft imbalance - fuel transferred to left wing tank when not expected' SEVERITY 6 DETECTABILITY 8

Fig. 42.3 FIL language-based functional model description fragment

functional language is tailored for FMEA and captures risk associated with function failure as severity and detectability using the interpretation required by the engineer.

42.4.1 Generating FMEA Results FMEA unlike fault tree analysis (FTA) (Vesely 2011) normally considers the effects of all possible individual component faults. While it is possible to consider multiple faults in an automated FMEA analysis, the effort involved precludes this in a manual analysis. In an automated FMEA, the number of concurrent faults considered is limited only by the simulation time available since it grows exponentially with the number of faults. One approach is to limit multiple fault analysis to combinations of faults with combined component failure likelihood above some threshold. This will have the effect of considering only combinations of less reliable components. The increase in output is generally not as great as the increase in simulation time since many combined faults will produce effects no worse than the individual failures or the combined effects of individual failures and thus add nothing extra to the FMEA output. It is only combinations that lead to effects that are not part of the effects of the individual faults that require additional reporting. An FMEA report can be automatically generated by comparing the results from a simulation of nominal system behavior with the results of a simulation of a version of the system for each possible component fault. The differences between the two simulations comprise the effect of the component failure (e.g., a function occurs at

1036

N.A. Snooke

some point in the simulation of nominal system behavior, but fails to occur in the same state when there is a specific component failure). More details of this technique are given in Price et al. (2006a). The system needs to be simulated in its different operating modes in order to identify the possible effects of a failure. That means that it is necessary to decide how the system should be exercised during simulation in order to cover all operating modes. Normally an engineer will provide a set of inputs that cause the system to enter all of its operating modes and configurations. For example, in an aircraft fuel system such as the example in Fig. 42.4, the valve positions may be changed to encounter normal fuel feed, fuel transfer operations, and fuel cross feed configurations. The set of inputs chosen is referred to as the scenario. On occasion an engineer may decide to only deal with a subset of system operation, and this will limit the effects that will be observed to those observable during the selected operations. For simplicity the fuel supply system fragment in Fig. 42.5 will be used to illustrate the FMEA generation. A fuel tank is used to supply an engine via several pipes using a single pump. A flow sensor, a pressure sensor, and an isolation value are present in addition to a three-way valve that is not used in this system fragment. The FMEA generation system operates in the following way. The system is simulated for the chosen scenario with no failures present, and all observable values are recorded for each step of the simulation. The system model is reconfigured for each component failure. Each step of a failure simulation is compared with the nonfailure simulation and the differences are recorded. The automatically generated FMEA provides a description of the observable effects of each fault. These effects are consistent, providing all potential effects in qualitative terms. Figure 42.6 shows a fragment of an FMEA generated from the example in Fig. 42.5. Due to the qualitative nature of the analysis, the results are naturally presented as qualitative differences (such as value higher than expected). The FMEA results are ordered by significance and are inspected by engineers to identify how significant failures can be mitigated, either by decreasing the effect of the failure (e.g., by providing backup systems) or improving the detection of the failure (e.g., by adding sensors) or reducing the occurrence of the failure (e.g., by using more robust components). Risk priority numbers (RPN) can be generated for each failure item if required for a criticality analysis and comprise the product of severity, detectability, and occurrence. The likelihood of the failure is stored as part of the component failure model, and the detectability and severity are associated with the failure of function and are therefore part of the functional model as can be seen in Fig. 42.3. The FMEA report provides a consistent report that covers all component failures of specific types, and this consistency and coverage provides a basis for generating diagnostics that will be explored in the next section. The selection of the scenario is an area where the engineer can have an influence on the FMEA. The aim is to exercise all of the system functions that are of interest for the faults to be considered. If the behavior associated with a function is not exercised, any fault in the components that are involved in producing the behavior

42 Automated Failure Effect Analysis for PHM of UAV

Generic fuel system schematic

Fig. 42.4 Fuel system example

1037

1038

N.A. Snooke PT_FL_LH

FT_FL_LH

IV_FL_LH ENGINE Pipe7

Pipe6

Pipe5

Pipe4

Pipe2

TJ_FL_LH

Pipe3 CP_FL_LH

OC_WT_LH

Pipe1

V_FL_LH

Fig. 42.5 Example fuel system fragment

will not produce effects, and the report will not be able to include that component fault as a cause of the function failing. For most systems it is only necessary to exercise each function individually since the worst effects of component faults are already present. It is possible that specific combinations of function activations or function operating modes could cause faults to have an effect that is worse (higher risk function failures) than the individual function activations, and in these cases the relevant function states need to be included in the FMEA. Including additional function combinations in the scenario has no disadvantage other than increased execution time, since the results will only need to include those that produce additional effects. One strategy sometimes used is to exercise individual functions and then “all functions” together if this is possible or largest physically possible subset if not. Systems that contain backup and/or redundant functions need to be identified in the functional model with the relevant triggering conditions and the FMEA and functional model configured accordingly; the FMEA will then note that the backup function has been activated when a relevant fault occurs. A selective multiple fault analyses would be needed to determine the effects of faults in both the primary and nonredundant systems. Functional models such as (Bell 2006) identify fundamental meta-function types such as warning/telltale, fault mitigating, interlocking, and recharging, providing strong indications that the scenario and components involving these functions need particular consideration.

42.5

Sensor Selection and Diagnosability Analysis

Diagnosis is a fertile application area for artificial intelligence methodologies. Techniques range from rule-based expert systems and case-based reasoning to black-box approaches such as neural networks and have been applied to a wide variety of applications, for example, Trav´e-Massuy`es and Milne (1997). This section will focus on model-based techniques because unlike black-box techniques

42 Automated Failure Effect Analysis for PHM of UAV

1039

Fig. 42.6 FMEA fragment, for example, system

they can explain how a result was obtained and unlike rule-based expert systems or case-based techniques, model-based method do not require large amounts of historical data, training sets, or rules to be obtained from engineers. The traditional consistency-based diagnosis approach is outlined in Fig. 42.7 and is flexible and generic with widely used tools available (Forbus 1990; Forbus and de Kleer 1993; Kuipers 1994; Struss and Price 2003); however, these approaches do require that the model executed in real time, probably onboard the aircraft, to enable the discrepancy in actual and predicted behavior to be used to detect faults. This approach does not necessarily require fault models for components: however, this can lead to the possibility of the production of physically impossible

1040

N.A. Snooke Inputs

Design Model of the system

Actual System

Textbook First principles (physics)

Observed behaviour

Predicted behaviour

Diagnosis

Fig. 42.7 Model-based diagnosis

diagnoses; for example, the classic example of three lamps in parallel with a battery, in the case where two lamps fail as well as a correct diagnoses, the possibility that the battery has failed and the lamp has failed (produces light when no power is applied) is produced. The kinds of problems stem from only having models of correct behavior, since the diagnostic system has no knowledge of how a component can fail. Qualitative models significantly reduce the amount of work required to specify fault models of components with most faults being modeled as simple changes to structural or behavioral elements. Automated FMEA relies on having fault models available for all component failure modes of interest. Onboard model execution and certification concerns have led to a different approach to model-based diagnosis based on automated FMEA results that uses offline models and simulation to produce sets of diagnostic rules that can be used as the basis of an onboard system. The rule set can then be used to perform diagnosability analysis. Automated FMEA data provides a great deal of diagnostic data. A comprehensive set of possible component faults is included, and the resulting effects for wide coverage of the system functional states will be available. Furthermore, and unlike a manually generated FMEA, an automated FMEA is produced by simulation and will provide accurate, consistent, and detailed effects for each fault. The results can therefore be processed automatically to produce diagnostic rules (symptoms) for each fault. In addition, the automated FMEA has been validated by an engineer providing higher confidence that the failure modes and causes are reasonable. Once a set of diagnostic rules is produced, it can be tuned and combined with other techniques such as Bayesian networks (BN) (Ben-Gal 2007) to produce an efficient onboard system that is able to calculate the probability a fault will exist given a set of symptoms and thus provide ranked sets of potential faults. A BN (Fig. 42.8) is able to accommodate an incomplete set of diagnostic rules

42 Automated Failure Effect Analysis for PHM of UAV

Faults Probability of observing 1 symptom given fault

F1

1041

F2

F3

F4

0.5

Symptoms

S1

S2

S3

S4

S5

Fig. 42.8 Bayesian diagnostic network

or uncertainty in the observations, thus making the system less sensitive to the numerical thresholds set for qualitative values and also to the states produced the FMEA scenario. Due to its statistical nature, a BN can operate with whatever symptom and fault data is available, for example, symptoms produced by hand from manual FMEA or symptoms that are incomplete or contain contradictory evidence. Other than thresholds, the network uses three basic numerical measures: Symptom leak – probability that the symptom will be observed even though there are no faults. By default this is set to 0 for symptoms generated from an automated FMEA symptoms to indicate the symptom will not be observed unless a relevant fault exists. It may be decided that certain observations are more likely to be spurious than others due to sensor or system characteristics. Prior – probability that the fault has occurred prior to any symptoms. This is the fault occurrence number in FMEA (or 0.01 default) Fault – probability that the symptom will be observed given the fault. By default this is set to 1 for the auto FMEA symptoms to indicate the symptom will be observed if the fault exists. Other values may be used for some symptom or fault categories because the effects predicted by the FMEA are less or more significant than can be determined qualitatively. For example, leaks may not always produce a measurable pressure drop, even though theoretically any leak would produce some drop in pressure. Engineering experience can also be captured and added into the model over time. For example, if a valve was found to make a characteristic sound (which can be detected), then this symptom can be added and conditional probabilities added based on how often the sound was reported for each fault. When one or more symptoms are observed, these values are propagated through the network to produce an ordered list of possible faults. A confidence value is associated with each predicted fault allowing decisions to be made within the higherlevel control elements of the system.

42.5.1 Generating Symptoms from an FMEA The automated FMEA contains a lot of diagnostic information in the form of component fault – effect relationships. In addition these relationships are validated

1042

N.A. Snooke

once an engineer has considered the FMEA. For convenience the symptoms are categorized into three types, although algorithms that generate the symptoms generate whichever type is necessary based on the number of observations needed and to ensure symptoms are able to exonerate faults if required: • Single value, for example, pressure transducer value = out of range low. In this situation a symptom is a simple value that does not occur during normal operation and directly indicates one or more faults. • Multiple value, for example, level = empty AND low level switch = on. Two values may be linked to indicate an inconsistency in system operation caused by a fault. Typically, this is because two values measure physically related properties and should be consistent in the absence of a fault. • Conditional IF (pump = on), pressure monitor = high or IF (Valve SOV4 commanded OPEN), Open Sensor is not responding OPEN. Conditional symptoms may be used when a value is only indicative of a fault in specific operating modes or configuration. For example, a low-pressure reading might only be significant when a pumping operation in progress. If no pumping is carried out, then the symptom does not apply (pressure will be expected to be low). The condition specifies if the observation is valid. For the diagnostic net (discussed in following paragraph), the distinction between multiple value and conditional symptoms is important because the symptom observations are not entered to the net for an invalid symptom. During diagnosis symptoms with a satisfied condition are valid, and valid symptoms may either indicate a fault if the overall symptom is satisfied or exonerate a fault if the overall symptom is false. Invalid symptoms are not applicable in the state or operating mode being considered. The FMEA report provides differences of observable values, for example, when PIPE5 is blocked, PT FL LH pressure is below normal when normal was expected. This type of information can be used to produce the kind of fault-symptom information needed by the online diagnostic system. The symptoms should be as simple as possible while still providing a definite indication of the associated fault. The FMEA produces output in the form of the difference between nominal and failure operation at each step of a scenario, exercising the system. An example may be After switch X closed and switch Y closed, Obs A=L when Obs 1=H expected. The FMEA information does not directly provide a symptom because it is usually not known explicitly what is expected. The symptoms must therefore be based only on identification of abnormal observations. The FMEA item above plus additional simulation data generated during the FMEA production is used to produce a symptom of the form: condition = f(X, closed), (Y, closed), (Obs B, Active)g; observation = f(Obs B, L)g; faults = fPipe1 Blocked, Pipe2 blockedg. To produce symptoms, additional nominal observations are often required such as the value of Obs B in the example. Valid symptoms must satisfy two constraints. Firstly, a symptom should only be present when the associated fault is present. Secondly, it is desirable to detect as many faults as possible in as many operating states as possible. The first constraint requires that any potential symptom be identified such that it is not satisfied in any

42 Automated Failure Effect Analysis for PHM of UAV

1043

nominal operating mode; however, it can be satisfied for other faults if general symptoms are required. There are efficient set theory algorithms for performing this kind of search; one specific example used for this application can be found in Snooke and Price (2012). The second constraint implies symptoms should be general and therefore contain the fewest measurements, such that the first constraint is satisfied. Simply using symptoms that include every available observable value produced from the component failure simulations will create a very limited diagnostic system because faults will only be diagnosable in the states contained in the FMEA scenario. The simulations carried out to produce an automated FMEA provide a subset of the possible failure and nominal system states. To produce symptoms that diagnose faults over the majority of system states requires generalization of the effects of faults. Such generalization is achieved by reducing the number of observations in the symptom, in particular by excluding those that are not relevant to the fault being diagnosed.

42.5.2 Associating Measurement to Functions The functional model (also used to interpret the FMEA) is the enabling concept that indirectly provides abstract-guiding knowledge-facilitating extrapolation of symptoms to system states that were not present in the FMEA. Measurements are associated with system functions based on the achievement and failure of functions across the entire FMEA. Given a comprehensive set of failures, this provides a simulation-derived description of which functions each component/measurement affects and is used to indicate a measurement has some behavioral or structural “causal” relationship to a function based on evidence of failure effect from the FMEA. Often these will simply be different value measurements from the same sensor: however, this is not assumed. By associating these abnormal measurements with functions that simultaneously fail, a mapping of function-associated measurements is created. Even measurements structurally adjacent to external triggers will be affected by faults in the connecting components. For example, switch.position ==on is a trigger; however, the current flow though the switch.contact will be affected by faults such as the contact being stuck or wiring faults in the switch circuit and will be associated with any functions that are affected by any of the switches failure modes. This produces a set of measurements that agree with engineering expectation of all the components used within the implementation each function. However, even if this is not clear, there must be a structural or behavioral relationship between the function and associated measurements, as required for measurement selection. A measurement-function mapping in turn allows an assessment of the relevance of measurements associated with a fault; measurements associated with functions that are never affected during the exercising of the fault (which should include at least all individual system functions) can be considered irrelevant and are removed from the set of measurements to be used as candidates for any symptom that will

1044

N.A. Snooke

diagnose the fault. This prevents symptoms being generated that rely on spurious correlations between measurements and faults due to the incomplete coverage of the system state space during the FMEA.

42.5.3 Fault Exoneration Diagnostic symptoms may be required to be negated to perform fault exoneration – a symptom expression evaluating to false needs to indicate fault absence. For example, consider a lamp with a plausible symptom: lamp == inactive AND switch == on implies faults flamp blown, switch dirty, lamp wire fracturedg If the switch is off, then switch == on is false resulting in a false symptom expression, but this does not imply the lamp is OK – it could be blown. Manually crafted symptoms are often conditional so that symptom absence will exonerate associated faults. This is because engineers consider the implication of not seeing the symptom as well as its presence. In some applications such as a BN-based diagnostic system may require that symptoms exonerate faults. This also allows better fault detection by allowing evidence from nominally operating parts of a system to contradict evidence from very general symptoms, hence reducing the set of possible faults. That is, a symptom expression must not exonerate a fault for an observation where the fault exists and to achieve this requires additional checks in the symptom generation algorithm. Fault exoneration requires that the algorithms include additional features to ensure that observations are included in the conditional part of the symptom as necessary to ensure the symptom is only valid in operating states such that the symptom can be negated.

42.5.4 FMEA Coverage and Symptom Generation The coverage provided by the FMEA can have an influence on the selection measurement strategy required to generate symptoms. For example, consider Fig. 42.9. Assume the functions A and B were not exercised simultaneously during the FMEA. There will be an entry in the FMEA for fault w2 fracture when the switch sw1 is closed providing the effect function A failed (function B is inoperative). One of the symptoms generated from this simulation requires the state of w1 and sw1. Wire w1 is allowed because it is associated with Function A (as are all the other wires on the path through sw1 and l1). The symptom generation search algorithm determines that state of the switch is included in the conditional part of the symptom because if the switch is not closed, electrical activity in w1 can say nothing about the fault in w2 (it depends on the state of Function B). The symptom correctly predicts the fault w2 fracture when function A only is active but falsely exonerates the fault when both functions are triggered because w1 is active when the condition is true but the symptom is false. Using a BN the probabilities could be adjusted to reflect this;

42 Automated Failure Effect Analysis for PHM of UAV

1045

Function A: sw1=closed triggers l1=active Function B: sw2=closed triggers l2=active

sw2

Vcc

l2

gnd

w1 w2

sw1

l1

Symptom when sw1=closed: w1=inactive indicates w2.fracture Symptom generated if A and B not exercised concurrently in scenario. Will provide false exoneration of w2 when both switches closed

Fig. 42.9 Exoneration of faults

however, more accurate consideration of the operating modes of the system will lead to more accurate diagnoses (higher confidences in the predicted fault). One approach to achieving this is to exercise both of the functions simultaneously in the FMEA scenario, forcing the symptom generation to include additional measurements in the symptom, for example, the position of switch sw2. Thus by extending the scenario to close both switches and exercise both functions simultaneously will produce the symptom when sw1 closed and sw2 open: w1 inactive indicates w2 fracture to be generated since A is failed and B is active, allowing w1 to be used. There may be many possible permutations of functions and particularly function states available making exercising of all in the FMEA an unattractive proposition. In addition it is necessary to detect any spurious symptoms, although this can be done by further simulation. An alternative is to disallow measurements associated with several functions if any of the functions are inactive or unexpected. For the w2 fault, this will remove w1 as a candidate measurement for the symptom since w1 is associated with the inoperative function B. Thus, w1 will not be used in any symptoms generated for Fig. 42.9 unless both switches are closed in the FMEA scenario. The symptom when sw1 closed: l1 inactive indicates w2 fracture and similar for the other wires on that branch will be generated. Notice that both these revised symptoms apply to more specific operating conditions than the original and are only valid (usable) when function B is not activated. The symptom can only exonerate w2 when function B is not activated (sw 2 open) and function A is activated (sw1 closed) and w1 is not inactive. Measurements of w1are less useful than measurements of l1 in diagnosing w2 fractured. The measurement selection technique described addresses the incomplete set of low-level operating states inherent in the FMEA and ensures that irrelevant measurements are not included in symptoms, and hence that symptoms are extrapolated

1046

N.A. Snooke

Fig. 42.10 Example symptoms

to be as general as possible. Conversely the decision as to which measurements are relevant is derived from the entire set of failure and non-failure states encountered during the FMEA. Given a reasonably comprehensive FMEA that exercises all system functions and includes faults for the majority of components, symptoms can be produced that are general enough to cover a great deal of the system state that is not explicitly included in the FMEA without producing spurious symptoms (Snooke and Price 2012). Analysis of the actual coverage of the derived symptom set is covered in a later section on diagnosability assessment. Since a diagnostic system based on the symptoms will use positive and negative evidence of each failure to derive a final likelihood of each fault, a “complete” set of symptoms is not necessary, and probabilistic techniques such as BN allow for a degree of uncertainty in the symptoms. Figure 42.10 shows an example set of symptoms generated from the simple example system shown in Fig. 42.5. Most of these symptoms are single abnormal values; however, notice that several conditional symptoms have been generated (S11–S14) to allow values that occur when the system is inactive to become symptoms when the operating conditions change. For more complex systems, the combinations of values and conditions required become greater and provide symptoms that detect unusual effects and are particularly useful when a system has many operating modes with symptoms specific to each. Once the set of symptoms have been generated, the FMEA is used to produce a matrix of fault-symptom pairs. An example fragment is shown in Fig. 42.11,

42 Automated Failure Effect Analysis for PHM of UAV

1047

Fig. 42.11 Fragment of an example fault-symptom mapping

example fault-symptom mapping, showing fault number, fault name, symptom number, symptom name, and (default) confidence value. In a real system, there are likely to be one or more faults associated with a symptom, and each fault may be associated with one or more symptoms.

42.6

Diagnosability Assessment

Ideally, any system fault could be diagnosed to the individual component that has failed; however, during design, there is a trade-off between the amount of sensing possible for a system and the diagnostic capability. It is useful to be able to investigate the relationship between the number and placement of sensors and the resultant ability to detect faults and subsequently isolate faults to a specific component or line replaceable unit (LRU). One of the benefits of model-based simulation is easy access to many system parameters, allowing analysis of those that might provide the required diagnosability at a given cost. Alternatively the question might be reversed to analyze how many sensors would be required to be able to provide a given level of diagnosis. On the ASTRAEA project, two types of analysis were performed; the first provides an ordered list of sensors that can provide the maximum number of diagnosed faults, and the second provides an ordered list prioritized by the ability to isolate a fault. In practice a whole variety of compromises must be made, and these “what if” analyses are used to inform higher-level engineering decisions. Symptom ranking is carried out using a recursive procedure starting with no included symptoms or detected faults and consists of the following steps: 1. From the remaining symptoms that have not been considered, find the symptom(s) that provide the maximum number of additional faults, that is, the number of faults detected by the symptom that are not already located by previously considered symptoms. These are termed the “next best symptoms” 2. Add each of the “next best symptoms” in turn to the included symptoms and also add any failures it detects to the faults detected list. 3. If the symptom has not already been found and included in the overall results, then include it in the results in the correct place (according to the number of

1048

N.A. Snooke

symptoms) and carry out the procedure again excluding the current symptom from the available symptoms. This will generate sets of symptoms for each number of observations that can diagnose the most faults. For further information and examples of the software tools used, see (Snooke 2009; Snooke and Price 2012). Where several symptoms each diagnose the same number of faults, they are all outputted. For the example circuit in Fig. 42.1, the following summary is produced: Ordered symptom information 1 combinations of 1 SYMPTOMS indicate 11 FAILURES (1 partitions) 1 combinations of 2 SYMPTOMS indicate 18 FAILURES (2 partitions) 1 combinations of 3 SYMPTOMS indicate 24 FAILURES (3 partitions) 1 combinations of 4 SYMPTOMS indicate 28 FAILURES (5 partitions) 8 combinations of 5 SYMPTOMS indicate 28 FAILURES (6--7 partitions) 28 combinations of 6 SYMPTOMS indicate 28 FAILURES (7--9 partitions) 56 combinations of 7 SYMPTOMS indicate 28 FAILURES (8--10 partitions) 70 combinations of 8 SYMPTOMS indicate 28 FAILURES (8--11 partitions) 56 combinations of 9 SYMPTOMS indicate 28 FAILURES (9--11 partitions) 28 combinations of 10 SYMPTOMS indicate 28 FAILURES (10--12 partitions) 8 combinations of 11 SYMPTOMS indicate 28 FAILURES (11--12 partitions)

The first of these provides: ENGINE.engine{\_}fuel{\_}feed = none AND CP{\_}FL{\_}LH.Control = on INDICATES FAULTS: Pipe5.blocked; Pipe5.fracture; Pipe2.blocked; Pipe3.blocked; Pipe1. blocked; Pipe6.blocked; Pipe6.fracture; Pipe7.blocked; Pipe7.fracture; Pipe4. blocked; Pipe4.fracture

The second entry considers the best pair of sensors: ENGINE.engine{\_}fuel{\_}feed = none AND CP{\_}FL{\_}LH.Control = on INDICATES FAULTS: Pipe5.blocked; Pipe5.fracture; Pipe2.blocked; Pipe3.blocked; Pipe1. blocked; Pipe6.blocked; Pipe6.fracture; Pipe7.blocked; Pipe7.fracture; Pipe4. blocked; Pipe4.fracture ENGINE.engine{\_}fuel{\_}feed = low fuel OC{\_}WT{\_}LH.tank{\_}level = higher than expected FT{\_}FL{\_}LH.flow = low INDICATES FAULTS: Pipe5.partialblocked; Pipe2.partialblocked; Pipe3.partialblocked; Pipe1.partialblocked; Pipe6.partialblocked; Pipe7.partialblocked; Pipe4.partialblocked

In this case, there are three symptoms that all indicate the same set of failures and any one could be used. Sometimes there may be several sets of (nonequivalent) symptoms able to discriminate the same number of faults. In the example above, this occurs for 5 symptoms, where 8 different combinations of symptoms indicate 28 faults (though not necessarily the same 28 faults). They may have different

42 Automated Failure Effect Analysis for PHM of UAV

1049

fault-isolating capability dividing the faults into six or seven partitions provided by different combinations of the chosen symptom set. These partitions each contain a set of one or more faults that are indistinguishable using the selected set of observations used by the selected symptoms. From the above table, it is clear that 4 symptoms are adequate to identify all 28 faults present in the FMEA of the system: ENGINE.engine{\_}fuel{\_}feed = none AND CP{\_}FL{\_}LH.Control = on INDICATES FAULTS: Pipe5.blocked; Pipe2.blocked; Pipe3.blocked; Pipe1.blocked; Pipe6.blocked; Pipe7.blocked; Pipe4.blocked ENGINE.engine{\_}fuel{\_}feed = low fuel OC{\_}WT{\_}LH.tank{\_}level = higher than expected FT{\_}FL{\_}LH.flow = low INDICATES FAULTS: Pipe5.partialblocked; Pipe2.partialblocked; Pipe3. partialblocked; Pipe1.partialblocked; Pipe6.partialblocked; Pipe7.partialblocked; Pipe4.partialblocked ENGINE.engine{\_}fuel{\_}feed = air INDICATES FAULTS: Pipe2.fracture; Pipe2.leak; Pipe3.fracture; Pipe3.leak; Pipe1. fracture; Pipe1.leak ENGINE.engine{\_}fuel{\_}feed = none AND CP{\_}FL{\_}LH.Control = on OC{\_}WT{\_}LH.tank{\_}level = lower than expected INDICATES FAULTS: Pipe5.fracture; Pipe6.fracture; Pipe7.fracture; Pipe4.fracture OC{\_}WT{\_}LH.tank{\_}level = lower than expected INDICATES FAULTS: Pipe5.leak; Pipe6.leak; Pipe7.leak; Pipe4.leak

This analysis demonstrates that the fuel feed detected by the engine is the most important available diagnostic indicator, followed by an incorrect fuel level in the tank. The flow and pressure sensors are not even necessary to detect a full set of faults. However, the engineer could decide that the engine fuel feed is not a feasible observation by using knowledge outside the scope of the modeling and simulation; removing it and running the analysis again, the flow meter then becomes an important sensor. It is also possible to maximize the fault isolation capability using a modified version of the symptom-ranking algorithm that maximizes the number of fault partitions instead of simply the number of faults. Selecting the most useful measurements leads to 7 symptoms that can identify all 28 faults and divides them into 10 separate sets of (indistinguishable) faults. Maximizing for fault isolation leads to 12 partitions of faults using various combinations of the following 12 symptoms: (Pipe2.partialblocked Pipe3.partialblocked Pipe1.partialblocked Pipe6.partialblocked Pipe7.partialblocked) (Pipe5.partialblocked Pipe4.partialblocked) (Pipe4.leak) (Pipe5.leak Pipe6.leak Pipe7.leak) (Pipe2.leak Pipe3.leak Pipe1.leak) (Pipe2.fracture Pipe3.fracture Pipe1.fracture) (Pipe5.blocked Pipe4.blocked) (Pipe6.blocked Pipe7. blocked) (Pipe6.fracture Pipe7.fracture) (Pipe5.fracture) (Pipe2.blocked Pipe3.blocked Pipe1.blocked) (Pipe4.fracture)

1050

N.A. Snooke

This in fact includes all the symptoms in Fig. 42.4 because S2, S3, and S5 turn out to be identical as mentioned above. The above paragraphs assume that all faults are equally important in terms of diagnosis. In practice, there are several additional considerations. Often there is a limit to the granularity required by fault isolation due to the presence of LRUs. There is no requirement to be able to isolate a fault beyond a single LRU. Observations may also fall into categories including basic system sensing that must be available for nominal operation and groups of additional observations that only make sense to provide as single units (sensors). Including these additional factors in the sensor selection allows further structuring of the symptoms, observations, and measures used to select symptom permutations.

42.7

Conclusion

This chapter is concerned with methods and technologies needed to be able to routinely fly UAVs safely in commercial airspace. The techniques described in this chapter contribute to this goal in several ways: • They automate the generation of failure effects for an FMEA report, providing consistent results for a complete set of component failure modes. It can be guaranteed that results are produced for all component failure modes that are modeled. • The generated FMEA results can be arranged as failure-symptom pairs with the same consistency and can be linked to specific symptoms that are observable by an online system. • The failure-symptom pairs have been integrated into a larger diagnostic system, where other failure-symptom pairs will have been produced by other methods. • Further analysis can indicate the most effective points within the system being diagnosed to place sensors, assisting in the decision of where sensors should be placed when designing the system. A number of existing modeling and analysis techniques can be integrated to provide models and design analysis that allow diagnosability and diagnosis to be designed in rather than bolted on, and this can form the basis of a lightweight verifiable onboard diagnostic system. Both of these aspects are essential when comprehensive system health information is required to enable mission level objectives to be decided in an automated environment.

References Astraea (2009), http://www.astraea.aero/. Accessed 20 Nov 2011 Bell, Interpretation of simulation for model based design analysis of engineered systems. Ph.D. Thesis, University of Wales Aberystwyth, 2006. http://cadair.aber.ac.uk/dspace/handle/2160/ 177 J. Bell, N. Snooke, C.J. Price, A language for functional interpretation of model based simulation. Adv. Eng. Inform. 21(4), 398–409 (2007)

42 Automated Failure Effect Analysis for PHM of UAV

1051

I. Ben-Gal, in Encyclopedia of Statistics in Quality and Reliability, ed. by F. Ruggeri, R.S Kennett, F.W Faltin (Wiley, Chichester, 2007). ISBN 978-0-470-01861-3 D. Bobrow (ed.), Special issue on qualitative reasoning about physical systems. Artif. Intell. 24, 1–5 (1984) W. Borutzky, Bond Graph Methodology (Springer, New York, 2010). ISBN 978-1-84882-881-0 J.S. Brown, R. Burton, J. de Kleer, Pedagogical and knowledge engineering techniques in SOPHIE I, II and III, in Intelligent Tutoring Systems, ed. by D. Sleeman, J.S. Brown (Academic, New York, 1982), pp. 227–282 J. de Kleer, Multiple representations of knowledge in a mechanics problem-solver, in Proceedings of the IJCAI-77 (Morgan Kaufmann, Los Altos, 1977), pp. 299–304 B.H. Far, A. Halim Elamy, Functional reasoning theories: problems and perspectives. Artif. Intell. Eng. Des. Anal. Manuf. 19(2), 75–88 (2005). Publisher Cambridge University Press, New York. ISSN: 0890–0604 K. Forbus, The qualitative process engine, in Readings in Qualitative Reasoning About Physical Systems, ed. by D. Weld, J. de Kleer (Morgan Kaufmann, San Mateo, 1990), pp. 220–235 K.D. Forbus, J. de Kleer, Building Problem Solvers (MIT, Cambridge, 1993). ISBN 978-0-26206157-5 W.C. Hamscher, J. de Kleer, L. Console (eds.) Readings in Model-Based Diagnosis. (Morgan Kaufmann, San Mateo, 1992) D. Harel, M. Politi, Modeling Reactive Systems with Statecharts: The STATEMATE Approach (McGraw-Hill, New York, 1998). ISBN 978-0070262058. Out of print. Downloadable from http://www.wisdom.weizmann.ac.il/harel/books.html B.J. Kuipers, Qualitative simulation. Artif. Intell. 29, 289–338 (1986) B. Kuipers, Qualitative Reasoning – Modelling and Simulation with Incomplete Knowledge (MIT, Cambridge, 1994). ISBN 978-0-262-11190-4 M.H. Lee, Qualitative circuit models in failure analysis reasoning. Artif. Intell. 111, 239–276 (1999) G.F. Luger, Artificial Intelligence – Structures and Strategies for Complex Problem Solving, 4th edn. (Addison Wesley, Reading, 2002). ISBN 0-201-64866-0 A. Mukherjee, R. Karmakar, Modeling and Simulation of Engineering Systems Through Bondgraphs (CRC Press LLC/N.W. Corporate, Boca Raton, 1999, 2000). ISBN 978-0849309823 B. Peischl, F. Wotawa, Model-based diagnosis or reasoning from first principles. IEEE Intell. Syst. 18, 32–37 (2003) PHM Technology, http://www.phmtechnology.com/. Accessed 18 Nov 2011 C.J. Price, Function directed electrical design analysis. Artif. Intell. Eng. 12(4), 445–456 (1998) C.J. Price, N.A. Snooke, S.D. Lewis, A layered approach to automated electrical safety analysis in automotive environments. Comput. Ind. 57, 451–461 (2006a) C. Price, L. Trav´e-Massuy`es, R. Milne, L. Ironi, K. Forbus, B. Bredeweg, M. Lee, P. Struss, N. Snooke, P. Lucas, M. Cavazza, G. Coghill, Qualitative futures. Knowl. Eng. Rev. 21(4), 317–334 (2006b). Cambridge University Press N. Snooke, An automated failure modes and effects analysis based visual matrix approach to sensor selection and diagnosability assessment, in Proceedings of the Prognostics and Health Management Conference (PHM09), SanDiego, CA, Sept 2009 N. Snooke, C.J. Price, Automated FMEA Based Diagnostic Symptom Generation, Proc. Advanced Engineering Informatics, AEI 2012, 26, 870–888. doi 10.1016/j.aei.2012.07.001 N. Snooke, C.J. Price, An effective practical model-based detectability tool, Proceedings of the DX-11, Murnau, Germany, 2011, pp. 180–187 P. Struss, C.J. Price, Model-based systems in the automotive industry. AI Mag. 24(4), 17–34 (2003) The Unified Modelling Language, http://www.uml.org/. Accessed 29 Nov 2011 L. Trav´e-Massuy`es, R. Milne, TIGERTM: gas turbines condition monitoring using qualitative model based diagnosis. IEEE Expert Intel. Syst. Appl. 12(3), 21–31 (1997) M. van Wie, C.R. Bryant, M.R. Bohm, D.A. McAdams, R.B. Stone, A model of function-based representations. Artif. Intell. Eng. Des. Anal. Manuf. 19, 89–111 (2005) W. Vesely, Fault Tree Handbook with Aerospace Applications, NASA, http://www.hq.nasa.gov/ office/codeq/doctree/fthb.pdf. Accessed 3 Sept 2011

Prognostics Applied to Electric Propulsion UAV

43

Kai Goebel and Bhaskar Saha

Contents 43.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1054 43.1.1 Prognostics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1054 43.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1055 43.2.1 Nominal Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1056 43.2.2 Damage Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1056 43.3 Prognostics Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1057 43.3.1 Model-Based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1057 43.3.2 Particle Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1057 43.3.3 Data-Driven Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1059 43.4 Case Study: Prognostics for Batteries Used in Electric UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1060 43.4.1 Battery Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1060 43.4.2 Model Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1062 43.4.3 Battery Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065 43.4.4 UAV Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065 43.4.5 Implementation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1067 43.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1069 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1069

Abstract

Health management plays an important role in operations of UAV. If there is equipment malfunction on critical components, safe operation of the UAV might possibly be compromised. A technology with particular promise in this arena is equipment prognostics. This technology provides a state assessment of the health

K. Goebel () NASA Ames Research Center, Moffett Field, CA, USA e-mail: [email protected] B. Saha Palo Alto Research Center, Palo Alto, CA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 47, © Springer Science+Business Media Dordrecht 2015

1053

1054

K. Goebel and B. Saha

of components of interest, and if a degraded state has been found, it estimates how long it will take before the equipment will reach a failure threshold, conditional on assumptions about future operating conditions and future environmental conditions. This chapter explores the technical underpinnings of how to perform prognostics and shows an implementation on the propulsion of an electric UAV. An accurate run-time battery life prediction algorithm is of critical importance to ensure the safe operation of the vehicle if one wants to maximize in-air time. Current reliability-based techniques turn out to be insufficient to manage the use of such batteries where loads vary frequently in uncertain environments. A Particle Filter is shown as the method of choice in performing state assessment and predicting future degradation. The method is then applied to the batteries that provide power to the propeller motors.

43.1

Introduction

Prognostics Health Management (PHM) is an engineering discipline that aims at maintaining nominal system behavior and function and assuring mission safety and effectiveness under off-nominal conditions. PHM encompasses a set of wideranging subdisciplines that address the design, development, operation, and life cycle management of subsystems, vehicles, and other operational systems. Starting with simple time-temperature recorder for the engine hot section on the F-8 aircraft (during deployment in Vietnam), and later the A-7 aircraft engine health monitoring program of the early 1980s, PHM principles found their way in various aircraft. Analysis of vibration and acoustic emissions data from rotorcraft drivetrains has led to breakthroughs in predicting impending failures of complex mechanical systems, resulting in the development of relative mature Health and Usage Monitoring Systems (HUMS) for rotorcraft (Revor and Bechhoefer 2004). Service providers like GE, PW, and Rolls-Royce have employed PHM principles to remotely monitor jet engines around the clock to detect early signs of damage as part of guaranteed uptime service agreements.

43.1.1 Prognostics Prognostics is a core element of PHM. It is a younger member in the family of health management techniques but has recently received considerable attention due to its game-changing potential. Prognostics is the science of determining the remaining useful life of a component or subsystem given the current degree of wear or damage, the component’s load history, and anticipated load and environmental conditions. A quantification of the degree of a component’s wear or damage and the estimate of end of life gives decision makers important information about the health of a system. This information can be used on UAV for risk reduction in go/no-go decision, cost reduction through the scheduling of maintenance as needed, and improved asset availability. Prognostics employs techniques that are

43 Prognostics Applied to Electric Propulsion UAV

1055

often based on a detailed analysis of historical data or an analysis of the fault modes and the modeling of the physics of both the component itself and the attributes that characterize the fault. For the latter, the idea is to model the progression of damage which includes the effects of damage accelerators or stressors (such as load or environmental conditions). Next, algorithms that estimate the remaining life use estimation techniques that propagate the anticipated degradation into the future and provide as output the point where the component does no longer meet its desired functionality. The algorithms use these physics-based models as well as measurements from the system as input. Output of the algorithms is the remaining life estimate. An alternative to physics-based models are data-driven techniques. These synthesize a behavioral representation from a large number of example runto-failure trajectories via machine learning methods. Prognostics can be developed for almost any critical component as long as one has either some knowledge about the underlying physics or a sufficient amount of run-to-failure data exists. The efficacy of prognostics has been demonstrated for a wide range of diverse components ranging from mechanical components to electrochemical components to electronics. Specific application areas range from rotating machinery (Marble and Tow 2006) to batteries (Saha et al. 2007), from printed circuit boards (Gu et al. 2007) to solid rocket motors (Luchinsky et al. 2008). In the U.S. military, two significant weapon platforms were designed with a prognostics capability as an integral element of the overall system architecture: the Joint Strike Fighter Program (Hess et al. 2004) and the Future Combat Systems Program (Barton 2007). Prognostic technology is also finding its way into future NASA launch vehicles and spacecraft (Osipov et al. 2007) as well as UAV (Valenti et al. 2007). As the technology matures further, it is expected that prognostics will play an important role in the design and operation of commercial systems such as passenger aircraft, automobiles, ships, the energy infrastructure, and consumer electronics.

43.2

Modeling

Underlying any prediction is a model that describes how the component of interest behaves under nominal conditions and how it will evolve as it experiences wear or a fault condition. To that end, one needs to capture that process in the form of a mathematical model. The model may be derived from laws of physics, captured by empirical relations, or learned from data. Models can also use a combination of these approaches. For example, parts of a component that are well understood may be constructed using physics-based models, with unknown physical parameters learned from data using appropriate system identification techniques. Modeling of physics can be accomplished at different levels, for example, microand macro-levels. At the microlevel, physical models are embodied by a set of dynamic equations that define relationships, at a given time or load cycle, between damage (or degradation) of a component and environmental and operational conditions under which the component is operated. The microlevel models are often referred to as damage propagation model, for example, Yu and Harris’s fatigue life

1056

K. Goebel and B. Saha

model for ball bearings, which relates the fatigue life of a bearing to the induced stress (Yu and Harris 2001), Paris and Erdogan’s crack growth model (Paris and Erdogan 1963) and stochastic defect-propagation model (Li et al. 2000) are other examples of microlevel models. Since measurements of critical damage properties (such as stress or strain of a mechanical component) are rarely available, sensed system parameters have to be used to infer the stress/strain values. Microlevel models need to account in the uncertainty management the assumptions and simplifications, which may pose significant limitations of that approach. Macro-level models are characterized by a somewhat abridged representation that makes simplifying assumptions, thus reducing the complexity of the model (typically at the expense of accuracy). An example is a lumped parameter model which assumes that the attributes of the component have idealized behavior and the nonideal characteristics are characterized with equivalent elements that suffice for a first-order approximation. When such a system is designed well, it will often (but certainly not always) result in satisfactory results, depending on performance requirements. It should be noted that for complex systems (e.g., a gas turbine engine), even a macro-level model may be rather time-consuming and labor intensive. The resulting simplifications may need to be accounted for via explicit uncertainty management.

43.2.1 Nominal Behavior The system model describes the characteristics of the system under nominal conditions. Ideally, such a model should be able to factor in the effects of operational and environmental conditions as well as any other conditions that cause different system response under nominal conditions. The model should also be able to adapt to changes of the system that are not considered abnormal. To that end, the system model could learn system behavior from examples, for instance, using machine learning techniques or it could integrate domain expertise and be implemented using rules.

43.2.2 Damage Modeling A damage propagation model describes how the damage is expected to grow in the future. It should, similar to the system model, account for operational and environmental conditions as well as any other conditions that have an impact on the damage. While one often thinks of damage as a monotonically increasing phenomenon, it is possible for the domain in which damage is evaluated to have nonmonotonic attributes. These could be either intrinsic attributes (e.g., recovery effects in batteries or power semiconductors) or extrinsic effects such as partial repair actions. Depending on the fault mode, damage propagation may exhibit different symptoms and it may be necessary to consider dedicated damage propagation models for different fault modes.

43 Prognostics Applied to Electric Propulsion UAV

43.3

1057

Prognostics Algorithms

The role of the prognostic algorithm is applying the damage propagation model into the future. It needs to ensure that it properly considers the effects of environmental and operational conditions, healing phenomena, as well as how to account for the different sources of uncertainty. Depending on the implementation, the damage propagation model and the prognostic algorithm may not be separable. For the general case, the damage propagation model and the prognostic algorithm will be treated as separate.

43.3.1 Model-Based Algorithms In order to make a prediction, the current health state of the system must first be known. Determining the current state of the system health is generally known as health state estimation. Given the current state of health, the model-based prediction algorithm propagates this estimate forward in time using damage propagation model equations up to the end-of-life (EOL) threshold. As mentioned earlier, the evolution of the state depends on future stressors. Therefore, these needs to be stated to the degree possible. For many applications, some knowledge about future stressors exists where tasks are scheduled or repeated. In other cases, one has only statistical information about future stressors. The accuracy and precision of the predictions depend on both the quality of the model and the uncertainty of future inputs. The prediction algorithm of choice depends on the type of model used, on what information is needed to describe the EOL (which could be in the form of a distribution), and the amount of computation that may be performed. This section takes a look at Particle Filters, one of the most prevalent algorithms for model-based prediction.

43.3.2 Particle Filter Particle methods assume that the state equations can be modeled as a firstorder Markov process with additive noise and conditionally independent outputs (Arulampalam et al. 2002). Let xk D fk1 .xk1 / C !k1

(43.1)

zk D hk .xk / C vk :

(43.2)

While there are several flavors of Particle Filters, the focus here is on Sampling Importance Resampling (SIR), in which the posterior filtering distribution denoted as p.xk jZk / is approximated by a set of N weighted particles fhxpi ; wip iI i D 1; : : : ; N g sampled from a distribution q.x/ that is “similar” to .x/, i.e., .x/ > 0 ) q.x/ > 0 for all x 2 Rnx . The importance weights wik are then normalized

1058

K. Goebel and B. Saha

Initialize PF Parameters Propose Initial Population , 〈x0,w0〉 Propagate Particles using State Model , xk−1→xk Measurement zk

Update Weights, wk−1→wk Weights degenerated?

No

Yes Resample

Fig. 43.1 Particle Filtering flowchart

wik

 ı    xik q xik D N P  j .  j  q xk  xk

(43.3)

j D1

such that †i wik D 1, and the posterior distribution can be approximated as p.xk jZk / 

N X

  wik ı xk  xik :

(43.4)

i D1

Using the model in Eq. (43.1), the prediction step becomes p.xk jZk1 / 

N X

  wik1 fk1 xik1 :

(43.5)

i D1

The weights are updated according to the relation w N ik

D

wik D

p wik1

 ˇ i  i ˇ i  zk ˇxk p xk ˇxk1   ˇ ; q xik ˇxik1 ; zk

wN ik : N P i wN k

(43.6) (43.7)

j D1

It is possible that all but a few of the importance weights degenerate such that they are close to zero. In that case, one has a very poor representation of the system state (and also wastes computing resources on unimportant calculations). To address that, resampling of the weights can be used (Saha et al. 2007). The basic logical flowchart is shown in Fig. 43.1.

43 Prognostics Applied to Electric Propulsion UAV Fig. 43.2 Prediction flowchart

1059

Start Prediction at tp Estimate Initial Population, 〈xp,wp 〉

Propagate Particles using State Model , xp+k–1→xp+k

EOL threshold exceeded?

No

Yes Generate RUL pdf from {wp}

During prognosis this tracking routine is run until a long-term prediction is required, say at time tpnD, at which E point Eq. (43.1) o will be used to propagate the

xpi ; wip I i D 1; : : : ; N until xi fails to meet the system i  i specifications at time tEOL . The RUL pdf, i.e., the distribution p tEOL  tp , is given by the distribution of wip . Figure 43.2 shows the flow diagram of the prediction process. posterior pdf given by

43.3.3 Data-Driven Algorithms As the name implies, data-driven techniques utilize monitored operational data related to system health. Given the availability of data, data-driven approaches are appropriate when the understanding of first principles of system operation is not easy to come by or when the system is sufficiently complex that developing an accurate model is prohibitively expensive. The principal advantage of data-driven approaches is that they can often be deployed quicker and cheaper compared to other approaches and that they can provide system-wide coverage. On the other hand, data-driven approaches require a substantial amount of data for training which is a fundamental limitation for most systems, since full trajectories to failure are not recorded in large numbers for components in highvalue systems like aircraft. Data-driven approaches can be further subcategorized into fleet-based statistics and sensor-based conditioning. In addition, data-driven techniques also subsume cycle-counting techniques that may include domain knowledge.

1060

K. Goebel and B. Saha

There are two basic data-driven strategies that involve either (1) modeling cumulative damage (or, equivalently, health) and then extrapolating out to a damage (or health) threshold or (2) learning directly from data the remaining useful life. As mentioned, a principal bottleneck is the difficulty in obtaining run-to-failure data, in particular for new systems, since running systems to failure can be a lengthy and rather costly process. Even where data exist, the efficacy of data-driven approaches is not only dependent on the quantity but also on the quality of system operational data. These data sources may include temperature, pressure, oil debris, currents, voltages, power, vibration and acoustic signal, spectrometric data, as well as calibration and calorimetric data. Features must be extracted from noisy, highdimensional data.

43.4

Case Study: Prognostics for Batteries Used in Electric UAV

This section will use batteries used for energy storage in electric UAV as an illustrative example to show the principles of prognostics. The section will start with a discussion on battery characteristics, followed by the model chosen and the implementation on the UAV.

43.4.1 Battery Characteristics Batteries are essentially energy storage devices that facilitate the conversion, or transduction, of chemical energy into electrical energy, and vice versa (Huggins 2008). They consist of a pair of electrodes (anode and cathode) immersed in an electrolyte and sometimes physically divided by a separator. The chemical driving force across the cell is due to the difference in the chemical potentials of its two electrodes, which is determined by the difference between the standard Gibbs free energies the products of the reaction and the reactants. The theoretical open-circuit voltage, E ı , of a battery is measured when all reactants are at 25 ı C and at 1 M concentration or 1 atm pressure. However, the voltage during use differs from the theoretical voltage because of various passive components like the electrolyte, the separator, and terminal leads. The voltage drop due to these factors can be mainly categorized into Ohmic drop, activation polarization, and concentration polarization. Ohmic drop EIR refers to the diffusion process where Li-ions migrate to the cathode via the electrolytic medium. The internal resistance to this ionic diffusion process is also referred to elsewhere as the IR drop. For a given load current, this drop usually decreases with time due to the increase in internal temperature that results in increased ion mobility. Self-discharge is caused by the residual ionic and electronic flow through a cell even when there is no external current being drawn. The resulting drop in voltage has been modeled to represent the activation polarization of the battery EAP . All chemical reactions have a certain activation barrier that must be overcome in order

increasing voltage E

43 Prognostics Applied to Electric Propulsion UAV

1061

IR drop

E⬚

activation polarization concentration polarization

increasing current I Fig. 43.3 Typical polarization curve of a battery (Saha and Goebel 2009)

to proceed and the energy needed to overcome this barrier leads to the activation polarization voltage drop. The dynamics of this process is described by the Butler– Volmer equation (Bockris and Reddy 1973). This process was represented by an exponential function in Saha and Goebel (2009). However, a log function is a more accurate representation, as abstracted from the Butler–Volmer equation. The concentration polarization ECP represents the voltage loss due to spatial variations in reactant concentration at the electrodes. This is mainly caused when the reactants are consumed by the electrochemical reaction faster than they can diffuse into the porous electrode, as well as due to variations in bulk flow composition. The consumption of Li-ions causes a drop in their concentration along the cell, which in turn causes a drop in the local potential near the cathode. The value of this factor is low during the initial part of the discharge cycle and grows rapidly towards the end of the discharge or when the load current increases. Figure 43.3 depicts the typical polarization curve of a battery with the contributions of all three of the above factors shown as a function of the current drawn from the cell. The voltage drop usually increases with increasing output current. The output current plays a prominent role in determining the losses inside a battery and is therefore an important parameter to consider when comparing battery performance. The term most often used to indicate the rate at which a battery is discharged is the C-Rate (Huggins 2008). The discharge rate of a battery is expressed as C =r, where r is the number of hours required to completely discharge its nominal capacity. So, a 5 Ah battery discharging at a rate of C =10 or 0.5 A would last for 10 h. The terminal voltage of a battery, as also the charge delivered, can vary appreciably with changes in the C -Rate. Furthermore, the amount of energy supplied, related to the area under the discharge curve, is also strongly C -Rate dependent. Figure 43.4 shows the typical discharge of a battery and its variation

1062

K. Goebel and B. Saha

Fig. 43.4 Schematic drawing showing the influence of the current density upon the discharge curve (Reproduced from Fig. 1.14 in Huggins 2008)

Low Current Medium Current

Voltage

High Current

Charge Delivered

with C -Rate. Each curve corresponds to a different C -Rate or C =r value (the lower the r the higher the current) and assumes constant temperature conditions.

43.4.2 Model Adaptation For many engineered systems, models for nominal operation are available, but damage propagation models like Arrhenius model or Paris’ law are comparatively rare. Developing these models may require destructive testing for model validation which may not be possible in many cases. In some cases, testing may be done on subscale systems, but there may be difficulty in generalizing the models learned. Additionally, the parameter values of these models are often system specific and thus need to be relearned for every new application. The PF framework described above can help in these cases by adapting the prognostic/aging model in an online fashion. One of the key motivating factors for using Particle Filters for prognostics is the ability to include model parameters as part of the state vector to be estimated. This allows model adaptation in conjunction with state tracking and, thus, produces a tuned model that can be used for long-term predictions. Let system health xk state evolution model f and measurement model h with known noise distributions ¨ and , respectively. Additionally, the parameter values of h are assumed to be known (without lack of generality). The system health state is assumed to be one dimensional. Stationary or (better) non-stationary measurement models can be used to account for progressive degradation in sensors caused by corrosion, fatigue, wear, etc. The parameters of f, denoted by ’k D f’j;k I j D 1; : : : ; nf g nf 2 N , are combined with xk to give the state vector xk D Œxk ’k T , where T represents the transpose of a vector or matrix. Equations (43.1) and (43.2) can then be rewritten as

43 Prognostics Applied to Electric Propulsion UAV

1063

xk D f.xk1 ; ˛k1 / C !k1

(43.8)

zk D h.xk / C vk :

(43.9)

To formulate the state equations for ’k , one can choose a Gaussian random walk such that ˛j;k D ˛j;k1 C !j;k1 (43.10)   where !j;k1 is drawn from a normal distribution, N 0; ¢j2 , with zero mean and variance ¢j2 . Given a suitable starting point ˛j;0 , and variance ¢j2 , the PF estimate will converge to the actual parameter value ˛N j , according to the law of large numbers. In this way, model adaptation has been introduced into the PF framework, adding nf extra dimensions, yet achieving convergence (and without incurring the curse of dimensionality). The notion of a good proposal density, though, comes into play in the choice of the values of ˛j;0 and ¢j2 . If the initial estimate ˛j;0 is far from the actual value and variance ¢j2 is small, then the filter may take a large number of steps to converge, if at all. The variance value may be chosen to be higher in order to cover more state space, but that can also delay convergence. One way to counter this is to make the noise variance itself a state variable that increases if the associated weight is lower than a preset threshold, i.e., the estimated parameter value is far from the true value, and vice versa. Thus, ˛j;k D ˛j;k1 C !j;k1 I j;k D cj;k

  2 ; !j;k1  N 0; j;k1

8 wth ;  j;k1 I cj;k D 1; if wk1 D wth ; : cj;k > 1; if wk1 < wth :

(43.11)

The multiplier cj; k is a positive valued real number, while the threshold wth is some value in the interval (0, 1). The intent is to increase the search space when the error is high and tightening the search when one is close to the target. Note that although this produces a better proposal density, it introduces a further nf dimensions to the state vector. It is quickly evident that it is not feasible to take this approach for all the parameters of a sufficiently high-order model. This motivates the use of sensitivity analysis techniques (SA) to determine the more sensitive parameters that need to be estimated online. SA is essentially a methodology for systematically changing parameters in a model to determine the effects on the model output. There are several methods to perform SA like local derivatives (Cacuci 2003), sampling (Helton et al. 2006), and Monte Carlo sampling (Saltelli et al. 2004). Depending upon the form of the system model, any of these methods may be used to assess which parameters to target. Assuming that the model function f in Eq. (43.8) is differentiable, i.e., @f=@aj (the time index k dropped for the sake of generality), it can be computed at any

1064

K. Goebel and B. Saha

Fig. 43.5 Effect on f .x/ D ˛1 exp(˛2 x) due to 10 % variation in parameters ˛1 and ˛2 (Saha and Goebel 2011)

point in the state space defined by xk D Œxk ˛k T . If the partial derivative is positive, then the value of the function increases with an increase in the parameter value and vice versa. The magnitude of the derivative indicates the degree to which the parameter affects the output of f, as shown in Fig.43.5. This directs the choice of the parameters to estimate online. If the posterior error given by eki D xki 

N X

wik xki

(43.12)

i D1

is positive, then the parameters that have a positive local partial derivative need to be reduced and those with a negative one need to be increased. The opposite holds true if the error is negative. The amount by which the parameters need to be reduced or increased also depends on the magnitude of the local partial derivative. The higher the magnitude, the smaller the steps needed in order to prevent instability while approaching the true value. This notion can be formalized in the following way (the particle index i has been dropped for the sake of generality): ˛j;k D ˛j;k1 C Cj;k C !j;k1 I

  !j;k1  N 0; j2 :

(43.13)

Cj;k / ek ;

ˇ @f ˇˇ / ; @˛j;k ˇxk

D K:

ek ˇ : @f=@˛jj;k ˇx k

(43.14)

43 Prognostics Applied to Electric Propulsion UAV

1065

Note that in this model adaptation scenario, the noise variance parameter is not added to the state vector since the search process is directed and not random as discussed previously.

43.4.3 Battery Model For the purposes of the electric UAV BHM, the battery model design space is explored at a high level of abstraction with respect to the underlying physics. To predict the end of discharge (EOD), it is required to model the SOC of the battery. For the empirical charge depletion model under consideration here, the output voltage E.tk / of the cell is expressed in terms of the effects of the changes in the internal parameters (Saha et al. 2011): E.tk / D E o  EIR .tk /  EAP .tk /  ECP .tk /

(43.15)

where E ı is the Gibb’s free energy of the cell, EIR is the Ohmic drop, EAP is the drop due to activation polarization, and ECP denotes the voltage drop due to concentration polarization. These individual effects are modeled as     EIRC .tk / D Ik ˛6 1  exp ˛7 tk  tIk  ˛1 tk

(43.16)

EAP .tk / D ˛2;k ln.1 C ˛3;k tk /

(43.17)

ECP .tk / D ˛4;k exp.˛5;k Ik tk /

(43.18)

where Ik is the step change in current at time tIk and the ’’s represent the set of model parameters to be estimated. The PF representation of the battery state for predicting EOD is given by ’j;k D ’j;k1 C !j;k ; j D 1; : : : ; :7

(43.19)

xk D E.tk / C !k

(43.20)

xk D Œxk ’j;k T ; j D 1; : : : ; :7

(43.21)

zk D E.tk / C k

(43.22)

where all but the parameters ’3 and ’5 are learnt from training data, while ’3 and ’5 are estimated by the PF online.

43.4.4 UAV Application Electric aviation concepts are receiving increasing attention because of their potential benefits for emissions, fuel savings, and possibly noise reduction. While electric propulsion is used in concept manned aircraft, electric UAV is more common due

1066

K. Goebel and B. Saha

Fig. 43.6 Edge 540 UAV

to their lighter weight and lower loss consequences (compared to manned flight). However, like ground vehicles, battery-powered electric propulsion in aircraft suffer from uncertainties in estimating the remaining charge and hence most flight plans are highly conservative in nature. Different flight regimes like takeoff/landing and cruise have different power requirements, and a dead stick condition (battery shut off in flight) can have catastrophic consequences. To tackle this issue the battery health management algorithm was designed and implemented on an embedded hardware platform and integrated it into an UAV airframe to provide real-time onboard battery life predictions (Saha et al. 2012). The particular UAV platform is a COTS 33 % scale model of the Zivko Edge 540 T as shown in Fig. 43.6. The UAV is powered by dual tandem mounted electric outrunner motors capable of moving the aircraft up to 85 knots using a 26 in. propeller. The gas engine in the original specification was replaced by two electric outrunner motors which are mounted in tandem to power a single drive shaft. The motors are powered by a set of four Li-Poly rechargeable batteries, each rated at 6,000 mAh. The tandem motors are controlled by separate motor controllers (Quach et al. 2013). A 12 channel JR radio system is used to control the airplane. The system communicates in the 2.4 GHz band using a proprietary DSM2 protocol. Control surfaces are manipulated by seven actuators. The airplane is equipped with a number of sensors to collect structure, propulsion, and navigation. The health of the structure is monitored using a series of strain gauges and accelerometers. Navigation data consist of GPS location, ground speed, altitude, true heading, and magnetic heading. Power plant data (such as motor RPMs, currents, battery voltages, and temperatures) help to assess adequacy of the thrust from the motors and 26 in. propeller, the relevant parameters being. Data from sensors are logged by two separate data systems instantiated in a combination of RCATS and PC104. The RCATS is a turn-key system with proprietary software that creates an ASCII log of connected sensors. It provides telemetry data to a laptop receiver which displays the data for callout to the pilot. It measures flight-related parameters such as motor RPM, motor temperature, airspeed, and z-axis acceleration at 10 Hz and interleaves GPS position and altitude data at 1 Hz. The PC104 stack consists of a CPU board, a DC/DC converter, an IO card, and a signal conditioning card for strain gauges. It runs MathWorks xPC Target

43 Prognostics Applied to Electric Propulsion UAV

1067

Fig. 43.7 The BHM system installed onboard

operating system. The data are acquired using a Simulink model that is compiled to an xPC target OS. It records strain, accelerometer, battery temperature, and motor current at 200 Hz. It also outputs a 0.5 Hz sine wave to the RCATS system for synchronizing data. The BHM system is designed to be a relatively low cost analog-to-digital data acquisition system. The design philosophy behind the first BHM system is to use COTS solutions which would have a light weight compact footprint. Figure 43.7 shows the system installed onboard the Edge 540 airframe. The BHM system itself utilizes several small COTS boards to convert TTL signal voltages into PC RS-232 signal voltages.

43.4.5 Implementation Results Testing on the Edge 540 UAV platform was initially carried out with the airframe restrained on the ground. The propeller was run through various load regimes indicative of the intended flight profile (takeoff, climb, multiple cruise, turn, and glide segments, descent, and landing) by varying the propeller rotational speed. Figure 43.4 shows the voltages during a typical profile. It is desired to predict when the battery will run out of charge, i.e., the EOD event indicated by the end of the voltage plots after landing. In order to evaluate the prognostic algorithm, multiple predictions were made at the time instants 13, 15, 17, and 19 min. It is not desired to completely discharge the batteries in flight since there needs to be some time for the UAV pilot to land the aircraft with some safety margin on the remaining battery life (Fig. 43.8).

1068

K. Goebel and B. Saha

Fig. 43.8 Predictions (vertical lines) during ground test (Saha et al. 2012)

22 21

Volts

20 19 18 17 16 15

0

200

400

600

800 secs

1000

1200

1400

22 particle estimates 21 measured voltage

19

amber alert

voltage threshold 17

Fig. 43.9 Battery voltage prediction using Particle Filter during flight test (Saha et al. 2012)

16

0

5

10

15 mins

20

red alert

18

EOD estimates

Volts

20

25

30

In order to validate the learned prognostic model, several dozen actual flight tests were conducted using the UAV with randomized flight profiles. The prediction performance was accurate to within 2 min, i.e., jtEOD  tNRUL;p j < 2 min, over multiple flights of durations between 15 and 25 min. Figure 43.9 shows the profile of one such flight. The blue line indicates the measured voltage, while the green dots indicate the state values of the PF. The grey lines denote the RUL;p values plotted every second, while the amber and red lines represent early alerts for the pilot to land the plane before the dead stick condition.

43 Prognostics Applied to Electric Propulsion UAV

43.5

1069

Conclusion

This chapter lays out a technique for predicting component degradation as exemplified on a battery used for propulsion of an electric UAV. The approach chosen here is model-based where the form of the model has been linked to the internal processes of the battery and validated using experimental data. The model was used in a PF framework to make predictions of EOD. By profiling the power required for different flight regimes like cruise segments, banked turns, and landings, one can estimate the mission completion probability by calculating the RUL pdf. Since the prediction result is in the form of a pdf, it is easy to integrate the BHM routine into a higher-level decisioning algorithm that can provide advance warning about when to land the aircraft. Generally, a similar process can be followed for other components where a damage progression model is built for the component at hand. Naturally, the model would be different, if, say, a bearing, an electronics component, or a structural element were the object of interest. The next step in Prognostics Health Management is to integrate the health information, in this case remaining life, into a decision-making unit that reacts appropriately. Depending on the prognostic horizon, the reaction could be to autonomously change controller settings, reconfigure system resources to ensure primary mission goals (and perhaps extend the life of the stressed component), invoke a replanning or rescheduling routine, or provide the information for later processing in a maintenance setting. Acknowledgments This work was performed as a cross-center collaboration between NASA Ames and Langley Research Centers (ARC and LaRC) and Dryden Flight Research Center (DFRC). The authors would like to especially thank Patrick Quach, Sixto L. Vazquez, Edward F. Hogge, Thomas H. Strom and Boyd L. Hill at LaRC, and Edwin Koshimoto at DFRC for their contributions. The funding for this work was provided by the NASA System-Wide Safety and Assurance Technologies (SSAT) project under the Aviation Safety Program of the Aeronautics Research Mission Directorate (ARMD).

References S. Arulampalam, S. Maskell, N.J. Gordon, T. Clapp, A tutorial on Particle Filters for on-line nonlinear/non-Gaussian Bayesian tracking. IEEE Trans. Signal Process. 50(2), 174–188 (2002) P. Barton, Prognostics for combat systems of the future. IEEE Instrum. Meas. Mag. 10, 10–14 (2007) O’M. Bockris, A.K.N. Reddy, Modern Electrochemistry, vol. 2 (Plenum, New York, 1973), pp. 845–1136 J.O’M. Bockris, A.K.N. Reddy, A. Gamboa-Aldeco, Modern Electrochemistry 2A: Fundamentals of Electrodics, 2nd edn. (Kluwer Academic/Plenum Publishers, New York, 2000) D.G. Cacuci, Sensitivity and Uncertainty Analysis: Theory, vol. I (Chapman & Hall, Boca Raton, 2003) J. Gu, D. Barker, M. Pecht, Prognostics implementation of electronics under vibration loading. Microelectron. Reliab. 47, 1849–1856 (2007)

1070

K. Goebel and B. Saha

J.C. Helton, J.D. Johnson, J.C. Salaberry, J.B. Storlie, Survey of sampling based methods for uncertainty and sensitivity analysis. Reliab. Eng. Syst. Saf. 91, 1175–1209 (2006) A. Hess, G. Calvello, T. Dabney, PHM a key enabler for the JSF autonomic logistics support concept, in: Proceedings of the IEEE Aerospace Conference Big Sky, 2004 R. Huggins, Advanced Batteries: Materials Science Aspects, 1st edn. (Springer, New York/London, 2008) Y. Li, T. Kurfess, S. Liang, Stochastic prognostics for rolling element bearings. Mech. Syst. Signal Process. 14(5), 747–762 (2000) D. Luchinsky, V. Osipov, V. Smelyanskiy, Model based IVHM system for the solid rocket booster, in: Proceedings of the IEEE Aerospace Conference, Big Sky, 2008 S. Marble, D. Tow, Bearing health monitoring and life extension in satellite momentum/reaction wheels, in: Proceedings of the IEEE Aerospace Conference, Big Sky, 2006 V. Osipov, D. Luchinsky, V. Smelyanskiy, In-flight failure decision and prognostics for the solid rocket booster, in: AIAA 43rd AIAA/ASME/SAE/ASEE Joint Propulsion Conference, Cincinnati, 2007 P. Paris, F. Erdogan, A critical analysis of crack propagation laws. Trans. ASME J. Basic Eng. 85, 528–534 (1963) C.C. Quach, B. Bole, E. Hogge, S. Vazquez, M. Daigle, J. Celaya, A. Weber, K. Goebel, Battery charge depletion prediction on an electric aircraft, in Proceedings of Annual Conference of the PHM Society (PHM’13), New Orleans, 2013 M. Revor, E. Bechhoefer, Rotor track and balance cost benefit analysis and impact on operational availability, in: Proceedings of the American Helicopter Society 60th Annual Forum, Baltimore, 2004 B. Saha, K. Goebel, Modeling Li-ion battery capacity depletion in a Particle Filtering framework, in: Proceedings of the Annual Conference of the Prognostics and Health Management Society, San Diego, 2009 B. Saha, K. Goebel, S. Poll, J. Christophersen, An integrated approach to battery health monitoring using Bayesian regression and state estimation, in: Proceedings of the IEEE Autotestcon, Baltimore, 2007, pp. 646–653 B. Saha, K. Goebel, S. Poll, J. Christophersen, Prognostics methods for battery health monitoring using a Bayesian framework. IEEE Trans. Instrum. Meas. 58(2), 291–296 (2009) B. Saha, C. Quach, K. Goebel, Exploring the model design space for battery health management, in: Proceedings of the Annual Conference of the Prognostics and Health Management Society, Montreal, 2011 B. Saha, P. Quach, K. Goebel, Optimizing battery life for electric UAVs using a Bayesian framework, in: Proceedings of the 2012 IEEE Aerospace Conference, Big Sky, 2012 A. Saltelli, S. Tarantola, F. Campolongo, M. Ratto, Sensitivity Analysis in Practice: A Guide to Assessing Scientific Models (Wiley, Chichester/Hoboken, 2004) M. Valenti, B. Bethke, D. How, D. de Farias, J. Vian, Embedding health management into mission tasking for UAV teams, in: Proceedings of the American Control Conference, New York, 2007 W.K. Yu, T. Harris, A new stress-based fatigue life model for ball bearings. Tribol. Trans. 44, 11–18 (2001)

Actuator Fault Detection in UAVs

44

Guillaume Ducard

Contents 44.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1072 44.1.1 Definition of Fault and Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1072 44.1.2 Different Approaches for FDI Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074 44.1.3 Interaction Between Flight Controllers and FDI Systems . . . . . . . . . . . . . . . . . . . . . . . 1076 44.1.4 Other Practical Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077 44.2 Aircraft Configuration and Dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077 44.2.1 Aircraft Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1078 44.2.2 Aircraft Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1078 44.3 Residual Generator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1079 44.3.1 EKF Equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1080 44.3.2 Filter’s Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1081 44.4 Multiple Model Approaches for FDI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1084 44.4.1 Modeling Actuator Faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1085 44.4.2 The EMMAE Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1086 44.4.3 Designing the EKF for the No-Fault Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1088 44.4.4 Augmenting the State Vector with the Faulty Actuator Parameter ıNi . . . . . . . . . . . 1089 44.4.5 Designing the EKF for the Case of a Failure on Aileron 1 . . . . . . . . . . . . . . . . . . . . . . 1090 44.4.6 Actuator Fault Isolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1090 44.4.7 Simulation Results of the EMMAE-FDI with No Supervision System. . . . . . . . . 1095 44.4.8 Remarks on the First Attempt to Use the EMMAE-FDI System . . . . . . . . . . . . . . . 1098 44.4.9 Techniques to Improve Actuator Fault Diagnosis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1098 44.4.10 Computational Complexity of the EMMAE-FDI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101 44.4.11 Realistic Flight Scenario and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1102 44.5 A Single Model Active (SMAC) FDI System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103 44.5.1 Residual Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103 44.5.2 Fault Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106 44.5.3 Excitation Signals Generator (ESG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1109

G. Ducard I3S CNRS-UNS, Sophia Antipolis, France ETH Zurich, IDSC, Zurich, Switzerland e-mail: [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 43, © Springer Science+Business Media Dordrecht 2015

1071

1072

G. Ducard

44.5.4 Actuator Health Evaluator (AHE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1110 44.5.5 Fault Isolator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1112 44.5.6 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114 44.5.7 Properties of the SMAC-FDI System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117 44.5.8 Computational Load Requirements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117 44.5.9 Conclusions About the SMAC-FDI System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117 44.6 Chapter General Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1118 Appendix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1118 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1119

Abstract

Future unmanned aerial vehicles (UAVs) will be designed to achieve their missions with increased efficiency, safety, and security. To this end, an efficient fault detection and isolation (FDI) system should be capable of monitoring the health status of the aircraft. Fault-tolerant control systems for small and low-cost UAVs should not increase significantly the number of actuators or sensors needed to achieve the safer operation. This chapter is dedicated to actuator fault detection systems for UAVs, with two main requirements: realtime capability and modularity. After defining the terminology employed in this field, this chapter reviews some commonly used techniques in FDI systems. The chapter continues by presenting briefly the mathematical model of a UAV which will serve as a basis for the design of two actuator FDI systems. The first method presents and illustrates the multiple-model approach, whereas the second method presents an FDI system which is based on a single model. Both methods have been enhanced by a mechanism that actively tests actuators in order to efficiently detect and isolate actuator faults and failures. This chapter explains the advantages and drawbacks of each method and discusses issues of robustness against model uncertainties and external perturbation. In addition, aspects of computational load are addressed. Finally, the FDI systems of this chapter are applied to a realistic model of an unmanned aircraft, and the performance of the methods is shown in simulation.

44.1

Introduction

44.1.1 Definition of Fault and Failure According to the definition found in Isermann (2006), “a fault is an unpermitted deviation of at least one characteristic property or feature of the system from the acceptable, usual, standard condition.” Based on this definition, a fault corresponds to an abnormal behavior of the system, which may not affect the overall functioning of the system but may eventually lead to a failure (defined below). Moreover, a fault may be small or hidden and therefore difficult to detect and estimate. For example, consider the temperature of an engine. If this temperature exceeds a certain accepted limit, say 100 ı C, there is a fault in the system. Although this

44 Actuator Fault Detection in UAVs

1073

a

b

δmax

δmax δ tF

time

tF

δmin

δmin

c

d

δmax

δmax

desired actuator position true actuator position

δ tF δmin

time

time

tF

time

δmin

Fig. 44.1 Several types of actuator failures: (a) floating around trim, (b) locked-in-place, (c) hardover, and (d) loss of effectiveness (actuator fault occurring after tF )

excessive temperature does not prevent the engine from working properly for a while, it may eventually damage components of the engine and possibly lead to its breaking down. In this chapter, an actuator fault corresponds to any abnormal behavior. This includes bias or loss of effectiveness as shown in Fig. 44.1d. A failure is a permanent interruption of a system’s ability to perform a required function under specified operating conditions. (Isermann 2006)

Resulting from one or more faults, a failure is therefore an event that terminates the functioning of a unit in the system. On an aircraft, actuators are used to deflect control surfaces such as ailerons, elevators, and rudders, and also to actuate the engine throttle or the landing-gear mechanism. An actuator is declared failed when it can no longer be used in a controlled manner. For a control surface, there are two major types of failures (Boskovic and Mehra 2003). As shown in Fig. 44.1a, the control surface may become ineffective and float at the zero-moment position. The control surface can also be locked at any arbitrary intermediate position (Fig. 44.1b) or reach and stay at the saturation position as shown in Fig. 44.1c. Mechanical failures may also happen. This is the case when the mechanical link between the control surface and its corresponding actuator or servo breaks. The engine may also fail. Finally, there are other sources of possible irreversible damage to the aircraft that may be classified as structural failures. They correspond to the scenarios where a piece of the aircraft is missing, such as an aileron, a tail rudder, an elevator, or part of a wing.

1074

G. Ducard

44.1.2 Different Approaches for FDI Systems In the context of reconfigurable flight control using an explicit online fault detection and isolation system (FDI), results from the FDI are used to modify the control system’s structure, laws, parameters, or even the trajectory of the aircraft. There are two families of FDI systems, namely, passive and active FDI systems. Passive FDI systems “wait” until a fault or failure occurs (Maybeck and Stevens 1991; Maybeck 1999), whereas active FDI systems will artificially excite the aircraft, either by flying health-check maneuvers (Azam et al. 2005; Elgersma et al. 1998) or by injecting test signals in the actuator commands and then assessing the individual health status of actuators and sensors. Very few papers have discussed this technique so far. The work published by Honeywell in 1998 (Elgersma et al. 1998) and 2001 (Elgersma and Glavaski 2001) is among the first occurrences of using artificial exciting signals for FDI purposes. Test signals are injected into the null space of the inputs using redundant control surfaces such that these signals (ideally) cancel one another and thereby do not excite aircraft motion (Elgersma and Glavaski 2001) but contribute to better fault diagnosis; see also Buffington et al. (1999). The use of artificial signals was demonstrated to improve significantly the performance of an FDI system based on the extended multiple model adaptive estimation (EMMAE) method (Ducard and Geering 2006, 2008). This method is described in Chap. 4 of this book (Ducard 2009). In 2007, the authors of Boskovic et al. (2007) suggested an adaptive fault-tolerant controller with self fault-diagnosis actuators. This is done by generating high-frequency signals for actuators with suspected failures and minimizing the effects of those signals on the system state using the remaining healthy actuators. In Bateman et al. (2011), the isolation procedure between redundant actuators is inspired from the excitation mechanism from Ducard and Geering (2006). Latest contributions in active FDI are reported in Ducard and Geering (2010) and extended in the second part of this chapter. Table 44.1 provides a list of common and recent techniques that are encountered in the literature for the design of FDI systems.

44.1.2.1 Trends in Filter Design for FDI System In the mid-1990s, several implementations of recursive least-squares (RLS) algorithms were used in FDI systems and successfully flight tested. For example, the work by Ward et al. (1998) describes a computationally efficient real-time parameter identification and reconfigurable control algorithm. The identification algorithm is based on a modified sequential least-squares (MSLS) found in Ward and Barron (1995), the recursive version of which is found in Bodson (1995). The MSLS parameter identification algorithm is based on RLS techniques and incorporates additional constraints to take into account a priori information and to adjust the size of the data window used in the regressor of the filter (Ward et al. 1994). Many FDI filters have also been designed using mathematical models of the system being monitored. Model-based FDI methods have been enhanced using

44 Actuator Fault Detection in UAVs

1075

Table 44.1 List of some recent and popular techniques used to design FDI systems for flight applications Technique Example of recent books/papers using this technique (ordered chronologically) (Modified -) recursive least-squares (RLS) Kalman filters (KF) (bank of -)

Extended Kalman filters (EKF) (bank of -)

Unscented Kalman filters (UKF)

Linear parameter varying (LPV) filters Interaction matrix Particle filters Neural networks

Statistical methods Wavelet analysis H1 Robust model-based system Parity space approach

Unknown input observer

Ward et al. (1994, 1998), Bodson (1995), and Shore and Bodson (2005) Urnes et al. (1990), Maybeck and Stevens (1991), Eide and Maybeck (1996), Maybeck (1999), Ni (2001), Hajiyev and Caliskan (2003), and Fekri et al. (2006) Hajiyev and Caliskan (2003), Tanaka et al. (2006), Ducard and Geering (2006, 2008), and Ducard (2007) Julier et al. (2000), Brunke and Campbell (2002), Campbell et al. (2007), and Perea and Elosegui (2008) Szaszi et al. (2005) Koh et al. (2005) Rapoport and Oshman (2005) Calise et al. (1998), Younghwan (1998), Wise et al. (1999), Brinker and Wise (2000), and Azam et al. (2005) Isermann (2006) and Samara et al. (2008) Adams (1994) and Azam et al. (2005) Collins and Song (2000), Marcos et al. (2005), and Rotstein et al. (2006) Chen and Patton (1999) and Patton et al. (2000, 2008) Chow and Willsky (1984), Frank (1994), Gertler (1997), Chen and Patton (1999), and Isermann (2006) Patton et al. (1989), Chen et al. (1996), and Patton and Chen (1997)

robust FDI techniques as defined in Chen and Patton (1999). It consists of incorporating during the design of FDI systems the effects of disturbance signals, model uncertainties, and measurement noise (Patton et al. 2000). It is often the case that several model-based filters are organized in a bank in which one filter is sensitive to a specified failure but the other filters remain insensitive to that failure. A recent example of this technique can be found in Patton et al. (2008), where a robust fault diagnosis for a spacecraft attitude control system is designed. Many different variants of Kalman filters (KFs) have been constructed for detecting and isolating faults or for state estimation and state reconstruction. The use of extended Kalman filters (EKFs) applied to nonlinear systems for FDI purposes has also gained recent interest in Ducard and Geering (2008). A recent paper presented a method that uses EKFs to estimate online the aircraft’s aerodynamic parameters and the components of the wind velocity. These estimates are used to update the parameters of the flight controller (Tanaka et al. 2006).

1076

G. Ducard

The unscented Kalman filter (UKF) is among the latest extensions of Kalmantype filters and seems to provide remarkable results for systems that are particularly nonlinear. The paper by Campbell et al. (2007) discusses the implementation of a sigma point filter (SPF) which was originally introduced as the UKF (Julier et al. 2000), where the distributions are approximated by a finite set of points. It is used to estimate aircraft states and aerodynamic derivatives in real time. This is a nonlinear estimation algorithm that can be performed online, which possesses robustness properties against parameter uncertainties, against filter tuning, and against initial conditions. The discussion in Julier et al. (2000) explains that the SPF has similar performance to a truncated second-order EKF but without the need to calculate the Jacobian matrices. A comparison between EKF and SPF can also be found in Brunke and Campbell (2002). The main results of this chapter indicate that the SPF filter has equal or better performance than an EKF for real-time estimation application for the following reasons: the SPF is more robust against initial uncertainties and against jumps in the data, is less sensitive to tuning of the process noise, is less susceptible to divergence, is more accurate from one time step to the next, and, finally, requires equivalent computational load. Recent contributions in Perea and Elosegui (2008) focused on a new formulation for the state update equation of the filter for improved accuracy. Recently, linear parameter-varying (LPV) filters gained the attention of some researchers in the fault-tolerant control community. For example, a design of LPVbased FDI filters is found in Szaszi et al. (2005). An example of an H1 control law that minimizes command tracking errors under actuator fault occurrence combined with an FDI filter based on an affine LPV model of a Boeing 747 is found in Shin et al. (2006).

44.1.2.2 Challenges of Designing Reliable FDI Systems A reliable FDI system provides accurate information about the health status of the aircraft. In order to achieve such a result, the FDI system needs to be robust against external disturbances, model uncertainties, and sensor noise. In addition, the FDI system should not trigger false alarms and should still be sufficiently sensitive to detect the faults. Therefore, robustness is a fundamental issue in the performance of FDI systems and reconfigurable flight controller. FDI systems may experience significant performance reduction if the model uncertainties are not properly considered. A robustness analysis framework for failure detection and accommodation is provided in Chen and Patton (1999), Zhang and Jiang (2000), and Belcastro and Chang (2002).

44.1.3 Interaction Between Flight Controllers and FDI Systems It is often the case that a reconfigurable flight control system incorporates an FDI system and a flight controller. The FDI system monitors the aircraft’s behavior and identifies relevant parameters that are usually used by the flight controller to

44 Actuator Fault Detection in UAVs

1077

synthesize the control commands. Therefore, the performance of the flight controller is dependent on the results provided by the FDI system and vice versa. Thus, the interactions between these two systems should be rigorously investigated. The following observation is made in Shin et al. (2006): “it is fairly common for integration of failure detection and accommodation systems to be problematic if they are designed separately.” Challenges exist when some aircraft parameters need to be identified during the flight in real time and under feedback control (Ward et al. 1998). This task is even more difficult and delicate when an actuator or a sensor fault happens. Moreover, the robustness of the flight controller can mask some aircraft faults and failures and make the detection problem more difficult. There exist already many examples of integrated fault-tolerant control, sometimes referred to as IFTC (Zhang and Jiang 2000; Schierman et al. 2004; Boskovic et al. 2005; Shin et al. 2006). New generations of reconfigurable flight control systems will not only rely on a fault-tolerant controller but will include complete and integrated systems that reconfigure the flight controllers, adapt the guidance system, and reshape online the vehicle trajectories (Schierman et al. 2004).

44.1.4 Other Practical Challenges Usually, the flight control system relies on some nominal values for the mass, the moments of inertia, and the aerodynamic coefficients to generate the control signals. When the aircraft experiences an actuator failure or airframe damage, the aircraft becomes asymmetric. It is thus not trivial to determine which of these parameters need to be (online) reestimated to keep good flying performance. It is often the case that the available onboard processing power is limited, in particular for small or micro UAVs. The design of a reconfigurable flight controller is therefore a trade-off between performance, complexity, and available processing power. Finally, challenges are many when the fault diagnostic, the reconfiguration of the control system, and the reconfiguration of the path plan or of the mission are to be done autonomously under limited or no human supervision.

44.2

Aircraft Configuration and Dynamics

This section presents the mathematical model of an unmanned aircraft, whose actuators are to be monitored by an FDI system. This model has been constructed, validated, and identified through real flight tests of an aerobatic UAV and is reported in M¨ockli (2006). This model will serve as a basis for the design of the different filters created for the FDI systems of the subsequent sections. The numerical values of the aircraft model parameters are found in the Appendix.

1078

G. Ducard rudder: δr right elevator: δe2 left elevator: δe

1

right aileron: δa

2

yb

Ob

left aileron: δa1

yn East xb

On

xn North

zn Down

zb

Fig. 44.2 Aircraft configuration

44.2.1 Aircraft Configuration The five control surfaces of the aircraft under consideration are one left aileron, one right aileron, one left elevator, one right elevator, and one rudder, as shown in Fig. 44.2. All actuators are fully independent, which means that ailerons (or elevators) can individually move up or down. This configuration permits some pitch torque to be produced with ailerons or some roll torque to be produced with elevators. The control vector of the aircraft involving only actuator deflections is ı D Œıa1 ıa2 ıe1 ıe2 ır .

44.2.2 Aircraft Dynamics Among the nonlinear equations which describe the dynamics of an aircraft, those involving the turn rates are of primary interest. Indeed, as soon as any control surface exhibits a faulty behavior, the aerodynamic moments applied to the airframe are altered. The fault detection system constructed in this chapter is based on a simplified aircraft model that gives the explicit relationship between turn rates (p, q, r), the inertia matrix I b , and the torques applied to the aircraft, that is, ŒL M N T , expressed in the body-axes frame .xb ; yb ; zb / of the aircraft: 02 3 b 2 3 2 3 2 31 pP p p L C b 4 qP 5 D .I b /1 B @4 M 5  4 q 5  I 4 q 5A : rP r r N

(44.1)

44 Actuator Fault Detection in UAVs

1079

In the context of this work, the aircraft is a small UAV for which the aerodynamic moments have been modeled as follows (see Stevens and Lewis (2003), Stengel (2004), M¨ockli (2006), and Ducard (2009)): L D qS N bCL .ıa1 ; ıa2 ; ıe1 ; ıe2 ; p; r; ˇ/ ; M D qS N cC N M .ıa1 ; ıa2 ; ıe1 ; ıe2 ; ˛; q/ ; N D qS N bCN .ıa1 ; ıa2 ; ıe1 ; ıe2 ; ır ; r; ˇ/;

(44.2)

V 2

where the dynamic pressure is qN D 2T , the total airspeed of the aircraft is VT , the air density is , the wing total surface is S , the wing span is b, and the mean aerodynamic wing chord is c. N The aerodynamic derivatives are expressed as a linear combination of the state elements and control inputs as CL D CLa1 ıa1 C CLa2 ıa2 C CLe1 ıe1 C CLe2 ıe2 C CLpNN pNN C CLrNN rNN C CLˇ ˇ ; CM D CM1 C CMa1 ıa1 C CMa2 ıa2 C CMe1 ıe1 C CMe2 ıe2 C CM qNN qN C CM˛ ˛ ; CN D CNır ır C CN rNN rNN C CNˇ ˇ;

(44.3)

bp cq N br ; qNN D 2V ; rNN D 2V : The last two with the dimensionless angular rates pNN D 2V T T T differential equations concern the angle of attack ˛ and the sideslip angle ˇ as follows (see Ducard (2009) for the derivation of the following two formulae):

˛P  q C

g VT

 1C

 qS N .ŒCX1 C CZ˛ ˛ C CZ1 / ; mg

qS N CY 1 ˇ; ˇP  r C mVT

(44.4)

with the drag derivative CX1 , the side force derivative CY 1 , and the lift derivatives CZ1 , CZ˛ being constant terms. Finally, the inertia matrix expressed in the body-fixed frame is I b D 2 3 Ixx 0 Ixz 4 0 Iyy 0 5, with Ixz D Izx . The numerical values of the parameters of the airIzx 0 Izz craft model used throughout this chapter are summarized in Table 44.2 in Appendix.

44.3

Residual Generator

In this chapter, the detection of a fault in the system is achieved by monitoring a signal that is within certain limits in normal conditions and that goes beyond these limits in faulty situations. This is the reason why a mathematical model has

1080

G. Ducard

been developed to describe the behavior of the aircraft. This model and the real aircraft receive the same actuator commands issued by the flight controller. The real aircraft dynamics are measured by sensors mounted aboard the vehicle. They will be compared with the predicted aircraft dynamics that are computed with the aircraft model. The difference between the real and predicted dynamics constitutes signals that are called residuals. However, the mathematical model of the aircraft will never be perfect, leaving many effects unmodeled. Also, this model will only be approximated by the computer implementation. Furthermore, several parameters of the model will not be known exactly, and the sensor measurement data will be corrupted by noise and biases. For all these reasons, Kalman filtering techniques are employed in this work to take into account such system dynamics and measurement noise, errors, and uncertainties. Kalman filters (KF) are well suited when the real world can be described by linear differential equations expressed in state space form and when the measurements are linear functions of the states. However, in most realistic problems, the real world is described by nonlinear differential equations. In order to take into account these nonlinearities, the residual generator of the FDI systems presented in this chapter uses extended Kalman filters (EKF), which are recalled below.

44.3.1 EKF Equations The EKFs are designed based on a set of continuous nonlinear differential equations that describe the plant under consideration as follows (see Zarchan and Musoff (2005)): xP D f.x; u/ C w; (44.5) where the state vector is x, the control input vector is u, the set of nonlinear functions of the state and control vectors is f .x; u/, and the random zero-mean process noise vector is w. The continuous process noise covariance matrix describing the random process w is given by R w D EfwwT g: (44.6) Equation 44.5 is first linearized around the current operating point and then discretized using the Euler integration method. Note that the Euler integration method is also used for the simulations presented in the following sections. The discretized form of (44.5) is expressed in state space form as xkC1 D k xk C Gk uk C wk ;

(44.7)

where the state vector is evaluated at the discrete-time instant tk D kTs , with the constant Ts being the sampling period of the system. The control input vector at time step k is uk , and the discrete random zero-mean process noise wk is used to describe uncertainties in the model.

44 Actuator Fault Detection in UAVs

1081

Finally, the discrete form of the measurement equation, either a linear or a nonlinear function of the states, is yk D h.xk / C vk ;

(44.8)

where the discrete zero-mean random noise vk is described by the measurement noise covariance matrix Rv;k D Efvk vTk g and consists of the variances of each of the measurement noise sources. The discrete transition matrix k is approximated by k  I C F.k/Ts ;

(44.9)

where the continuous system dynamics matrix F.k/ is obtained by linearizing the continuous nonlinear equations and is evaluated at the latest available state estimate xO kjk according to ˇ @f.x; u/ ˇˇ F.k/ D : (44.10) @x ˇxDOxkjk ; uDuk Similarly, the continuous measurement matrix H.k/ is computed by linearizing the (possibly nonlinear) measurement equation h.x/ and is successively evaluated at the latest available state estimate xO kjk (see mechanization shown in Fig. 44.3) according to ˇ @h.x/ ˇˇ H.k/ D Hk D : (44.11) @x ˇxDOxkjk The equations used in the EKF are described below (Kalman 1960; Brown and Hwang 1997; Zarchan and Musoff 2005). The schematic overview of the computation steps is shown in Fig. 44.3.

44.3.2 Filter’s Equations The equations used in the filter are as follows: 1. The Kalman gain matrix Lk is computed as  1 Lk D †kjk1 HTk Hk †kjk1 HTk C Rv;k

(44.12)

and is a function of the last-propagated state error covariance matrix †kjk1 and of the measurement noise covariance matrix Rv;k . 2. The measurement update of the state estimate is as follows: xO kjk D xO kjk1 C Lk Œyk  h.Oxkjk1 /;

(44.13)

1082

G. Ducard

Fig. 44.3 EKF mechanization

where the last extrapolated state estimate is xO kjk1 , the measurement vector is yk , and the estimated measurement vector is h.Oxkjk1 /. The set of continuous nonlinear measurement equations is h.:/. 3. The third step concerns the update of the state error covariance matrix †kjk D Efekjk eTkjk g with ekjk D x.k/  xO kjk , where x.k/ is the true (unknown) value of the state vector at the discrete instant k. The matrix †kjk is recursively computed

44 Actuator Fault Detection in UAVs

1083

as a function of the last predicted state error covariance matrix †kjk1 and the last computed Kalman gain matrix Lk as follows: † kjk D ŒI  Lk Hk  †kjk1 :

(44.14)

4. The forward propagation of the state error covariance matrix is †kC1jk D k †kjk kT C Rw;k ;

(44.15)

where the matrix Rw;k represents the covariance of the discrete process noise acting on the elements of the state vector. The value of Rw;k is found from the continuous process noise covariance matrix Rw and the continuous transition matrix  according to Z Rw;k D

Ts 0

./Rw  T ./d ;

(44.16)

where the matrix .t/ is defined as follows: .t/ D I C F.k/t:

(44.17)

There is an alternative solution of introducing process noise into the system through the control input matrix. Indeed, if the plant is described by the following equation: xkC1 D k xk C Gk uk C Gk wk ;

(44.18)

then the process noise covariance matrix is defined by Rw;k D Gk Rw GTk ;

(44.19)

where the covariance matrix of the discrete process noise acting on the elements of the ˇ state vector is Rw;k , and the discrete control input matrix is Gk D ˇ @f ˇ . Ts @u ˇ VT .k/

5. The propagation forward of the state estimate does not have to be done with the discrete transition matrix k , but rather it is done directly by integrating the actual nonlinear differential equations forward at each sampling interval. If the Euler integration technique is used, the extrapolated state estimate is computed with xO kC1jk D xO kjk C xPO kjk Ts ; (44.20) where the state time derivative is obtained from   xPO kjk D f xO kjk ; uk :

(44.21)

1084

G. Ducard

Remarks: • Note that the EKF presented in this section keeps track of the total estimates and not of the incremental ones (deviation from a nominal trajectory) as would be the case in a linearized KF. Indeed, the residuals are built from the difference between the true measurement vector and the predicted measurement vector using the set of nonlinear measurement equations acting on the total state estimate h.Oxkjk1 /. • Moreover, the measurement update of the state estimate is done using the total estimate xO .kjk  1/, and the propagation forward of the total state estimate is achieved using the set of nonlinear differential equations acting on the total state estimate f.Oxkjk ; uk / instead of using the discrete transition matrix acting on incremental state quantities. • Nevertheless, the computation of the Kalman gains and state error covariance matrices makes use of the recursively updated linear model through k , Hk , and Rw;k . • The schematic overview shown in Fig. 44.3 clarifies the mechanization of the EKF implemented in this work.

44.4

Multiple Model Approaches for FDI Systems

One popular approach to detect and isolate actuator or sensor faults is the multiple model adaptive estimation (MMAE) method (Magill 1965; Maybeck and Stevens 1991) as depicted in Fig. 44.4. It is based on a bank of parallel Kalman filters (KF), each of which is matching a particular fault status of the system. A hypothesis u y

KF based on No Failure KF based on Failure 1

xˆnf ∑nf rnf xˆ1

xˆ

∑1 r1

KF based on Failure i

xˆi ∑i ri pi Hypothesis Conditional Probability Computation

Fig. 44.4 Classical MMAE scheme

p1 pnf

44 Actuator Fault Detection in UAVs

1085

testing algorithm uses the vector of residuals r i and state error covariance matrix † i from each KF to assign a conditional probability to each fault hypothesis. Several papers have demonstrated how the MMAE method can be used in the context of FDI systems for aircraft (Maybeck and Stevens 1991; Eide and Maybeck 1996; Maybeck 1999) and underwater vehicles (Ni 2001). The main advantage of the MMAE method lies in its responsiveness to parameter variations, potentially leading to faster fault isolation than that obtained by other methods without a multiple model structure. However, the MMAE method possesses three limitations to its successful application as explained in Ducard and Geering (2008). The first limitation concerns the number of filters that must be designed in order to span the range of possible fault scenarios, which must be limited due to computational load. The second limitation appears when an actuator is locked at an arbitrary nonzero position that biases the residuals of the KFs, leading to inaccurate fault detection and state estimation. Third, most of the implementations of an MMAE method only work efficiently around predefined operating conditions. This chapter shows how the MMAE method is modified to compensate the three limitations of the classical MMAE implementation by employing extended Kalman filters (EKF) to estimate the deflection of a faulty control surface (or actuator). The resulting method is called thus “extended multiple model adaptive estimation” (EMMAE), and its architecture is shown in Fig. 44.6 (Rupp et al. 2005).

44.4.1 Modeling Actuator Faults A lock-in-place or floating actuator fault in the system can be seen as if the desired control input ıj was disconnected and replaced by a faulty control signal ıNj that takes control over the plant, as shown in Fig. 44.5. In a concise manner (Tao et al. 2004), the true input of the plant can be written as ui .t/ D ıi .t/ C Ai .ıNi .t/  ıi .t//:

σA1

δ1(t)

yref (t)

δ1(t)

Reconfigurabe Controller

ˆ , y(t) , δˆi (t) x(t)

EMMAE

y(t)

Fig. 44.5 Modeling of actuator faults

δm(t)

u(t)

1 − σA1

δm(t) δ(t)

(44.22)

1 − σAm

σAm

Aircraft

y(t)

1086

G. Ducard

In the case of actuator failure(s), the vector of the (unknown) inputs is N D ŒıN1 .t/ ıN2 .t/ ı.t/ with  D diagfA1

A2

::: 

Aj D

1; 0;

:::

ıNm .t/T ;

(44.23)

Am g, where  if the j th actuator fails : otherwise

(44.24)

In the method presented below, the unknown parameters ıNj are constantly estimated by their respective EKF. The conditional fault-hypothesis probabilities pj assign the value for Aj .

44.4.2 The EMMAE Method The MMAE method is to be made applicable for any arbitrary lock-in-place faults or uncontrolled varying faults and at all flying conditions. Therefore, the original MMAE algorithm is modified by replacing the linear KFs by EKFs used as nonlinear estimators of the state vector and a fault parameter, namely, the deflection of a faulty control surface (or actuator) ıNj . The implementation of the EMMAE is depicted in Fig. 44.6. Contrary to the FDI designs with

y

xˆ

xˆ nf

u EKF based on No Failure

∑nf rnf δˆ1

δˆ1 EKF based on Failure 1

xˆ 1 ∑1 r1 δˆ i

δˆi EKF based on Failure i

xˆ i ∑i ri pi

Hypothesis Conditional Probability Computation

p1 pnf

Fig. 44.6 EMMAE-FDI scheme: each EKF monitors its assigned actuator

44 Actuator Fault Detection in UAVs

1087

the classical MMAE method where several KFs are designed for several faulty deflections for one actuator, in the EMMAE method, only one EKF is responsible for completely monitoring one actuator’s health. Therefore, the EMMAE method drastically reduces the number of filters required for actuator health monitoring. The addition of the actuator deflection estimate in the system state vector enables the EMMAE method to work for all the possible positions where an actuator can be locked or floating. In order to better illustrate the difference between the EMMAE and MMAE methods, Eqs. (44.25) and (44.26) recall how the models are defined in the regular MMAE method (Eide and Maybeck 1996; Maybeck 1999; Maybeck and Stevens 1991; Ni 2001) for an actuator or a sensor failure. The MMAE scheme considers a bank of linear models of the form xP D Ax C Bu; y D Cx C Du;

(44.25)

where each model matches a fault scenario. For example, the model that describes a failure of the j th actuator will have its B matrix modified such that the j th column of the B matrix is replaced by the very same column times a factor j that varies from zero (complete loss of the actuator) to one (fully functioning actuator), see (44.26). Any intermediate value of j indicates a reduction in the effectiveness of the j th actuator to modify the dynamics of the aircraft as shown in (44.26): xP D Ax C Bj u; 2 b11 6 :: 6 : 6 xP D Ax C 6 6 bl1 6 : 4 :: bp1

   b1j j    :    ::       blj j    :    ::   

32 3 u1 b1N :: 7 6 :: 7 6 7 : 7 76 : 7 7 7 blN 7 6 6 uj 7 : 6 : 7 :: 7 : 5 4 :: 5

   bpj j    bpN

(44.26)

uN

However, this kind of approach for the modeling of actuator failure is very restrictive. Indeed, in the case of a total loss of the j th actuator, the factor j equals zero. This means that whatever control input the controller generates for the j th actuator, it has no influence on the dynamics of the aircraft, and the faulty actuator deflection is considered to be zero. Note that if the j th actuator is actually locked at a nonzero deflection angle, the faulty actuator deflection has an influence on the dynamics of the aircraft. This condition results in an unknown bias term that will prevent the j th KF in the MMAE method from working properly. Therefore, the residuals will be biased, and the state estimation and the computation of the probabilities will be incorrect as well.

1088

G. Ducard

In the EMMAE method, both the control input matrix and the dynamics matrix are modified. Indeed, in order to define a model that describes a failure of the j th actuator, the j th column of the control input matrix is zeroed and the state vector is augmented with the j th actuator deflection ıNj . The dynamics matrix is also augmented with the original j th column of the control input matrix. In this way, the control inputs from the controller to the j th actuator are totally ignored, but the faulty deflection ıNj that is constantly estimated (ıONj ) in the state vector contributes to modifying the dynamics of the aircraft model of the j th filter. It yields residuals that are the smallest for the filter matching the occurring fault. The next section illustrates how the filters in the EMMAE are constructed in practice.

44.4.3 Designing the EKF for the No-Fault Scenario In the EMMAE method, the residuals r.k/ are obtained as follows: r.k/ D y.k/  xO kjk1 ;

(44.27)

OT where the filter’s last-propagated state estimate vector is xO kjk1 D ŒpO qO rO ˛O ˇ .kjk1/ T and the measurement vector is y.k/ D Œp; q; r; ˛; ˇ.k/ . The continuous system dynamics matrix for the no-fault filter Fnf .k/ evaluated at time step k can be explicitly derived from the nonlinear model detailed in Sect. 44.2 as 3 2 S bŒIzz CLˇ Ixz CNˇ  0 qN D1 7 6 7 6 F S cC N M˛ q N 0 7 6 1 Iyy 7 6 6 S bŒIxx CNˇ I xzCLˇ  7 Fnf .k/ D 6 ; (44.28) 7 0 q N D1 7 6 7 6 7 6 0 1 0 VT S CZ˛ 0 2m 5 4 VT S CY 1 0 0 1 0 2m xO .kjk/;q.k/ nf

with the submatrix F1 defined as 2

Izz S b 2 CLpQ N 2D1 VT q



N1 D1 q

6 6 Ixx Izz I xz F1 D 6 6  Iyy r C 2 Iyy p 4 S b2 C I N3  2D1LVpQT xz qN C D q 1

N1 D1 p

C

N2 D1 r

S cN2 CM qQ qN 2VT Iyy N3 D1 p

C

N1 D1 r

.Izz CLQr Ixz CN rQ /S b 2 qN 2D1 VT Izz  IxxIyy p



C

N2 D1 q

2Ixz r Iyy

S b 2 ŒIxz CLQr CIxx CN rQ  qN 2D1 VT

C

N1 D1 q

3 7 7 7; 7 5 (44.29)

2 2 2 where N1 D Ixz .Ixx  Iyy C Izz /, N2 D Iyy Izz  Ixz  Izz2 , N3 D Ixz  Ixx Iyy C Ixx 2 and D1 D Ixx Izz  Ixz .

44 Actuator Fault Detection in UAVs

1089

The discrete transition matrix for the no-fault filter is calculated with nf;k D I C Fnf .k/Ts . The control input matrix of the no-fault filter Gnf .k/ is computed as follows: 2

S bIzz CLa1 D1

6 N M a1 6 S cC 6 Iyy 6 S bIxz CLa1 Gnf .k/ D qNk 6 6 D1 6 6 4 0 0

S bIzz CLa2 D1

S bIzz CLe1 D1

S bIzz CLe2 D1

S bIxz CNır D1

S cC N M a2 Iyy

S cC N M e1 Iyy

S cC N M e2 Iyy

0

S bIxz CLa2 S bIxz CLe1 S bIxz CLe2 D1 D1 D1

0 0

0 0

0 0

S bIxx CNır D1

0 0

3 7 7 7 7 7: 7 7 7 5 (44.30)

The discrete control input matrix for the no-fault filter is Gnf;k D Gnf .k/Ts .

44.4.4 Augmenting the State Vector with the Faulty Actuator Parameter ıNi The state vector of the i th filter is augmented to monitor the occurrence of the i th actuator fault. The deflection of the failed actuator is included in the state vector in a way to be estimated by the EKF. Therefore, the state vector for each filter i is x zi D N : ıi 

(44.31)

The augmented state vector leads to the following state space equations for each filter i : zi .k C 1/ D fzi .zi .k/; ı.k// C wk ; yi .k/ D h.zi .k// C vk ;

(44.32)

Using the above equations, the linearized system evaluated at each sampling time can be written as 

  .0;i / x.k/ F.k/ Gi .k/ G .k/ ı.k/ ; C ıNi .k/ 0 1 0  x.k/ : y.k/ D ŒH 0 N ıi .k/

x.k C 1/ ıNi .k C 1/





D

(44.33)

where G.i / represents the i th column of G and G.0;i/ represents the matrix G with its i th column set to zero.

1090

G. Ducard

44.4.5 Designing the EKF for the Case of a Failure on Aileron 1 This section provides an example of how to derive the matrices for the EKF corresponding to the scenario of a lock-in-place or floating actuator failure. Aileron 1 is taken as an example. The system dynamics matrix for the aileron1-fault filter Fıa1 .k/ is explicitly derived from the nonlinear model as 2 6 6 F 1 6 6 6 6 Fıa1 .k/ D 6 6 60 1 0 6 6 6 0 0 1 4

0 S cC N M˛ qN Iyy

0 VT S CZ˛ 2m

0

0 0 0

S bŒIzz CLˇ Ixz CNˇ  qN D1

S bIzz CLa1 D1

0

S cC N M a1 Iyy

qN

3

7 qN 7 7 7 S bIxz CLa1 7 S bŒIxx CNˇ I xzCLˇ  7 q N q N D1 D1 7 7 7 0 0 7 7 VT S CY 1 7 0 5 2m

0

0

1

:

zO 1 .kjk/

(44.34) The discrete transition matrix for the aileron1-fault filter is calculated with k;ıa1 D I C Fıa1 .k/Ts . The control input matrix of the no-fault filter Gıa1 .k/ is computed as follows: 2

0

6 6 60 6 6 6 Gıa1 .k/ D 6 0 6 6 60 6 40 0

S bIzz CLa2 D1

qN

S bIzz CLe1 qN D1

S bIzz CLe2 D1

qN

S bIxz CNır D1

qN

3

7 7 7 7 7 S bIxz CLa2 S bIxz CLe1 S bIxz CLe2 S bIxx CNır qN qN qN qN 7 : (44.35) 7 D1 D1 D1 D1 7 7 0 0 0 0 7 7 5 0 0 0 0 0 0 0 0 q.k/ N S cC N M a2 Iyy

qN

S cC N M e1 qN Iyy

S cC N M e2 qN Iyy

0

The discrete control input matrix for the aileron1-fault filter is Gıa1 ;k D Gıa1 .k/Ts . All the other filters monitoring the other actuators are designed in a similar way.

44.4.6 Actuator Fault Isolation 44.4.6.1 Hypothesis Testing A hypothesis testing algorithm uses the residuals and the state error covariance matrix from each EKF to assign a conditional probability to each fault scenario.

44 Actuator Fault Detection in UAVs

1091

The estimated state vector of the system is the sum of the state vector of each EKF weighted by its corresponding probability xO Œk D

X

xO i Œk  pi Œk;

(44.36)

i

where xO i Œk is the state estimate computed by the EKF that assumes the fault scenario i . The index i covers all the fault scenarios implemented, including the no-fault case. The term pi Œk denotes the probability that the i th fault scenario is occurring. Now, the main difficulty lies in the online computation of the probability pi Œk. In order to determine which fault scenario the actual plant is the closest to, it is needed to consider the measurement data from the sensors. The last available measurement vector is yŒk, sometimes also written yk in the following. The sequence of the last measurement vectors is defined as Yk D fyk ; yk1 ; yk2 ; : : : ; y0 g. The fault probability pi Œk can be expressed as the a posteriori conditional probability pi Œk D p. D i jYk /, that is, the probability that the actual plant can be categorized in the scenario i given the sequence of the last measurements Yk . The Bayes’ law states that pŒYk j D i pŒ D i  ; (44.37) pi Œk D pŒ D i jYk  D pŒYk  where the probability pŒYk  can be decomposed as pŒYk  D pŒYk j D 1   pŒ D 1  C    C pŒYk j D N   pŒ D N  ; D

N X

pŒYk j D j   pŒ D j :

(44.38)

j D0

Combining (Eq. 44.37) and (Eq. 44.38) yields pi Œk D pŒ D i jYk  D PN

pŒYk j D i pŒ D i 

j D0 pŒYk j

D j   pΠD j 

;

(44.39)

with N being the number of different scenarios under consideration. In order to make a recursive form appear in the probabilities, the measurement data sequence Yk is rewritten as the sequence fyk ; Yk1 g: pŒYk j. D j / D pŒyk ; Yk1 j. D j /    D p yk j Yk1 ;  D j  pŒYk1 j. D j /   D p yk j. D j ; Yk1 /  pŒ. D j /jYk1    D p yk j. D j ; Yk1 /  pj Œk  1: Using the result (Eq. 44.40) in (Eq. 44.39) yields

(44.40)

1092

G. Ducard

p Œyk j. D i ; Yk1 /  pi Œk  1  pŒ D i  :   j D0 p yk j. D j ; Yk1 /  pj Œk  1  pŒ D j  (44.41)

pi Œk D pŒ. D i /jYk  D PN

Since a fault may occur at any time, regardless of which actuator may fail, the same probability is assigned to all the scenarios, that is, pŒ D j  D 1=N for j D 1; : : : ; N . Therefore, the equation above simplifies to the following recursive expression: p Œy D yk j. D i ; Yk1 /  pi Œk  1 : (44.42)   j D0 p y D yk j. D j ; Yk1 /  pj Œk  1

pi Œk D pŒ D i jYk  D PN

Remarks: • By examining the probabilities, the “health status” of the system can be determined. It is either the no-fault case or a case where an actuator is lockedin-place or floating. An actuator fault is declared valid if the corresponding fault probability exceeds 90 % for a certain amount of time. A fault is declared removed when the corresponding fault probability is below 5 % for a certain amount of time. • One may notice that the denominator of (Eq. 44.42) corresponds to the sum of each scenario probability’s numerator, such that fault probabilities add up to one. • In practice, in order to prevent the possibility that the recursive computation of the fault probability in (Eq. 44.42) stays at zero forever as soon as the probability reaches zero, the lower bound of each probability is set to 0.001. • This method that uses probabilities for fault isolation is sometimes called a Bayes classifier (see Isermann (2006), Chap. 16).

44.4.6.2 Gaussian Conditional Probability Density This section provides an explicit formula for the term p Œy D yk j. D i ; Yk1 /, which corresponds to the probability of obtaining the measurement data yŒk at time tk D kTs , assuming the scenario i exists and given the sequence of the last measurements Yk1 . The probability density is chosen to be a Gaussian function (Maybeck 1994) with its characteristic bell-shaped curve according to the following formula: T

p Œy D yk j. D i ; Yk1 / D i Œke ri Œk

†1 i Œkri Œk=2

;

(44.43)

1 with i Œk D .2 /m=2 j† 1=2 , where j : : : j denotes the determinant of the matrix, m i Œkj represents the measurement dimension, and †i Œk is the residual covariance matrix calculated at time step k by the i th EKF. The term ri Œk corresponds to the residuals of the i th EKF, when the measurement update step occurs according to ri Œk D yk  h.Oxi .kjk  1//:

44 Actuator Fault Detection in UAVs Fig. 44.7 Conditional probability density in the scalar case

1093 probability density: f (ri | (q = qi ,Yk-1))

0.4 p(y = yk | (q = qi ,Yk-1))

0.35 0.325 0.3 0.25 0.2 0.15 0.1 0.05 0 −8

−6

−4

−2

0

yˆi k

2

4

y k

ri

y

yˆ i y

ri k

Intuitive Explanation of the Probability Density In the case of a single-input single-output problem, in which the state and measurement vectors reduce to scalars, the measurement data h.xO i .kjk1// that are expected according to the model can be seen as the mean value of the measurement data computed by the i th EKF, that is, yOi Œk in Fig. 44.7. The width of the conditional Gaussian density is governed only by the covariance term †i Œk. Figure 44.8 shows the shapes of the Gaussian function for several standard deviations  2 D †.jj /. The residual ri Œk determines the relative position of the peak of the probability density with the actual measurement yk . In the multivariable case, which applies in this FDI system, the fault probability p Œy D yk j. D i ; Yk1 /

(44.44)

is given by f.ri D ri Œkj. D i ; Yk1 //, with the probability density defined as a function of the residual ri with f .ri j. D i ; Yk1 // D

1 T 1  e ri †i Œkri =2 : .2 /m=2 j†i Œkj1=2

(44.45)

Therefore, the filter that corresponds to the fault scenario produces an estimate for the measurement vector yO i Œk D h.Oxi .kjk  1// very close (apart from noise) to the actual value of the measurement data vector yŒk. The residual ri Œk D yk  h.Oxi .kjk  1// will be small and close to zero. This means that the corresponding probability p Œy D yk j. D i ; Yk1 / is the highest for the filter matching the fault scenario.

1094

G. Ducard

1.4

(a)

σ=0.3 σ=0.5 (c) σ=0.7 (d) σ=1 (e) σ=2 (a)

1.2

(b)

1 (b)

0.8

(c)

0.6

(d)

0.4

(e)

0.2

0 −8

−6

−4

−2

0

2

4

6

8

Fig. 44.8 Gaussian functions for zero-mean value and several standard deviations 

By examining the probabilities computed with (Eq. 44.46), the health status of the system is determined; it is either the no-fault case or actuator-failure case: p Œy D yk j. D i ; Yk1 /  pi Œk  1 : (44.46)   j D0 p y D yk j. D j ; Yk1 /  pj Œk  1

pi Œk D pŒ D i jYk  D PN

Remarks: • The hypothesis testing uses a Gaussian density function, which assumes that the residuals from the EKFs are Gaussian distributed. When this is not the case, there is a little inconsistency with the application of the theory. However, the assumption that these residuals are Gaussian distributed is still reasonable, especially when the aircraft dynamics are slow. • The reason why a Gaussian distribution is used to describe the probability density of the current measurement to take on the value yk based on the fault hypothesis i and the previous measurements Yk1 is to make the mathematics tractable. As mentioned in the book by Maybeck (1994), “the Kalman filter, which propagates the first and second-order statistics [mean and variance of a process], includes all information contained in the [Gaussian] conditional probability density, rather than only some of it, as would be the case with a different form of density.” If another probability density function was known, then (Eq. 44.45) could be changed accordingly.

44 Actuator Fault Detection in UAVs

1095

44.4.7 Simulation Results of the EMMAE-FDI with No Supervision System 44.4.7.1 Simulation Conditions In order to obtain realistic simulations, the sensor measurement data are corrupted on purpose with zero-mean white Gaussian noise corresponding to typical specifications of low-cost sensors. For the turn rate sensors, the standard deviation is p;q;r D 5ı /s = 0.0873 rad/s, which corresponds to a noise covariance of †p;q;r D 0:0076  I3 (rad2 /s2 ). For the airflow angle sensors, the noise standard deviation is ˛;ˇ D 2ı =0.0349 rad (†˛;ˇ D 0:0012  I2 [rad2 ]). The airspeed sensor noise has a standard deviation of VT D1 m/s (†VT D 1 m2 =s2 ). Poor sensor quality adversely affects the FDI reliability. When there is little excitation and when the faulty actuator deflection is close to the trim conditions, an actuator failure becomes even more difficult to detect. Indeed, due to large sensor noise, the control signals (see Fig. 44.9) become noisy as well, which reduces the difference between the actual faulty actuator deflection and its corresponding control signal. The larger this difference, the easier it is to detect the fault. The EKF process noise covariance matrix and the sensor noise covariance matrix are selected as follows: Rw D 0:002  I5 and Rv D diagŒ0:1  I3 0:02  I2 . 44.4.7.2 Scenario The scenario to test the fault detection method is chosen to put the FDI in the most difficult conditions, which are those of minimum excitation of the system. This is achieved when the aircraft is flying straight and level (no maneuvers, no wind) at a constant speed of 30 m/s. The actuator faults are simulated by blocking the control surfaces close to the trim deflections corresponding to straight level flight conditions, because those faults are harder to detect and to estimate. For example, the ailerons and the rudder are intentionally made to fail close to the neutral deflection (0ı ), and the elevators are made to fail close to 2ı . Fault detection with the EMMAE method is tested on a 6 degree-of-freedom nonlinear aircraft model of a UAV used in M¨ockli (2006) and Ducard (2009). Simulations were performed in MATLABR /SimulinkR on closed-loop control architecture, with a nonlinear autopilot which regulates the speed, altitude, and the attitude of the aircraft. The actuator configuration of the aircraft is depicted in Fig. 44.2. The EMMAE-FDI system is therefore composed of six EKFs, one for monitoring the no-fault case, two for monitoring each aileron (one on each wing), two EKFs for monitoring each of the two independent elevators, and one EKF for the rudder. As depicted in Fig. 44.9, a sequence of consecutive faults is generated. From t = 10. . . 40 s, aileron 1 fails and is locked at 1ı deflection; for t = 70. . . 100 s, aileron 2 fails and is “floating” between the two positions 1ı and 1ı in a square-wave fashion. For t = 130. . . 160 s, the rudder fails and is locked at 1ı .

1096

G. Ducard

Aileron2

5 0 −5 −10

Rudder

5 0 −5

30 20 10 0

Thrust

4 2 0 −2 −4

Elevator1

2 0 −2

Elevator2

Aileron1

control signal in deg real deflection in deg

70 60 50

50

100

150

200

250

300

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

Time (s)

Fig. 44.9 Control signals and actual actuator deflections

For t = 190. . . 220 s, elevator 1 gets locked at 0:5ı , and finally, for t = 250. . . 280 s, elevator 2 is floating between two uncontrolled positions 1ı and 3ı in a squarewave fashion. After this sequence of faults, the aircraft continues to fly straight and level, and no more faults are introduced.

44.4.7.3 Comments on the Simulation Results Figure 44.10 shows the results obtained by the FDI system after the sequence of faults. The top plot labeled “no fault” has a probability of 1 when the EMMAE-FDI system does not detect any fault in the aircraft. An actuator fault is declared valid if the corresponding fault probability exceeds 90 % for a certain amount of time. A fault is declared removed when the corresponding fault probability is below 5 % for a certain amount of time. When aileron 1 fails at t = 10 s, the “No-Fault” filter needs about 6 s for its probability to go down to almost 0, which means that a failure

44 Actuator Fault Detection in UAVs

1097

Fault probability

Pb af5

Pb af4

Pb af3

Pb af2

Pb af1

No Fault

Expected probability 1 0.5 0

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150 Time (s)

200

250

300

1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0

Fig. 44.10 Probabilities from each filter of the EMMAE-FDI after a sequence of faults (no supervision system)

occurred somewhere. After the aileron 1 fault is introduced, both probabilities for aileron 1 and 2 to fail (Pbaf1, Pbaf2) start rising at t = 11.5 s. At t = 17 s, the FDI system begins to distinguish between the two ailerons which one has failed, and Pbaf2 returns to zero while Pbaf1 rises up to 90 % at t = 34 s. Therefore, it took 24 s for the FDI system to indicate that aileron 1 experiences a failure. At t = 40 s, the fault of aileron 1 is removed, and the actuator behaves normally again. Figure 44.10 shows that the probability Pbaf1, indicating whether aileron 1 fails, decreases slowly and reaches 0 again after 10 s, while the “No-Fault” probability rises accordingly. It thus takes 10 s for the FDI system to indicate that the fault has been removed. As Fig. 44.9 shows, at t = 70 s, aileron 2 fails and has an uncontrolled squarewave motion between 1ı and 1ı . Figure 44.10 shows that the FDI system takes 20 s to detect such a failure. An ambiguity exists between the two ailerons for a few

1098

G. Ducard

seconds before the probability Pbaf2 finally reaches 90 % at t = 90 s. At t = 100 s, the fault of aileron 2 is removed, and this actuator behaves normally again. However, the FDI system is not capable of detecting quickly that the fault has been removed. It takes about 8 s to do so. For the rudder, the fault is introduced at t = 130 s and is isolated by the FDI system at t = 131 s when Pbaf5 exceeds 90 %. The fault removal is detected in less than 5 s. It takes less time to isolate a rudder fault and to detect its removal because there is only one rudder, unlike the other actuators which are redundant (two ailerons, two elevators). Therefore, a malfunctioning rudder cannot be compensated by a redundant rudder, thus resulting in no actuator-fault ambiguity. The introduction of the elevator 1 fault at t = 190 s is isolated by its corresponding filter (probability signal Pbaf3 exceeding 90 %) at t = 208 s. After the fault removal, the FDI system takes 8 s to indicate the fault removed. The behavior of elevator 2 is similar to the behavior of elevator 1. Finally, after the last fault has been removed, 5 s are needed by the FDI system to slowly build up probability in the “No-Fault” filter to indicate that no more faults are present in the system.

44.4.8 Remarks on the First Attempt to Use the EMMAE-FDI System • The results plotted in Fig. 44.10 indicate that the current implementation of the method is able to detect the fact that a failure occurred, even in a very low-excitation case, but it could not tell quickly and reliably which actuator experienced the fault. In cases of redundant actuators having the same influence on the aircraft aerodynamics, the EMMAE method has difficulty quickly resolving ambiguities between redundant actuators when they cannot be properly excited. • It appears that failures of actuators near trim deflection are more difficult to detect and isolate. • Moreover, whenever a fault is removed, the EMMAE method alone requires a long time to detect that fact. It is critical, however, that the FDI quickly detects the removal of a failure or quickly recognizes that a false alarm has been triggered due to possible external perturbations, such as strong wind gusts. • Finally, it is critical that the probabilities quickly reach the “expected values” which correctly describe the fault scenario. Indeed, the estimated state vector of the system, which is the sum of the state vector of each EKF weighted by its corresponding probability, must be sufficiently correct and accurate if this state estimate is fed back to the controller.

44.4.9 Techniques to Improve Actuator Fault Diagnosis This section describes some techniques added to enhance the performance of the FDI system, when there is very low excitation of the system, particularly during steady level flight. In order to improve the speed and the accuracy of the fault isolation, a supervision module is designed whose tasks are detailed below.

44 Actuator Fault Detection in UAVs

1099

xˆ

Supervisor Probability actuator 1 failed

New upper/lower limits for actuator 1

Estimated deflection of actuator 1 New upper/lower limits for actuator i Excite actuator 1 Excite actuator i

Control Allocation Actuator deflection limits Virtual control inputs

xˆ Ref

Probability actuator i failed Estimated deflection of actuator i

Controller

CL Cu CM CN δ Thrust

p1 δˆ1

FDI: EMMAE

pi δˆi

δexi δc

Actuators

Aircraft

Fig. 44.11 EMMAE-FDI system in a reconfigurable flight control system

44.4.9.1 Design of an Active Supervision Module (Supervisor) The supervision module shown in Fig. 44.11 is designed to monitor probability signals from the FDI. If an actuator-failure probability exceeds a certain threshold for some time, then the supervisor is designed to superimpose an artificial control signal on the corresponding actuator. If an actuator fails, the additional signal will have no effect on the aircraft dynamics, but it will help the FDI confirm more quickly the failure of this actuator. On the other hand, if the actuator has not actually failed, the aircraft will respond according to the additional signal, and the FDI will then remove the fault assigned to this actuator. If the corresponding fault probability falls below a certain threshold for a defined period of time, the supervisor removes the superimposed signal. The excitation signals can also be adaptively controlled within certain limits, for example, from 1ı to 4ı , with the function ıex i .t/ D Œ1ı C3ı .1pi .t// cos.2 fi t/, see Fig. 44.12. Note that an actuator is only excited when its corresponding fault probability pi exceeds 5 %. Most of the time, only one actuator is excited. If several actuators are to be excited simultaneously, better detection performance may be expected if the excitation signals to each actuator are independent and uncorrelated. The frequency of the signal is to be chosen within the range of the aircraft bandwidth; in the case of this chapter, fi D 1 Hz. The excitation signal has an adaptive amplitude dependent on the probability pi of actuator i to have failed. In this way, when the probability pi is low, the excitation amplitude is large and vice versa. Simulation results show that this adaptive amplitude for the excitation signal efficiently improves the accuracy and speed for fault isolation compared with a fixed amplitude excitation signal. This adaptive amplitude ensures that the actuator is excited as little as possible, but still enough to isolate the fault or to remove a false alarm. Figure 44.12 shows the practical implementation of the excitation signal generator and provides an example for aileron 1.

1100

G. Ducard

Check Aileron1 p1

pthreshold ?

No

Yes Count_Ail1_failed++

Count_Ail1_failed > Count Threshold?

Count_Ail1_failed– –

No

Yes

No

Count_Ail1_failed < 0? Yes

Count_Ail1_failed = Count Threshold

Count_Ail1_failed = 0

δex_1(t) = 1 + 3 [1– p1] cos(2π f1t )

δex_1(t) = 0

exit

Fig. 44.12 Generation of the actuator excitation signal (example for aileron 1)

This method is therefore a systematic way of testing each triggered failure alarm, to confirm it or to remove it, hence making the FDI more robust. Whereas other proposed schemes (Ni 2001; Azam et al. 2005) suggest having the aircraft perform a “health-check maneuver” or “diagnostic maneuver” as soon as a failure is detected, in the method presented in this chapter, the actuator is directly excited by the supervision module rather than by the aircraft autopilot, yielding much faster and more accurate fault isolation.

44.4.9.2 Fault Detection Performance of the EMMAE-FDI with the Supervision System Figure 44.13 shows how the aileron faults are accurately detected and isolated in less than 5 s. The rudder fault is isolated after 1 s. The elevator faults take longer (about 9 s) to be isolated. However, the removal of all the faults is detected in less than 5 s. Furthermore, there is no more ambiguity or false detection among the actuator faults. The results shown in Fig. 44.13 compared to those of Fig. 44.10 indicate that the performance and the robustness of the EMMAE-FDI system have greatly improved due to the supervision module. It is recalled that these results are obtained in the most difficult conditions for the FDI system. Indeed, there is no external disturbance such as wind gusts, the aircraft flies straight and level, and the actuators fail close to their trim deflection.

44 Actuator Fault Detection in UAVs

1101

No Fault

0.5

Pb af5

Pb af4

Pb af3

Pb af2

0.5

Pb af1

Fault probability Expected probability 1 0

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150 Time (s)

200

250

300

1 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0

Fig. 44.13 Probabilities from each filter of the EMMAE-FDI after a sequence of faults (with the supervisor)

44.4.10 Computational Complexity of the EMMAE-FDI Let us denote by N the number of actuators that are monitored by an FDI system. In the classical implementation of an MMAE-FDI, for each actuator, k filters are to be designed for k different possible positions of the failed actuator. Therefore, kN C 1 KFs are required (+1 refers to the no-fault filter). If the appearance of a second fault is to be checked as well, a new bank of KFs has to be reloaded, based on the knowledge of the first fault that occurred. In total N k C1 C.N 1/kN D N 2 k C1, KFs must be designed. One major advantage of the EMMAE-FDI method presented in this chapter over classical MMAE schemes is that it requires only one filter to completely monitor one actuator. Any possible actuator-fault scenario is taken into account by only one filter. Therefore, for the monitoring of a single actuator fault, only N C 1 filters are

1102

G. Ducard

required with the EMMAE-FDI method. For the monitoring of a second actuator fault with the EMMAE-FDI method, no other bank of filters has to be loaded. Indeed, if actuator i fails, it suffices to feed all the other filters with the estimate of the faulty control surface deflection ıONi instead of the input ıi . Thus, again only N C1 filters are needed with the EMMAE-FDI system to detect and isolate a second fault after a first actuator fault has been introduced. As a simple example, the UAV described in Sect. 44.2 is equipped with five actuators. A classical MMAE scheme designed for 3 possible faulty deflections per actuator requires 16 filters, whereas the EMMAE method needs only 6 filters for any lock-in-place and floating actuator fault scenarios. Finally, compared to other works (e.g., Ducard and Geering (2006)) with larger O T minimizes the number state vectors, the choice of the state vector x D Œp q r ˛ ˇ ı of relevant state elements for the satisfactory operation of the filters, thus limiting the number of computations required for the EKFs.

44.4.11 Realistic Flight Scenario and Conclusions More simulation results are reported in Ducard (2009), where the EMMAE FDI system is tested in a realistic flight scenario, in which the UAV takes off, tracks a predefined trajectory, and follows an altitude and a speed reference profile. First, the simulation is done without wind, and in a second phase, Dryden Wind turbulences are included in the simulation to test the behavior of the FDI system in the presence of external disturbances. The speed profile is chosen to cover a significant range of the aircraft speed in order to test the performance of the fault detection system at different operating conditions. The measured airspeed VT is corrupted by noise, like all the other measurement signals defined in Sect. 44.4.7.1 that are used in the FDI system. Also, the altitude reference signal is chosen such that the aircraft has a typical vertical motion during a short mission. The results show that the EMMAE-FDI algorithm, combined with an active supervision module, offers fast and accurate fault detection and isolation. Moreover, the addition of the estimation of the faulty control surface deflection in the state vector makes the method applicable for actuator failures such as frozen or floating at an arbitrary position. Only one filter is needed to monitor the health of one actuator. The filters used in the EMMAE-FDI are EKFs, which provide nonlinear state estimations at any flight-operating condition. An active FDI technique is developed, which generates appropriate artificial excitation of the aircraft when needed. In Ducard (2009), an additional filtering stage for the fault-probability signals has been designed to enhance the robustness of the diagnosis, even in the event of severe wind turbulence. The whole system has been demonstrated using nonlinear simulations of a realistic flight scenario. The FDI system was shown to be capable of handling two simultaneous actuator failures without increasing the computational load. Finally, when a fault is clearly isolated, the faulty actuator deflection estimate

44 Actuator Fault Detection in UAVs

1103

can be advantageously used to modify online the settings of a control allocator, making the whole system suitable for flight control reconfiguration without any change in the initial controller and any additional actuator position sensor.

44.5

A Single Model Active (SMAC) FDI System

In both the MMAE and EMMAE methods, actuator fault detection and isolation is done through the computation of fault probabilities. Although the EMMAE is more computationally efficient and reliable than the MMAE method, it still requires a bank of EKFs, where each EKF monitors one actuator in the system. This section presents an FDI system which is single model based and thus does not employ a bank of filters. This FDI system employs a single EKF to generate residuals, which indicate the deviation of the monitored system from its normal behavior (fault detection). The residuals are further used to assess the origin of the system’s behavior discrepancy. In order to isolate which actuator experiences a fault, an innovative mechanism that artificially excites actuators has been designed. The method presented next has been named single model active fault detection and isolation (SMAC-FDI) system (Ducard and Geering 2010). This section is organized in order to (1) present an innovative architecture of an FDI system based on a single filter for the generation of residuals independently on the fault scenario; (2) describe an active strategy that searches in the residuals of this single filter the signature of an actuator fault and to test the actuators in a systematic way; (3) show a new procedure to isolate actuator fault that is not based on residuals manipulations but on the observation of control signals only; (4) show that the SMAC-FDI method is highly computationally efficient and, therefore, can run on microcontrollers with low computational resources; and (5) demonstrate the performance of this simple method in simulation. Figure 44.14 shows the architecture of the SMAC-FDI system. It is composed of five main subsystems: a residual generator, a fault detector, an excitation signals generator (ESG), an actuator health evaluator (AHE), and a fault isolator. Each of these subsystems is described in the following:

44.5.1 Residual Generator The residual generator monitors the control signals sent to actuators and the behavior of the aircraft through the reading of selected measurement of the aircraft’s dynamics. In the SMAC-FDI system, the turn rates p, q, and r are measured and compared to the expected turn rates p, O q, O and rO , which are obtained using a model of the aircraft. This model receives the same actuator commands as the real aircraft. A filter is used to take into account model uncertainties and sensor measurement data noise. This filter is an EKF, whose useful output – for FDI purpose – is the vector of residuals !Q k defined as follows:

1104

G. Ducard Aircraft c

δc

Guidance System

Controller

Actuators

δT

y

Aircraft Dynamics

+ SMAC FDI System Residuals Generator Excitation Signals Generator (ESG)

δs

Actuator Health Evaluation (AHE)

˜ ω

˜ LP1(ω) Fault ThereIsaFault Detector

Λ

Fault diagnostic Fault Isolator Actuator(s) Select

Fig. 44.14 Architecture of the SMAC-FDI system

!Q k D yk  xO kjk1

3 2 3 2 p  pO.kjk1/ pQ 7 6 D 4 qQ 5 D 4 q  qO.kjk1/ 5 ; rQ r  rO.kjk1/

(44.47)

where the measurement vector at the discrete-time instant k is yk D Œp q rT and the filter’s state vector is the estimate of the turn rates. These residuals indicate whether the aircraft has dynamics that deviate from the model-expected dynamics. If the residual values are higher than a certain threshold, it is most likely that an actuator fault is present in the system. The EKF used in this residual generator has the structure described in Sect. 44.3. The matrices and vectors involved in this filter are described next.

44.5.1.1 Discrete Transition Matrix for the SMAC-FDI System The discrete transition matrix k is computed as k D I3 C F.k/Ts . The continuous transition matrix F.k/ 2 =T1

No

Yes Isolate the First Faulty Sensor

Select New C and Hy Block (2) Compute the Parity Vector

FD >=T2

No

Yes Isolate the Second Faulty Sensor

Select New C and Hy

Begin the Wavelet Transformation Block (3) Compute the Parity Vector

FD >=T3

No

Yes In-lane Monitoring Method

Isolate the Third Faulty Sensor

END

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

1195

In this way, triple successive faults can be treated during the flight, and the UAV can perform its mission continuously. A measurement matrix C , which is originally 6-by-3, is reduced to a 5-by-3 matrix after the first fault isolation and is downsized to a 4-by-3 matrix after the second fault isolation. A parity matrix Hy , which is originally 3-by-6, is reduced to a 2-by-5 matrix after the first fault isolation and is reset to a 1-by-4 matrix after the second fault isolation.

47.3

Hardware Configuration for the Flight Test

The skew-configured IMU and the fault signal generation board were constructed to verify the performance of the proposed hybrid FDI scheme. The fault signal generation board produces virtual fault signals instead of real fault signals. The hybrid FDI algorithm is processed in the flight control computer. The performance of the proposed FDI is verified through ground and flight experiments. Figure 47.4 describes the system configuration and the test procedure. The details of the design/calibration procedures and functions of the skew-configured IMU, the fault signal generation board, and the fixed-wing UAV are described below.

47.3.1 Skew-Configured Inertial Measurement Unit The primary IMU of the experimental UAV system is the Crossbow AHRS400. The secondary IMU is the XSENS MTi-68, which is a small and low-cost IMU compared to the primary IMU. The secondary IMU is mounted on the inclined plane with respect to the aircraft body coordinate, whereas the primary IMU is aligned with the aircraft body coordinate. The inclination angle of the secondary IMU is determined by optimizing the navigation index and the fault detection performance index. A detailed optimization procedure can be found in the references (Park 2004). Figure 47.5 illustrates the designed structure of the mounting surface for the secondary IMU. As shown in Fig. 47.5, the primary IMU is aligned with the aircraft body coordinate .X1 ; Y1 ; Z1 /, whereas the secondary IMU is aligned with the slope plane .X2 ; Y2 ; Z2 /. As a result, six inertial sensors are configured in a cone shape. With this configuration, the measurement matrix C .k/ and the parity matrix Hy .k/ in Eqs. (47.12) and (47.13) are analytically calculated as follows: 2

3T 1 0 0 0:6667 0:6667 0:3333 C .k/ D 4 0 1 0 0:3333 0:6667 0:6667 5 0 0 1 0:6667 0:3333 0:6667

(47.34)

1196 Fig. 47.4 System configuration and test procedure

S. Yoon et al.

Skew-Configured Inertial Measurement System

XBOW AHRS400

Fault Signal Generation Board

XSENS MTi-68

Normal Sensor Signal

Fault Scenario Command Faulty Sensor Signal

Flight Control Computer

HILS and Flight Test

Hybrid Fault Detection and Isolation using parity space method & in-lane monitoring method

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

1197

76 mm

Fig. 47.5 Configuration of two inertial measurement units

114 mm

X2

Z2

Z1

Y2

Y1

76

m

m

38 mm

X1

m

76 m

3 0:4714 0:2357 0:4714 0:7071 0:0000 0:0000 Hy .k/ D 4 0:4714 0:4714 0:2357 0:0000 0:7071 0:0000 5 0:2357 0:4714 0:4714 0:0000 0:0000 0:7072 2

(47.35)

In practice, the skew-configured IMU contains small alignment errors in C .k/ and Hy .k/. Therefore, the assembled IMUs should be calibrated with an initialization experiment, which compensated for the bias of the skew-configured IMU with a 5-min history of each signal in the stationary condition. The sensor system was placed on a single-axis rate table, Ideal Aerosmith 1291BR, and was rotated about three different axes .X1 ; Y1 ; Z1 / with six constant angular velocities: 100, 50, 20, 20, 50, and 100 deg/s. Using the signals from these experiments, the measurement matrix C .k/ in Eq. (47.12) with w .k/ D 0 and f .k/ D 0 was computed as C .k/ D z .k/ xr .k/ (47.36) 

where z .k/ is a 6-by-18 output matrix of the skew-configured IMU and xr .k/ is a pseudo inversion of an 18-by-3 reference input matrix. The columns of z .k/ comprise 18 rotating conditions; these are the rotating axis X1 with 100 deg/s, the rotating axis Y1 with 100 deg/s, the rotating axis Z1 with 100 deg/s, and so on. As a result of this initialization experiment, the calibrated C .k/ and Hy .k/ were obtained as follows: 2

3T 1:0018 0:0031 0:0063 0:6885 0:65889 0:3041 C .k/ D 4 0:0030 1:0018 0:0121 0:3163 0:6554 0:6865 5 0:0021 0:0028 0:9999 0:6555 0:3660 0:6635

(47.37)

1198

S. Yoon et al.

3 0:4845 0:2295 0:4613 0:7070 0:0012 0:0001 Hy .k/ D 4 0:4656 0:4644 0:2595 0:0002 0:7073 0:0012 5 0:2185 0:4809 0:4700 0:0007 0:0017 0:7072 2

(47.38)

47.3.2 Fault Signal Generation Board In practice, it is difficult to make the inertial measurement sensors partially fail during flight. A fault signal generation board is considered to produce virtual fault signals instead of real fault signals. The fault signal generation board receives normal sensor signals from the skew-configured IMU and adds virtual fault signals to normal sensor signals according to the failure scenario. The modified sensor signals are transmitted to a flight control computer that will perform FDI. In this way, inertial measurement sensors can be virtually broken by the fault generation board. The fault signal generation board was designed with digital signal processing units: Texas Instrument’s TMS320LF2407A and TL16C754BPN. It has 5 serial communication ports and 20 digital input/output ports. The fault signal generation board has a communication port and six light-emitting diodes on the outside. Through its 15-pin communication port, 2 IMUs are connected to the fault signal generation board. In addition, commands from the ground station and the corresponding virtual sensor signals are communicated through this port.

47.3.3 Experimental Unmanned Aerial Vehicle System The performance of the proposed FDI approach was verified with an experimental UAV, Durumi 2, developed and operated by the Korea Aerospace Research Institute. This UAV has performed various flight experiments for real-time system identification and fault-tolerant control systems (Park et al. 2006). Its wing span is 4.8 m, and its body length is 2.7 m as shown in Fig. 47.6. Its total weight is 37 kg, including a maximum payload weight of 12 kg. The maximum speed is 120 km/h, and its stall speed is 55 km/h. In addition, all of the control surfaces are fractioned into two pieces able to move independently for a fault-tolerant mission. The skew-configured IMU was loaded on the payload of this experimental UAV as shown in Fig. 47.7. A hexahedron in the lower right corner is the primary IMU, the Crossbow AHRS400. An upper flat unit on the slope is the redundant secondary IMU, XSENS MTi-68, which is smaller than the primary unit. Note that the secondary unit is cheaper and has a poorer resolution. These two IMUs are directly connected to the fault signal generation board. The modified sensor output in the fault signal generation board is transmitted to the flight control computer.

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

1199

Fig. 47.6 Experimental UAV, Durumi 2

Fig. 47.7 Skew-configured inertial measurement system in the payload

47.4

HILS Test and Flight Experiments

HILS test and flight experiments were performed to evaluate the performance of the proposed hybrid FDI system. Note that the operational inspection of the FDI algorithm was done in advance by a ground test using the single-axis rate table. The fault diagnosis performance of the HILS and the flight test are discussed below.

47.4.1 HILS Test Results The objective of the ground experiment was to verify the hybrid FDI algorithm and determine the parameters of the algorithm. The bias and misalignment errors of the skew-configured IMU were calibrated with the initialization experiment. The ground test also inspected the function of the fault signal generation board.

1200

S. Yoon et al.

Table 47.1 Parameters for the hybrid fault detection and isolation Parameter Variable Maximum standard deviation of the inertial sensor signals in Eq. (47.21) Safety factors of the fault detection in Eq. (47.21) Confirmation time of the fault detection function Safety factor of the threshold in Eq. (47.33) Sampling number for the wavelet transform in Eq. (47.33)

Value

max

0.81

.k1 ; k2 ; k3 / tc kw Nw

(1.5, 1.5, 1.3) 0.25 (s) 1.5 32

As a result of various ground experiments including constant rotations, oscillations, and random vibrations, the FDI parameters for the parity equation and the maximum absolute value of the wavelet transform at each sampling time (this will be called “the discrete wavelet transform” hereafter for simplicity) were determined as summarized in Table 47.1. Figure 47.8 shows one of the ground experimental results: the sensor signals, the parity norm, the discrete wavelet transform, and the isolated sensor number, respectively. In this test, 10 deg/s bias faults successively occurred in Sensors 1, 2, and 5. The first fault occurred in Sensor 1 at 6.2 s. As shown in the second plot, the fault detection function FD .k/, the norm of the parity vector, increased when the fault occurred. That is, the first fault was detected when the residual exceeded the prescribed threshold T1 . Simultaneously, the fault isolation function FI;1 .k/ declared Sensor 1 to be faulty, as shown in the bottom plot, and therefore, Sensor 1 is isolated from the measurement. In the same way, the second fault occurring in Sensor 2 at 8.4 s was detected when the residual exceeded the threshold T2 as shown in the second plot of Fig. 47.8. Among the remaining five sensors, the fault isolation function FD;2 .k/ declared Sensor 2 to be the faulty sensor as shown in the bottom plot. Then, Sensor 2 was isolated from the measurement. At that time, the wavelet transform of the remaining four sensor signals began as shown in the third plot. Whereas the first and the second faults are detected and isolated by the parity space method, the third faults are detected by the parity space method and isolated by the in-lane monitoring method. The second plot of Fig. 47.8 shows that the residual exceeded the threshold T3 when the third fault occurred in Sensor 5 at 11.2 s. Simultaneously, the faulty Sensor 5 was detected and isolated by monitoring the wavelet transform as shown in the third and fourth plots. In brief, the results of the ground HILS test showed the viability of the proposed approach and encouraged the necessity for real flight experiments.

47.4.2 Flight Experimental Result The fault diagnosis performance of the hybrid FDI system was also verified through the flight test. The command for the type and magnitude of the fault signal is transmitted from the ground station according to the sensor failure scenario.

Parity Norm, degs/s

Sensor Signals, deg/s

47 Fault Diagnosis of Skew-Configured Inertial Sensor System 15

#1 #2 #3 #4 #5 #6

10 5 0 −5

4

5

6

7

8

9

10

11

12

6

13 residual thr1 thr2 thr3

4 2 0 4

Isolated Faulty Sensor No. Wavelet Transform, deg/s

1201

5

6

7

8

9

10

11

12

13 #1 #2 #3 #4 #5 #6 thr4

6 4 2 0 4

5

6

7

8

4

5

6

7

8

9

10

11

12

13

9

10

11

12

13

6 4 2 0

Time, sec

Fig. 47.8 Ground test result of the hybrid fault detection isolation system

In the flight test, two types of sensor fault were considered, a bias fault and a stuck fault. The bias and stuck faults are basic errors that can be generated in sensors due to both hardware and software problems. If the sensor is badly calibrated before or during the flight, the measurement can be easily biased in the decoding process. After hardware failure, the sensor output might be stuck to any last value that was normal or biased in the computing algorithm. The detectable magnitude of the sensor fault is numerically computed according to the noise level and the configuration of the skew-configured IMU. Once the parity matrix Hy .k/

1202

S. Yoon et al.

in Eq. (47.13) and the threshold Ti in Eq. (47.21) are determined, the fault magnitude, which can be always detected, is computed as ı  kf .k/k D Ti hj min

(47.39)

where hj is the j -th column vector of Hy .k/ D Œh1 h2 : : : hm . When all six inertial sensors are normal, the constantly detectable magnitude is 5.9 deg/s. When five inertial sensors are normal, the constantly detectable magnitude is 6.6 deg/s. When four inertial sensors are normal, the constantly detectable magnitude is 10.9 deg/s. If the corresponding column vector hj is larger than the denominator of Eq. (47.39) in a certain case, the sensor fault smaller than the constantly detectable magnitude can be detected usually but not always. Note that the detectable fault magnitude increases from 5.9 to 10.9 as the number of normal sensors decreases. Thirty-six experiments were performed during the flight test according to the failure scenario, including seven sets of successive triple faults. Table 47.2 summarizes the fault types and magnitudes of each sensor failure scenario. Although the parameters of the hybrid FDI system had been shown to be tolerant to 10 deg/s bias errors in the ground test, the fault diagnosis performance in the flight test was superior than the expected performance from the ground test. Therefore, three additional failure scenarios (5 through 7) were tested. Table 47.3 summarizes the flight experimental results for each sensor failure scenario. In the flight test, the aircraft maneuvered to perform level flights, bank turns, climbs, and descents. The result of the fault isolation is simply displayed as follows: success, partial success, and fail. In failure scenarios 1 through 5, the proposed FDI technique successfully detected and isolated the successive sensor faults under various real flight situations. In failure scenarios 6 and 7, all of the first/second faults were successively detected and isolated by the proposed FDI system, but the third fault was partially isolated in some cases even though all of the third faults were detected by the parity space method without any problems. Those cases that missed the fault isolation will be discussed in detail in the next subsection. Next, one of the flight experimental results is discussed in detail. The UAV was in a coordinate turn for 23 s. The mean velocity and the mean altitude over the ground were 35 m/s and 170 m, respectively. The successive stuck and bias faults

Table 47.2 Sensor failure scenarios commanded from the ground station First fault Second fault Third fault Scenario number deg/s Type Sensor deg/s Type Sensor deg/s Type 1 15 Bias 1 15 Bias 3 15 Bias 2 15 Bias 2 15 Bias 4 30 Stuck 3 10 Bias 1 10 Bias 3 10 Bias 4 10 Bias 2 10 Bias 4 30 Stuck 5 8 Bias 1 10 Stuck 3 8 Bias 6 8 Bias 2 10 Stuck 4 8 Bias 7 10 Stuck 5 8 Bias 1 10 Stuck

Sensor 5 6 5 6 5 6 3

47 Fault Diagnosis of Skew-Configured Inertial Sensor System Table 47.3 Flight experimental results of the fault detection and isolation Scenario number Flight status FDI parameters 1 Level flight k1 D 1:8; k2 D 1:8; k3 D 1:5; tc D 0:12 Bank turn Climb/descent 2 Level flight Bank turn Climb/descent 3 Level flight k1 D 1:5; k2 D 1:5; k3 D 1:2; tc D 0:12 Bank turn Climb/descent 4 Bank turn 5 Level flight k1 D 1:3; k2 D 1:3; k3 D 1:1; tc D 0:12 Bank turn Climb/descent 6 Level flight Bank turn Climb/descent 7 Roll oscillation

1203

Fault isolation (success/total) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (2/2) Success (3/3) Success (2/2) Partial success (2/4) Success (2/2) Partial success (2/4) Partial success (1/3)

were generated in Sensors 5, 1, and 3 according to the failure scenario 7. Figure 47.9 shows the sensor signals, the parity norm, the discrete wavelet transform, and the isolated sensor number, respectively. The first fault took place in Sensor 5 at 1,274.5 s. Its signal was stuck at 10 deg/s as shown in the first plot. The fault detection function FD .k/ increased immediately when the fault occurred, as shown in the second plot. The first fault was detected when this residual exceeded the prescribed threshold T1 . Then, the fault isolation function FI;5 .k/ declared that Sensor 5 was the faulty sensor, as shown in the fourth plot. Consequently, Sensor 5 was isolated from the skew-configured inertial measurement system, and the fault diagnosis parameters were updated for a five-sensor configuration. The second fault, an 8 deg/s bias error in Sensor 1 at 1,283.5 s, was detected in the same way. The wavelet transform of the remaining four inertial sensor signals began after the isolation of the second faulty inertial sensor. When the third fault occurred in Sensor 3, which stuck at 10 deg/s after 1,287.8 s, the residual exceeded the threshold T3 , as shown in the second plot of Fig. 47.9. The parity space method could no longer identify the faulty sensor because of an insufficient number of remaining healthy sensors. However, the wavelet transform of each sensor signal could identify the timing of the abrupt signal change caused by the fault. As shown in the third plot, the wavelet-transformed signal of Sensor 3 exceeded the predefined threshold T4 and that sensor was thus successfully isolated. Other flight experiments in Scenarios 1 through 5 showed similarly satisfactory performance; the UAV was able to continue performing its mission with the skew-configured IMU composed of the remaining three sensors.

Sensor Signals, deg/s

1204

S. Yoon et al.

40

#1 #2 #3 #4 #5 #6

20 0 −20 1275

1280

1285

1290

1295

Parity Norm, deg/s

10 residual thr1 thr2 thr3

8 6 4 2 0

Isolated Faulty Sensor No.

Wavelet Transform, deg/s

1275

1280

1285

1290

8

1295

#1 #2 #3 #4 #5 #6 thr4

6 4 2 0 1275

1280

1275

1280

1285

1290

1295

1285

1290

1295

6 4 2 0

Time, sec

Fig. 47.9 Flight test results of the hybrid fault detection and isolation system

47.4.3 Analysis 47.4.3.1 Comparison of the State Vector with or Without FDI The performance of the proposed FDI technique against three successive faults can be analyzed by comparing the body-coordinate sensor signals collected during the faulty situation to those collected during a healthy, fault-free situation. The healthy body-coordinate angular rates are calculated from normal sensor signals, and the faulty body-coordinate angular rates are calculated from all six sensor signals. Figures 47.10 and 47.11 illustrate the difference between the healthy and the faulty angular rates of failure scenario 7 without and with application of the FDI scheme, respectively. As shown in Fig. 47.10, without the application of the

Roll Rate, deg / sec

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

1205

healthy without FDI

20 0 −20 −40

1275

1280

1285

1290

1295

Pitch Rate, deg / sec

20

10

0

Yaw Rate, deg/ sec

−10

Isolated faulty sensor no.

healthy without FDI

1275

1280

1285

1290

1295

healthy without FDI

10 0 −10 −20 1275

1280

1275

1280

1285

1290

1295

1285

1290

1295

6 4 2 0

Time, sec

Fig. 47.10 Comparison between healthy and faulty signals without FDI

hybrid FDI, the body-coordinate sensor output (blue line) could not provide accurate signals (red line) when successive faults occurred according to failure scenario 7 (Sensor 5 stuck at 10 deg/s, Sensor 1 bias of 8 deg/s, and Sensor 3 stuck at 10 deg/s, successively). Abrupt changes in the roll, pitch, and yaw rate signals due to the stuck faults in Sensors 5 and 3 can be seen at 1,274.5 and 1,288.1 s, respectively. Notably, the error between the healthy and faulty signals increased suddenly when the first stuck fault occurred. The error became even larger and more remarkable after the second and third faults.

Roll Rate, deg/sec

1206

S. Yoon et al.

healthy with FDI

20 0 −20 −40

1275

1280

1285

1290

1295

Pitch Rate, deg/sec

20 10

0

Yaw Rate, deg/sec

−10

Isolated Faulty Sensor No.

healthy with FDI

1275

1280

1285

1290

10

1295

healthy with FDI

0 −10 −20 1275

1280

1275

1280

1285

1290

1295

1285

1290

1295

6 4 2 0

Time, sec

Fig. 47.11 Comparison between healthy and faulty signals with FDI

Figure 47.11 shows that the proposed FDI scheme enhances the accuracy of the body-coordinate sensor output. The computed body-coordinate sensor outputs (blue line) are consistent with the healthy signals (red line) except the short interval when the faults occurred because the FDI algorithm requires almost a half second to detect and isolate the faulty sensor. Table 47.4 presents a quantitative error analysis based on the error norm between the healthy and faulty signals. The errors obtained without FDI decreased by up to 29 % when the proposed FDI is applied. Therefore,

47 Fault Diagnosis of Skew-Configured Inertial Sensor System Table 47.4 Error norm analysis between healthy and faulty signals

Axis Roll rate Pitch rate Yaw rate

w/o FDI 139.02 121.57 98.42

1207

w/ FDI 82.68 32.58 35.95

Ratio 1:0.59 1:0.29 1:0.37

the flight experiments demonstrated that the proposed hybrid FDI technique with the skew-configured IMU can continuously provide healthy sensor outputs despite successive stuck and bias faults in the individual sensors.

47.4.3.2 False Alarm Due to the Smooth Transition As summarized in Table 47.3, several partially successive tests of failure scenarios 6 and 7 failed to isolate the third fault in the skew-configured IMU. Figure 47.12 shows one of the partially successive FDI results of failure scenario 7. In this scenario, three successive faults occurred: Sensor 5 stuck at 10 deg/s, Sensor 1 experienced an 8 deg/s bias error, and Sensor 3 stuck at 10 deg/s. As depicted in the fourth plot, the first and second faulty sensors were correctly isolated but the third faulty sensor was not isolated because the wavelet-transformed signals did not exceed the threshold. As can be seen in the first plot of Fig. 47.12, the third fault in Sensor 3, which stuck at 10 deg/s after 1,395.5 s, did not generate a sufficiently remarkable signal change compared to the general sensor noise to be identified by the wavelet transform based on the Daubechies 2 wavelet function. Because the wavelet transform is a signal processing technique based on each sensor output, the fault may not be identified if the magnitude of the fault signal is smaller than the magnitude of the noise signal or if noise characteristics vary over time. Since the wavelet transform is primarily designed to avoid emergencies when a third fault occurs, the UAV should return to base by homing guidance if the parity space method detects a third fault but the in-lane monitoring system is unable to identify the failed sensor. If the UAV cannot return to base, it may be useful to generate a virtual sensor signal based on model-based propagation after the third fault. Nevertheless, this approach would yield errors if the discrepancy caused by the modeling uncertainties exceeds the fault signal. Alternatively, the wavelet transform could be run in parallel based on the higher scale’s wavelet basis function (e.g., Daubechies 9 or 10) to identify a slow or weak signal change. The capability of the flight control computer should be considered since this method requires a higher computation load. 47.4.3.3 False Alarm Due to the Computational Delay The computational delay in the wavelet transformation is another cause of the partially successful results. Confirmation of the presence of a faulty signal involves six steps of the fault detection algorithm and requires 0.12 s. A faulty signal that lasts less than 0.12 s is considered temporary noise and may be ignored. Moreover,

Sensor Signals, deg/s

1208

S. Yoon et al.

20

#1 #2 #3 #4 #5 #6

10 0 −10 −20 1380

1385

1390

1395

1400

Parity Norm, deg/s

10 residual thr1 thr2 thr3

8 6 4 2 0

Isolated Faulty Sensor No.

Wavelet Transform, deg/s

1380

1385

1390

1395

1400

8

#1 #2 #3 #4 #5 #6 thr4

6 4 2 0 1380

1385

1390

1395

1400

1380

1385

1390 Time, sec

1395

1400

6 4 2 0

Fig. 47.12 Partially successful result: insensitivity of the wavelet transform to the stuck fault

the discrete wavelet transform monitors signal changes based on 32 previous signal samples. If the change in the faulty signal is small, the parity norm increases slowly. Therefore, the 32 samples of the wavelet transform might not contain the faulty signal because the fault signal is detected by the parity norm. Figure 47.13 shows one of the partially successful FDI results from failure scenario 6. In this scenario, Sensor 2 experienced an 8 deg/s bias error, Sensor 4 stuck at 10 deg/s, and Sensor 6 experienced an 8 deg/s bias error. As shown in the third plot, the wavelet transformation of the signal from Sensor 6 exceeded the threshold due to the third bias fault at 945.9 s. The third fault could not therefore be detected because the fault detection function in the second plot increased slowly and did not guarantee the existence of the faulty sensor. This problem can be easily overcome by considering the previous history of the wavelet-transformed signal.

Sensor Signals, deg / s

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

20

Parity Norm, deg /s

#1 #2 #3 #4 #5 #6

10 0 −10 925 10

930

935

940

945

950 residual thr1 thr2 thr3

8 6 4 2 0 925

Isolated Faulty Sensor No. Wavelet Transform, deg / s

1209

930

935

940

945

950

8

#1 #2 #3 #4 #5 #6 thr4

6 4 2 0 925

930

935

930

935

940

945

950

940

945

950

6 4 2 0 925

Time, sec

Fig. 47.13 Partially successful result: computational delay in the fault detection function

47.4.3.4 False Alarm Due to the Aircraft Motion The proposed fault diagnosis method is theoretically independent of aircraft motion or trajectory. Four maneuvers were performed to verify that the aircraft motion does not affect fault diagnosis performance: level flight, bank turn, climb/descent, and oscillating roll motion. As summarized in Table 47.3, the fault diagnosis result was successful regardless of aircraft maneuver when FDI parameter ki was large. However, false alarms occurred as the FDI parameter decreased because the fault diagnosis became more sensitive to sensor noise or signal change. Even though aircraft states and actuator inputs do not affect the parity vector, a sudden maneuver does affect the time delay between the present aircraft state and the measured states in some inertial sensors. One of the fast maneuvers, an oscillating roll, was performed in flight scenario 7. If the response times of the inertial sensors are not

1210

S. Yoon et al.

synchronized, different responses to a fast maneuver might break the orthonormal properties of parity relation and cause a false alarm. Therefore, the redundant secondary IMU should be chosen to have similar response characteristics to those of the primary IMU.

47.5

Conclusion

A model-free hybrid FDI scheme was proposed to provide system tolerance to the occurrence of multiple successive faults in the inertial sensors during flight. The hybrid FDI scheme combines the parity space method with the in-lane monitoring method based on the discrete wavelet transform. This scheme was successfully verified through the hardware-in-the-loop simulation and flight experiments. The skew-configured inertial sensor system was developed with the primary IMU and the redundant secondary IMU. In addition, three successive fault signals were generated virtually with the fault signal generation board during the real flight experiments. The first and the second faults in arbitrary inertial sensors were detected and isolated by the parity space method, and the third sensor fault was detected by the parity space method and isolated by the in-lane monitoring method based on the discrete wavelet transform. The proposed fault diagnosis scheme is applicable to the robust navigation and control of robots, unmanned air/ground vehicles, and satellites. Acknowledgments The content of this chapter is based on the following journal paper: Yoon et al. (2011). This work was supported by the Korea Aerospace Research Institute under contract 0498-20060013 and was also partially supported by the Smart UAV Development Program, one of the twenty-first Century R&D Programs funded by the Ministry of Knowledge Economy of Korea.

References A. Bogges, F.J. Narcowich, A First Course in Wavelets with Fourier Analysis (Prentice Hall, Upper Saddle River, 2001) J. Chen, R.J. Patton, Robust Model-Based Fault Diagnosis for Dynamic Systems (Kluwer Academic, Boston, 1999) E.Y. Chow, A.S. Willsky, Analytical redundancy and the design of robust failure detection systems. IEEE Trans. Autom. Control 29(7), 603–614 (1984) I. Daubechies, Ten Lectures on Wavelets (SIAM, Philadelphia, 1992) F.E. DeAngelis, A threshold redundancy management algorithm for skewed sensor arrays, in Proceedings of IEEE/AIAA/NASA 9th Digital Avionics Systems Conference, Virginia Beach (1990) X. Ding, L. Guo, T. Jeinsch, A characterization of parity space and its application to robust fault detection. IEEE Trans. Autom. Control 44(2), 337–343 (1999) D. Dohono, I. Johnstone, Ideal spatial adaptation via wavelet shrinkage. Biometrika 81(3), 425– 455 (1994)

47 Fault Diagnosis of Skew-Configured Inertial Sensor System

1211

J. Gertler, Fault detection and isolation using parity relations. Control Eng. Pract. 5(5), 653–661 (1997) J. Gertler, Fault Detection and Diagnosis in Engineering Systems (Marcel Dekker, New York, 1998) J.P. Gilmore, R.A. McKern, A redundant strapdown inertial systems mechanization-SIRU, in Proceedings of the AIAA Guidance, Control, and Flight Mechanics Conference, Santa Barbara, CA (1970) I. Hwang, S. Kim, Y. Kim, C.E. Seah, A survey of fault detection, isolation, and reconfiguration methods. IEEE Trans. Control Syst. Technol. 18(3), 636–653 (2010) R. Isermann, Fault-Diagnosis Systems: An Introduction from Fault Detection to Fault Tolerance (Springer, New York, 2006) R. Isermann, P. Balle, Trends in the application of model-based fault detection and diagnosis of technical processes. Control Eng. Pract. 5(5), 709–719 (1997) J. Jankovitz, J. Perdzock, Laboratory and flight evaluation of the integrated inertial sensor assembly (IISA), in Proceedings of IEEE Position, Location and Navigation Symposium, Orlando, FL (1988) J.L. Junkins, Y. Kim, Introduction to Dynamic and Control of Flexible Structures (American Institute of Aeronautics and Astronautics, Washington, DC, 1993) S. Kim, Y. Kim, C.G. Park, I. Jung, Hybrid fault detection and isolation techniques for aircraft inertial measurement sensors, in Proceedings of the AIAA Guidance, Navigation, and Control Conference, Providence, RI (2004) S. Kim, Y. Kim, C.G. Park, Failure diagnosis of skew-configured aircraft inertial sensors using wavelet decomposition. IET Control Theory Appl. 1(5), 1390–1397 (2007) U.K. Krogmann, Failure management in spatio-temporal redundant, integrated navigation and flight control reference-systems, in Proceedings of IEEE Position, Location and Navigation Symposium, Las Vegas, NV (1990) X. Li, Discrete wavelet transform for tool breakage monitoring. Int. J. Mach. Tool Manuf. 39(12), 1935–1944 (1999) S. Mallat, Theory of multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11(7), 674–693 (1989) S. Mallat, A Wavelet Tour of Signal Processing (Academic, San Diego, 1999) C.R. McClary, J.R. Walborn, Fault-tolerant air data inertial reference system development results, in Proceedings of IEEE Position, Location and Navigation Symposium, Las Vegas, NV (1994) S. Park, Hybrid fault detection and isolation method for TDOF inertial sensors using hardware and analytic redundancy, M.S. Thesis, Department of Mechanical and Aerospace Engineering, Seoul National University, Republic of Korea, 2004 W. Park, E. Kim, K. Seong, Y. Kim, A study on the parameter estimation of DURUMI-II for the fixed right elevator using flight test data. J. Mech. Sci. Technol. 20(8), 1224–1231 (2006) R.J. Patton, Robustness in model-based fault diagnosis: the 1995 situation. Ann. Rev. Control 21, 103–123 (1997) R.J. Patton, J. Chen, Review of parity space approaches to fault diagnosis for aerospace systems. J. Guid. Control Dyn. 17(2), 278–285 (1994) R.J. Patton, P.M. Frank, R.N. Clark, Fault Diagnosis in Dynamic Systems: Theory and Application (Prentice Hall, Englewood Cliffs, 1989) R.J. Patton, P.M. Frank, R.N. Clark, Issues of Fault Diagnosis for Dynamic Systems (Springer, London, 2000) J.E. Potter, J.C. Deckert, Minimax failure detection and identification in redundant gyro and accelerometer systems. J. Spacecr. 10(4), 236–243 (1972) J.E. Potter, M.C. Suman, Thresholdless redundancy management with arrays of skewed instruments, in Integrity in Electronic Flight Control Systems. Agardograph-224 (NATO, Neuilly sur Seine, 1977), pp. 11–25

1212

S. Yoon et al.

A. Ray, M. Desai, A redundancy management procedure for fault detection and isolation. J. Dyn. Syst. Meas. Control 106(2), 149–156 (1984) S. Santoso, E.J. Powers, W.M. Grady, P. Hofmann, Power quality assessment via wavelet transform analysis. IEEE Trans. Power Deliv. 11(2), 924–930 (1996) J.J. Sudano, J.R. Preisig, J. Pokotylo, Improved fault detection using a selected grouping of parity equations for advanced flight control systems, in Proceedings of the IEEE 1988 National Aerospace and Electronics Conference, Dayton, OH (1988) D.H. Titterton, J.L. Weston, Strapdown Inertial Navigation Technology (Peter Peregrinus, Lavenham, 1997) S. Yoon, S. Kim, J. Bae, Y. Kim, E. Kim, Experimental evaluation of fault diagnosis in a skewconfigured UAV sensor system. Control Eng. Pract. 19(2), 158–173 (2011) Q. Zhang, Y.A. Yan, Wavelet-based approach to abrupt fault detection and diagnosis of sensors. IEEE Trans. Instrum. Meas. 50(5), 1389–1393 (2001)

Section X UAV Modeling, Simulation, Estimation and Identification Lora Weiss

48

UAV Modeling, Simulation, Estimation, and Identification: Introduction Kimon P. Valavanis and George J. Vachtsevanos

Modeling and simulation techniques are essential and crucial in the design, operation, and performance assessment of UAVs. Together with estimation and identification, they are integral components of the overall UAV design process. They afford the opportunity to develop, test, and evaluate sensing, control, and cooperative control algorithms, among others. Software-in-the-loop and hardwarein-the-loop simulations are typical prerequisites to flight testing. UAV Modeling, Simulation, Estimation and Identification presents a treatment of these topics from first principles to mechanisms that lead to UAV qualification and certification. Flight Dynamics Modeling of Coaxial Rotorcraft UAVs by Wang, Cui, Chen, and Lee aims at presenting a thorough approach related to systematic modeling of UAV flight dynamics by talking about the UAV model formulation, parameter identification, and model verification. The presented methodology of UAV model formulation and parameter identification is based on modeling two kinds of coaxial helicopters. This is a challenging problem as modeling of coaxial helicopters, despite the common governing kinematic and dynamic principles, deserves special attention due to their distinctive mechanical structure and aerodynamics behavior. The modeling procedures and parameter identification technique presented in this chapter may also serve as a guideline for modeling other types of aerial vehicles. Modeling of a Micro UAV with Slung Payload by Feng, Rabbath, and Su derives a mathematical and simulation model of a micro UAV carrying a payload.

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 141, © Springer Science+Business Media Dordrecht 2015

1215

1216

K.P. Valavanis and G.J. Vachtsevanos

When the load is slung underneath a UAV by cable, the flight dynamics of the UAV are altered, which makes the stability of the UAV disturbed. The unstable oscillation degrades UAV performance, and the accurate placement of the load will be affected. Unlike external disturbances, the negative effects are related to the characteristics of the UAV and the payload. In order for the UAV to be able to adopt the change of the system dynamics and reduce the effects caused by the swing of the load, one modeling method of a micro UAV with single slung payload is addressed. The slung payload is treated as a pendulum-like mass point, and the coupling factors between the UAV and the payload are considered. The derived model may be used to estimate the negative effects acting on the UAV, and it may also be used to estimate the trajectory of the load, which allows for improving the accuracy of placement of the loads. Command and Control of Autonomous Unmanned Vehicles by Scheidt describes a series of experiments, including 2011 experiments in which 16 fully autonomous unmanned vehicles, including 9 unmanned air vehicles, were used to simultaneously support mounted, dismounted, and maritime users. During these experiments users provided abstract mission-level ISR needs to the “vehicle cloud.” These needs were interpreted by the vehicles, which self-organized and efficiently achieved the user’s objectives. The starting point is using information theory to examine UAV command and control (C2). The information theoretic analysis provides a justification and use cases for autonomous UAVs. An autonomous unmanned vehicle system “Organic Persistent Intelligence Surveillance and Reconnaissance” (OPISR) is introduced. OPISR is an autonomous unmanned vehicle system that combines the immediate response to tactical ISR needs provided by organic assets with the time-on-station, minimal logistics provided by persistent unmanned systems. OPISR autonomous vehicles collectively interpret real-time tactical intelligence surveillance and reconnaissance (ISR) objectives submitted by any number of disadvantaged users, gather the required ISR data, and return the needed intelligence directly to the affected user. OPISR is an ad hoc, decentralized system that requires no central base or authority and is capable of functioning in communications-denied environment.

Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

49

Fei Wang, Jinqiang Cui, Ben M. Chen, and Tong H. Lee

Contents 49.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218 49.2 Working Principle of Coaxial Helicopters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1220 49.3 Model Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1222 49.3.1 Model Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1222 49.3.2 Kinematics and Rigid-Body Dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1224 49.3.3 Force and Torque Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1227 49.3.4 Force and Torque Formulation of Fixed-Pitch Coaxial Helicopter . . . . . . . . . . . . . 1228 49.3.5 Force and Torque Formulation of Variable-Pitch Coaxial Helicopter. . . . . . . . . . . 1234 49.4 Model-Based Parameter Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1239 49.4.1 Direct Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1239 49.4.2 Test Bench Experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1239 49.4.3 Flight Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1246 49.5 Model Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1252 49.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1255 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1255

Abstract

In many unmanned aerial vehicle (UAVs)-related engineering projects, flight dynamics modeling of the controlled platform usually forms the foundation of the whole project development. An accurate mathematical model of the controlled UAV not only makes high-performance model-based control law design possible but also provides insights into the mechanical design of the aerial platform so that radical improvements can be made at the beginning of the development. However, the topic of flight dynamics modeling is somehow not paid enough

F. Wang () • J. Cui • B.M. Chen • T.H. Lee Control and Simulation Lab, Department of Electrical and Computer Engineering, National University of Singapore, Singapore e-mail: [email protected]; [email protected]; [email protected]; [email protected]

K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 111, © Springer Science+Business Media Dordrecht 2015

1217

1218

F. Wang et al.

attention to across the general engineering audience. This chapter aims to disseminate the knowledge of systematic modeling of UAV flight dynamics by talking about UAV model formulation, parameter identification, and model verification. The methodology of UAV model formulation and parameter identification based on the case study of two kinds of coaxial helicopters will be explained. Modeling of coaxial helicopters, despite the common governing kinematic and dynamic principles, deserves special attention due to their distinctive mechanical structure and aerodynamics behavior. The modeling procedures and parameter identification process presented in this chapter also serve as a guideline for modeling other types of aerial vehicles.

49.1

Introduction

In recent years, unmanned aerial vehicles (UAVs) have been more actively involved in military and civil operations. Possible UAV applications include reconnaissance and intelligence gathering, forest patrol, coastline monitoring, and search and rescue (Tsach et al. 2010). In all kinds of UAV development, the flight dynamics modeling of the controlled platform always forms the cornerstone of the whole project. UAV modeling not only provides an accurate mathematical model so that advanced model-based control law design techniques can be used but also provides insights into the mechanical design of the aerial platform itself. Moreover, in contrast to those conventional manned aerial vehicles, UAV platforms are normally custommade or largely modified from off-the-shelf products to cope with the payload requirement and the mounting geometry of the onboard avionics. UAV platforms may possess different working principles, ranging from fixed-wing, rotorcraft, flapping-wing to mono-copters like the Samurai (Rosen and Seter 1991). To model some of the special types of aerial platforms, the well-established work for the conventional aerial platforms in literature cannot be applied directly. Hence, it is especially useful to let UAV developers understand how the models are derived and how to identify unknown parameters embedded within the models. Many works related to flight dynamics modeling and model-based parameter identification of aerial vehicles are conducted in literature. Some of them are done for the ultimate purpose of UAV development. In Heffley and Mnich (1988), a mathematical model for the conventional single-rotor helicopter with adequate complexity was derived, and the report highlighted the formulation of the main rotor thrust generation. In Johnson (1994), a fairly comprehensive and detailed coverage of helicopter working theory and design considerations were provided in aspects including helicopter vertical flight, forward flight, mathematics of rotating systems, rotary wing dynamics and aerodynamics, aeroelasticity, and stability and control. Based on the above works, Cai et al. (2012) obtained a comprehensive nonlinear model of a miniature single-rotor helicopter. This work has also later been extended in a book (Cai et al. 2011) with other UAV-related topics like UAV construction, software development, and controller design.

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1219

In more recent years, the development of UAV platforms has entered the era of miniature, microscale, and even nanoscale. The used-to-be efficient conventional fixed-wing or single-rotor helicopter structure may not be an optimum design anymore. The reduced form factor has also resulted in more innovative and unconventional aerodynamic designs, and for these new types of platforms, quite a few new modeling works have been published. In Nonami et al. (2010), the mathematical model of a miniature coaxial helicopter is derived in the transfer function form. The identified linear model is used for optimal controller design. In Bermes (2010), the design and dynamic modeling and the simulation of an autonomous coaxial micro helicopter platform (muFly) are investigated. A modular dynamic model is developed, which incorporates the active and passive flapping characteristics of a hingeless rotor system, the stabilizer bar dynamics, roll-pitch steering by swashplate or displacing center of mass, etc. Coaxial helicopter is an attractive UAV platform due to its small dimension, high thrust-to-weight ratio, and aerodynamic symmetry. If a coaxial helicopter and a single-rotor helicopter are of the same weight, the size of the coaxial helicopter can be 35–40 % smaller. In addition, the aerodynamic symmetry of coaxial helicopters successfully gets rid of the yaw moment and side forces commonly seen in singlerotor helicopters. Thus, the coaxial helicopter is much more effective in fast forward flights. These advantages make the coaxial helicopter ideal for UAV operation in confined environments such as indoor and cluttered outdoor. There are two types of coaxial helicopters. One has rotor blades with a fixed collective pitch, while the other is with variable collective pitch. In the following content of this chapter, they will be called the fixed-pitch coaxial and the variable-pitch coaxial in short. The dynamic modeling of these two kinds of coaxial helicopters will form the main discussion in this chapter. A major difference of modeling between the coaxial helicopter and the conventional single-rotor helicopter is the pair of concentric rotors, in which each of them rotates with the induced velocity affected by the other. Such relationship is described in Colin (1997). Detailed studies of the wake dynamics of the two rotors were also documented in Kim and Brown (2006), Rand and Khromov (2010), and Lim et al. (2009). Another specialty of coaxial helicopter is the stabilizer bar attached to the top rotor hub, which passively stabilizes the helicopter. It, however, causes strong influences to the rotor dynamics especially to the fixed-pitch coaxial configuration as the upper rotor is not linked to any servo. As a result, the cyclic pitch control of the upper rotor is solely induced by the stabilizer bar. The stabilizer bar dynamics is commonly modeled as a first-order lag system (Mukherjee and Waslander 2011). In the works shown in Mukherjee and Waslander (2011) and Schafroth et al. (2010), the tip-path-plane (TPP) dynamics is separated into the upper and lower portion, where only the lower TPP is controlled by the servo inputs. In a few recent works on the modeling of miniature coaxial helicopter, although fairly complete nonlinear or linear models are obtained, the works lack intuitive explanation of the model formulation. Moreover, their methods of parameter identification are not comprehensive enough. For example, in Neamtu et al. (2010),

1220

F. Wang et al.

the helicopter dynamics were treated as a black box, while the whole system is vaguely identified using the CIFER (Comprehensive Identification from FrEquency Responses) toolkit. To complement the existing work, this book chapter presents the detailed derivation of the nonlinear model for both the fixed-pitch and the variablepitch coaxial helicopters. The content of this chapter is organized as follows: Sect. 49.2 briefly gives the working principles of both the fixed-pitch and the variable-pitch coaxial helicopters. Next, Sect. 49.3 will comprehensively formulate the coaxial flight dynamics system, starting from kinematics, rigid-body dynamics, and then force/torque composition and generation. Some parts of the model formulation can be shared by both types of coaxial helicopters, while the differences in rotor thrust generation and rotor flapping dynamics will be separately explained. Model-based parameter identification will be introduced in Sect. 49.4 with the fixed-pitch coaxial as an example. Last but not least, Sect. 49.5 will verify the derived model of the fixed-pitch coaxial case and proves the fidelity of the overall modeling methodology.

49.2

Working Principle of Coaxial Helicopters

Before any rigorous derivation of the flight dynamic models of the aforementioned two types of coaxial helicopters, the general working principles of the two will first be introduced so that it will be easier for readers to understand the later detailed model formulation. First of all, both platforms have a pair of contrarotating rotors (upper and lower) to provide the fundamental lift force for the overall platform. This is indeed where the term “coaxial” comes from. For the fixed-pitch coaxial helicopter, as shown in Fig. 49.1, the collective pitch of the rotor blades cannot be changed. Hence, the heave and yaw motion of the platform can only be controlled by varying the rotational speed of the rotors, which are linked to two separate motors with step-down gears. In general, the summation of the motor speeds determines the helicopter vertical motion, while the difference of the two determines the yaw motion. Rolling and pitching are accomplished by introducing a cyclic pitch to the lower rotor via a dual-servo-controlled swashplate. In this way, a tilted flapping of the rotor blades can be induced, and the generated rotor thrust becomes non-vertical. In the case of the platform shown in Fig. 49.1, which is called the ESky Big Lama, the cyclic pitch of the upper rotor is not actively controlled. Instead, the rotor hub tethers together with a stabilizer bar. As the stabilizer bar is purposely constructed with relatively high moment of inertia, when the helicopter fuselage suddenly tilts, the stabilizer bar tends to remain rotating at the original level plane. This introduces a cyclic pitch to the upper rotor, and it is purposely designed in a way that the thrust tilting resulted from this cyclic pitch counters the instantaneous motion of the helicopter. Therefore, this kind of fixed-pitch coaxial helicopters is inherently stable with regard to their attitude angles.

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1221

Stabilizer Bar Upper Rotor Hub Lower Rotor Hub

Servo 2

Swashplate Servo 1

Motor 2 Motor 1 Headlock Gyro

Fig. 49.1 Description of the fixed-pitch coaxial helicopter

On the other hand, the variable-pitch coaxial helicopter, as shown in Fig. 49.2, is a mini coaxial helicopter customized according to the design of full-scale coaxial helicopter from the Kamov design bureau. Its rotor head is equipped with integrated hinges and shock-resistant dampers. This helicopter also consists of two contrarotating rotors. However, unlike the fixed-pitch coaxial, the rotational speeds of the variable-pitch coaxial helicopter’s rotors are maintained at the same constant speed in normal flight conditions. The dynamic motion of the helicopter is achieved by actively changing the pitch angles of both upper and lower rotors via the upper swashplate and the lower swashplate, respectively. The pitch angles of both rotors are constituted by collective pitch and cyclic pitch, which are mixed controlled by three servos linked to the lower swashplate. The two swashplates are always parallel to each other since they are circumferentially connected by three rigid linkages. The upper rotor is attached with a Bell-Hiller stabilizer bar which introduces damping to the rotor’s cyclic pitch. Collective and cyclic inputs from servos are transferred to the lower swashplate and also induced to the top swashplate, resulting in the dynamic movement of the helicopter in heave direction or pitch-roll direction. The yaw direction control is realized by changing the collective pitch angle of the lower rotor. For this particular platform, the upper rotor and lower rotor are driven by the same brushless direct current (DC) electric motor with the same gear ratio. Hence, the rotational speeds of the upper rotor and the lower rotor are always the same.

1222

F. Wang et al.

Upper Rotor

Stabilizer bar Top Swashplate Lower Rotor

Yaw Control Lower Swashplate Motor

ESC

Headlock Gyroscope

Fig. 49.2 Description of the variable-pitch coaxial helicopter

49.3

Model Formulation

49.3.1 Model Overview The mathematical model of any continuous physical dynamic system can be expressed in the following compact form: xP D f.x; u; w/;

(49.1)

where x is the state vector, u is the input vector, and w represents external disturbances. For the case of the fixed-pitch coaxial helicopter system, the system state and input vectors can be defined as x D .x y z u v w  

p q r aup bup adw bdw up dw rfb /T ;

(49.2)

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1223

Table 49.1 Physical meaning of state variables

Physical meaning

Unit

p, position in the ground frame

m

vb , linear velocity in the body frame

m/s

Symbol x y z u v w   p q r aup bup adw bdw up dw rfb

9 Roll = Pitch attitude angle ; Yaw !, angular velocity in body frame  Longitudinal flapping angle of upper blades Lateral  Longitudinal flapping angle of lower blades Lateral Rotational speed of the upper rotor Rotational speed of the lower rotor Controller state of yaw stability augmentation

rad

rad/s

rad rad

NA

u D .ıail ıele ıthr ırud /T ;

(49.3)

w D .!u !v !w / :

(49.4)

T

For the case of the variable-pitch coaxial helicopter system, the system state and input vectors can be defined as x D .x y z u v w  

p q r aup bup adw bdw rfb /T ;

(49.5)

u D .ıail ıele ıthr ırud /T ;

(49.6)

w D .!u !v !w / :

(49.7)

T

The physical meanings of the state, input, and disturbance variables are listed in Tables 49.1–49.3. It should be noted that the fixed-pitch coaxial helicopter has two additional state variables, up and dw , because its rotor rotational speeds keep changing during normal flights and its motor dynamics is not fast enough to be neglected. On the other hand, the rotors of the variable-pitch coaxial platform always rotate at the same speed. Thus, no dynamics need to be considered. With regard to the input definition, the conventional radio-controlled (RC) joystick signals, aileron (ıail ), elevator (ıele ), throttle (ıthr ), and rudder (ırud ), normalized to Œ1; 1 with respect to their corresponding minimum and maximum values, are chosen as the primary system inputs. This kind of input definition makes sure that the modeling methodology is also applicable to other types of

1224 Table 49.2 Physical meaning of input variables

F. Wang et al.

Variables ıail ıele ıthr ırud ıNrud

Table 49.3 Physical meaning of wind disturbance variables

Physical meaning Control deflection for lateral cyclic pitch Control deflection for longitudinal cyclic pitch Control deflection for collective pitch Control deflection for collective pitch of tail rotor Control deflection for yaw-stabilityaugmentation controller

Range Œ1; 1 Œ1; 1 Œ1; 1 Œ1; 1 Œ1; 1

Variables

Physical meaning

!u !v !w

Wind velocity in the helicopter x-axis Wind velocity in the helicopter y-axis Wind velocity in the helicopter z-axis

aerial platforms. Usually, the helicopter attitude angles (, ), heading ( ), and 3D position (x,y,z) are chosen to be the ultimate controlled outputs as they can comprehensively define the state of the helicopter in the full 6 degree-of-freedom (6DoF) space. Furthermore, to derive a mathematical model of a complex system, it is preferable to modularize the overall system into subsystems so that the divide-andconquer strategy can be used. Here, an overview of the two model structures is shown in Figs. 49.3 and 49.4. From the inputs to the state variables, there are numerous blocks representing all the subsystems involved. The two models share quite a few similarities but preserve their own distinctive features as well. On the one hand, the input and output definitions are unified for both, and the same kinematics and dynamics blocks can be used to relate the body-frame forces and moments to the helicopter 6DoF motion. On the other hand, the two models are different in the mechanism in generating the individual forces and moments. This results in two different sets of intermediate model blocks. In the following sections, mechanisms in all these blocks will be explained in detail. The similar blocks between these two types of platforms will be discussed together, while the different blocks will be discussed separately.

49.3.2 Kinematics and Rigid-Body Dynamics As a common practice of aeronautic analysis, model formulation of a flying vehicle normally assumes that the target platform is a rigid body. Thus, it follows the universal 6DoF kinematics equations and the Newton-Euler dynamics equations. Two main coordinate frames are generally involved to link the equations. One is the north-east-down (NED) frame, and the other is the helicopter body frame. While the NED frame is stationary with respect to a static observer on the ground, the body

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1225 aup, bup

Upper Rotor Flapping Dynamics

u, v, w

Forces and Moments Upper Rotor

x, y, z

F δthr

Mixer

Fuselage

+ M

δrud

Headlock Gyroscope

6-DOF Rigid-body Dynamics

Lower Rotor

Kinematics

φ, θ, ψ

mg p, q, r

δail δele

Lower Rotor Flapping Dynamics

adw, bdw

Fig. 49.3 Model structure of fixed-pitch coaxial helicopter

Fig. 49.4 Model structure of variable-pitch coaxial helicopter

frame is placed at the center of gravity (CG) of the coaxial helicopter, where its origin and orientation move together with the helicopter fuselage (see Fig. 49.5). It is worth noting that the listed formulas to describe the kinematics and dynamics of the coaxial helicopter are indeed universal to all other rigid-body vehicles. The relationship between the helicopter NED-frame position and its body-frame velocity is determined by the following navigation equation: 0 1 2 30 1 c c c s s  s c c s c C s s xP u @yP A D 4s c s s s C c c s s c  c s 5 @ v A ; s c s c c Pz w

(49.8)

1226

F. Wang et al.

Fig. 49.5 Coordinate frames and various forces and torques

Qup

Tup

Qdw

Tdw

x

NED Frame N

Body Frame

0 E

y

z

D

mg

where x, y, z are the NED-frame position components of the helicopter and u, v, w are the body-frame velocity components. , , are the conventional roll, pitch, yaw angles of the helicopter fuselage and s , c denote sin./, cos./, respectively. P P , are not P , It is also critical to point out that the Euler angle derivatives, , orthogonal to each other. They are related to the body-frame angular rates, p, q, r, by the following equation: 30 1 0 1 2 P p 1 s s =c c s =c @ P A D 40 c s 5 @ q A : P 0 s =c c =c r

(49.9)

Note that the above equation has singularity at  D 90ı . If full-envelope flight is required, a quaternion representation is recommended. However, normal maneuvering of a coaxial helicopter will not hit such an extreme condition. Thus, it is still adequate to use this relatively simple equation. Next, by treating the whole coaxial platform as a rigid mass, the 6DoF dynamics can be described by the following Newton-Euler equations: 0 1 0 1 0 1 0 1 u p uP Fx 1 @ vP A D @Fy A  @ q A  @ v A ; m w r w P Fz 8 0 19 0 1 0 1 0 1 p = p pP < Mx @ qP A D J1 @My A  @ q A  J @ q A ; ; : r r Mz rP

(49.10)

(49.11)

where Fx , Fy , Fz are projections of the net force, F, onto the helicopter bodyframe x-, y-, z-axis and Mx , My , Mz are projections of the net torque, M, onto the body-frame x-, y-, z-axis. The compositions of F and M come from

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1227

various parts of the coaxial helicopter. For the fixed-pitch and variable-pitch coaxial helicopters, although their force and torque compositions are more or less the same (explained in Sect. 49.3.3), the individual force and torque generation will have different formulations because of their different working principles (explained in Sects. 49.3.4 and 49.3.5, respectively). Till now, the unknown parameters that need to be identified are m, the total mass of the platform, and J, the moment of inertia of the platform, which is in the form of 2

3 Jxx Jxy Jxz J D 4Jxy Jyy Jyz 5: Jxz Jyz Jzz Since the coaxial helicopters are normally designed to be symmetric in both longitudinal and lateral directions, Jxy , Jxz , Jyz are extremely small and can be assumed to be zero. The identification method of Jxx , Jyy , Jzz will be explained in Sect. 49.4, which talks about various approaches in the problem of model-based parameter identification.

49.3.3 Force and Torque Composition As mentioned in Sect. 49.3.2, forces and torques acting on the coaxial helicopter come from various mechanical parts. First of all, the helicopter weight exerts a force of mg in the NED-frame z-axis. After converting it to the body frame, the vector is shown as the second term on the right-hand side of Eq. 49.12. Next, when the rotor blades spin, they generate thrusts, Ti (i = up, dw), in the direction perpendicular to their respective tip path plane (TPP). When the upper and lower TPPs deviate from their default orientation, the thrust vectors no longer pass through the CG of the helicopter, thus creating rotational torque. The torque vectors caused by the rotor thrusts can be calculated by lup  Tup and ldw  Tdw , where lup and ldw are the displacement vectors from helicopter CG to the upper rotor hub, and the lower rotor hub respectively. The deviation of the TPP can be described by the longitudinal flapping angle ai and the lateral flapping angle bi . The thrust decomposition to the body-frame axes can be approximated by Eq. 49.14. Nonzero ai and bi also directly result in flapping torque on the rotor hub. This torque can be simplified as the second term on the right-hand side of Eq. 49.13, where Kˇ is called the effective spring constant, and it has the same value for both the upper and lower rotors provided they are rotating at approximately the same speed. Furthermore, the rotation of the rotors creates the drag torque, Qup and Qdw , around the body-frame z-axis. When the coaxial helicopter hovers without yaw motion, the two torques have the same magnitude, thus canceling each other. Else, if the net drag torque is nonzero, yaw acceleration is generated. In addition, the change of rotational speeds of the rotors also generates the so-called reaction torques on the helicopter body (denoted by Qr,up and Qr,dw ). They are described in (Eq. 49.16),

1228

F. Wang et al.

where Jup and Jdw are the moment of inertia of the upper rotor (with stabilizer bar) and the lower rotor with respect to the rotor shaft. They can be calculated by measuring the mass and dimension of the rotor blades and stabilizer bar and assuming a regular geometric shape. Lastly, when the helicopter moves in air, its fuselage experiences drag forces, Xfus , Yfus , Zfus , due to air resistance. This drag force is usually related to the linear speed and up-front surface area of the aerial vehicle. Equations (49.12) and (49.13) have summarized all the forces and torques mentioned above, with (49.14)–(49.16) explaining how to evaluate the individual terms:

0

1 1 0 1 0 Fx Xfus s X @Fy A D Ti C mg @s c A C @ Yfus A ; Fz c c Zfus

(49.12)

0

1 0 1 Mx bi X X X X @M y A D @ li  Ti C Qd;i C Kˇ ai A C Qr;i ; Mz 0 1 0 li D jli j @ 0 A ; 1 0

(49.13)

0

1  sin ai A; Ti D jTi j @ sin bi  cos ai cos bi

(49.14)

Qd;up

0 1 0 1 0 0 D jQd;up j @0A ; Qd;dw D jQd;dw j @ 0 A ; 1 1

(49.15)

Qr;up

0 1 0 1 0 0 P up @0A ; Qr;dw D Jdw  P dw @ 0 A : D Jup  1 1

(49.16)

49.3.4 Force and Torque Formulation of Fixed-Pitch Coaxial Helicopter The generation of individual forces and torques on the fixed-pitch coaxial and the variable-pitch coaxial has quite different formulations. In this section, a full formulation of the fixed-pitch coaxial force and torque generation will be provided. That means, by tracing all the formulas listed in this section, the forces and torques exerted on the fixed-pitch coaxial helicopter can be exhaustively related to the four fundamental inputs in a rigorous way.

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1229

49.3.4.1 Thrust and Torque from Rotors Here, the magnitude of the rotor thrust and drag torque, jTi j and jQd;i j, will first be investigated. According to the aerodynamic actuator disk theory (Bramwell et al. 2001), the magnitude of thrust generated by the rotors can be formulated as follows: jTi j D CT;i A.i R/2 ;

(49.17)

where  is the density of air, CT;i is the lift coefficient, A is the rotor disk area, i is the rotational speed of the rotor, and R is the rotor blade length. Since this is a fixed-pitch coaxial helicopter, CT;i , like the other parameters in (49.17), is constant. The only variable is i . Hence, the equation can be simplified to jTi j D kT;i 2i ;

(49.18)

where kT;i is a lumped thrust coefficient that needs to be identified. Similar assumptions and formulation can be applied to the relationship between the drag torque and the rotational speed of the rotors: jQd;i j D kQ;i 2i :

(49.19)

49.3.4.2 Rotor Flapping Dynamics For this specific type of coaxial helicopter, the rotor collective pitch is fixed, while the cyclic pitch can be changed. For the lower rotor, the rotor hub is connected to the aileron and the elevator servos via a swashplate. When the swashplate tilts, it teeters the rotor hub and creates a cyclic pitch on the rotor. For every cycle of rotation, the rotor blade will reach the maximum angle of attack at a particular phase angle at which the lift on the blade is largest. This results in the flapping of the rotor disk. The whole mechanism is a combination of gyroscopic precession and aerodynamic precession. For the case of ESky Big Lama, if one observes the rotor blade in a slow motion, the maximum rotor flapping occurs roughly at 45ı lag with respect to the occurrence of maximum angle of attack. This explains why the aileron and elevator servos of the off-the-shelf coaxial platform are connected to the swashplate 45ı off the body-frame x-, y-axis. In this way, the aileron servo mainly controls the lateral flapping of the lower rotor, and the elevator servo mainly controls the longitudinal flapping. However, the flapping phase lag is not exactly equal to 45ı (slightly larger than 45ı from test bench observations) due to mechanical modifications to the original RC platform (original rotor blades have been replaced by stiffer ones for larger payload). This results in non-negligible coupling between the servo inputs and the lower rotor longitudinal and lateral flapping angles. As the lower rotor does not have any additional damping mechanism attached, its flapping process is almost instantaneous. By assuming a first-order dynamics, the time constant can be observed via a high-speed camera. The result turns out to be 0.0375 s (see Fig. 49.6), which is very small as compared to dynamics of other parts of the coaxial helicopter

1230 Fig. 49.6 Step response of servo motion attached to the lower rotor (top: t D 0; center, t  0:0375 s; bottom, t D 1). This figure looks dim because it was taken by a high-speed camera

F. Wang et al.

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1231

and thus can be ignored. Hence, the relationship between servo inputs and lower rotor flapping angles can be formulated in a static way: adw D Aa;dw ıele C Ab;dw ıail  Aq q;

(49.20)

bdw D Bb;dw ıail C Ba;dw ıele  Bp p;

(49.21)

where aileron (ıail ) and elevator (ıele ) are the servo inputs, Aa;dw and Bb;dw are the on-axis steady-state ratio from servo inputs to flapping angles, and Ab;dw and Ba;dw are the off-axis (coupling) values. The last term which depends on angular rates, p and q, comes from an effect called rotor damping, which was considered in literature either in linear or quadratic form. Here, a linear form is chosen because of its simplicity. For the upper rotor system, a stabilizer bar is attached to the rotor hub, so that they teeter together. As the stabilizer bar has large moment of inertia, it tends to remain at its original rotating plane. Hence, at the moment when the helicopter body tilts, the stabilizer bar TPP will remain at the level plane, thus creating a cyclic pitch on the upper rotor which leads to blade flapping. The torque generated by this flapping redresses the rotational motion of the helicopter and significantly stabilizes the whole platform attitude. Similar to the lower rotor system, the stabilizer bar is installed at 45ı phase lead to the rotor blade. In this way, the maximum flapping happens at the direction that precisely counters the rotational motion of the helicopter. Again, there is coupling between the longitudinal and lateral channels because the flapping phase lag is not exactly 45ı . The following equations describe the above mentioned dynamics: 1 .  sb /; Psb D sb

(49.22)

1 Psb D .  sb /; sb

(49.23)

aup D Aa;up .sb  / C Ab;up .sb  /  Aq q;

(49.24)

bup D Bb;up .sb  / C Ba;up .sb  /  Bp p;

(49.25)

where sb and sb are the roll and pitch angles of the stabilizer bar TPP, Aa;up and Bb;up are the on-axis steady-state ratio from the stabilizer bar teetering angles to the upper rotor flapping angles, and Ab;up and Ba;up are the off-axis (coupling) values. sb is the time constant of approximated first-order flapping dynamics. Again, the same rotor damping effects (terms depending on p and q) are considered for the upper rotor flapping dynamics.

49.3.4.3 Fuselage Drag When the helicopter fuselage moves in air, it experiences drag force acting on the opposite direction of the motion. For the body-frame horizontal directions, the rotor

1232

F. Wang et al.

downwash is deflected by u and v. In the situation when u (or v) is less than vi (the induced velocity of air at the lower rotor), the downwash effect needs to be taken into account. Otherwise, the downwash effect is relatively weak and can be ignored. The fuselage in all three directions is considered as a flat plate perpendicular to the helicopter motion; thus, the drag coefficient is approximately unity. As such, the horizontal fuselage drag forces are formulated in a quadratic form:  Xfus D  Sx u  max.vi ; juj/; 2  Yfus D  Sy v  max.vi ; jvj/; 2 s jTdw j ; vi D 2R2

(49.26) (49.27) (49.28)

where Sx and Sy are the effective drag area along the body-frame x- and y-axis, respectively. For the vertical direction, since the fuselage is constantly exposed to the lower rotor downwash, it is commonly formulated in the following form:  Zfus D  Sz .w  vi /jw  vi j: 2

(49.29)

However, as the lift coefficient test for identifying kT;i in Eq. 49.18 was done with the presence of the fuselage (so the term 2 Sz v2i has already been taken into account), the above equation needs to be compensated as  Zfus D  Sz w  max.vi ; jwj/; 2

(49.30)

where Sz is the effective drag area along the body-frame z-axis.

49.3.4.4 Motor Speed Dynamics Two brushless DC motors are used on the ESky Big Lama coaxial platform. Their rotational speed dynamics follows the well-known differential equation of electro motors: Jmot !P D

km U  km ke !  d!  ML ; Rmot

(49.31)

where Jmot is the motor moment of inertia, km and ke are the mechanical and electrical motor constants, U is the input voltage, Rmot is the resistance of the circuit, d is the friction coefficient, and ML is the external torque acting on the motor shaft. Here, ML is equal to the rotor drag torque Qd;i appeared in Eq. 49.19. If the helicopter operates at a near-hover condition, everything can be approximated as linear. ML can be assumed to be a combination of a constant trimming value, ML ,

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1233

and another term proportional to extra rotational speed as compared to the trimming speed,  : ML D ML C kL .   /:

(49.32)

By further considering that the rotational speed of rotor, , and the rotational speed of the motor, !, are perfectly proportional by the gear ratio, the rotor speed dynamics can be simplified to the following first-order equations: P up D 1 .mup ıup C up  up /;  mt

(49.33)

P dw D 1 .mdw ıdw C dw  dw /;  mt

(49.34)

where up and dw are the trimming values of the rotor rotational speed at hovering, mt is the time constant of the motor speed dynamics, and mup , mdw are the steadystate ratio between the change of rotor speeds and the change of motor inputs.

49.3.4.5 Mixer and Headlock Gyro Dynamics In order to largely decouple the throttle-heave and the rudder-yaw dynamics, the throttle and rudder signals are passed into a hardware mixer and transformed to dual motor control signals: ıup D ıthr C ıNrud ;

(49.35)

ıdw D ıthr  ıNrud :

(49.36)

It can be seen that when the throttle signal ıthr increases, inputs to both motors increase; when the rudder signal ıNrud increases, the input to the motor connected to the upper rotor increases while the input to the motor connected to the lower rotor decreases. Note that the rudder signal in the above mixer equation is not the original signal ırud . From ırud to ıNrud , there is a hardware headlock gyro which helps refine the rudder signal and acts as the most inner-loop yaw motion stabilizer. Usually, there is a PI controller embedded inside the headlock gyro, and it can be formulated as follows: rPfb D Ka ırud  r; ıNrud D KP .Ka ırud  r/ C KI rfb ;

(49.37) (49.38)

where rfb is the augmented state variable needed by the integral control. At this point, the full dynamics of a coaxial helicopter have been mathematically formulated, but the model parameters are yet to be identified. In Sect. 49.4, the identification methods will be comprehensively given.

1234

F. Wang et al.

49.3.5 Force and Torque Formulation of Variable-Pitch Coaxial Helicopter For the case of the variable-pitch coaxial helicopter, the formulation of force and torque generation will not be thoroughly listed as some of the formulas are very much like its fixed-pitch counterpart. For example, the formulas about fuselage drag and the headlock gyro PI controller are more or less the same. As such, only formulations that have significant differences when compared to the fixed-pitch coaxial case will be listed and explained here. It should also be noted that the main differences compared occur in two areas, namely, the rotor thrust generation and the rotor flapping dynamics. In this section, the formulations about thrust generation and rotor flapping dynamics of the variablepitch coaxial helicopter will be derived from a more fundamental perspective, thus capturing more aerodynamic details. This also reflects the fact that model formulation of aerial vehicles can be done at different levels of complexity. If the model is to be used for accurate structural analysis and optimization or controller design for aggressive maneuvering, then a more precise and complicated modeling approach should be adopted. However, if the planned flight missions are relatively stable and peaceful, a simplified model formulation is usually more than enough, such as that of the fixed-pitch coaxial case. Thorough aerodynamics study of coaxial rotor system is itself a big research topic. NASA researcher (Colin 1997) has provided a good survey covering the aerodynamics research worldwide, including America, Russia, Japan, and Germany. Of all the surveyed techniques, blade element momentum theory (BEMT) (Leishman 2006) is a standard method for preliminary rotor analysis before complex high-level analysis, such as free vortex models (FVM) and Lagrangian particle vortex methods (LPVM). In this chapter, the detailed derivation and reasoning of BEMT will not be repeated. Only the main computation procedures are listed for easier reference and understanding.

49.3.5.1 Thrust and Torque from Rotors The formal analysis of helicopter rotor motion usually starts from its thrust, torque, and power generation. The rotor thrust, torque, and power can be expressed as T D CT A2 R2 ; 2

3

(49.39)

Q D CQ A R ;

(49.40)

P D CP A3 R3 :

(49.41)

Here,  is the air density, A is the rotor disk area,  is the rotor rotational speed, and R is the rotor radius. CT , CP , CQ are the rotor thrust coefficient, power coefficient, and torque coefficient, respectively. It should be noted that the power is related to torque by P D Q. Hence, CP D CQ numerically. All the parameters in Eqs. 49.39–49.41 are constant except for the three coefficients CT , CP , CQ .

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1235

First of all, the incremental thrust coefficient is defined as 1 Cl˛ Cl r 2 dr D .u r 2  r/dr; 2 2

dCT D

(49.42)

where Cl˛ is the lift-curve slope of the airfoil section, which can be obtained by checking the official wind tunnel test results if the blade airfoil is standard. is the rotor solidity defined as the ratio of the blade area against the rotor disk area. u is the blade pitch distribution on the upper rotor. r is the nondimensional radial distance along the blade. is the nondimensional induced velocity which can be expressed in terms of r and 1 as s 

.r; 1 / D

1 Cl˛  16F 2

2

Cl˛ u r  C 8F



1 Cl˛  16F 2

 ;

(49.43)

where F is the factor to account for the Prandtl tip loss defined as   2 F D cos1 .e f /; 

(49.44)

where f is given in terms of the number of blades Nb and the radial position of the blade element r:   Nb 1  r ; (49.45) f D 2 r and  is the induced inflow angle, which equals to .r/=r. For both the upper and lower rotors, the same principles can be applied. However, as the inner part of the lower rotor operates in the vena contracta of the upper rotor, the analysis can be more complicated. From the flow visualization results of Taylor (1950), the author stated that the wake of the upper rotor contracts fully within 0:25R below the upper rotor. The ideal wake contraction ratio, rc , is 0.707, but in practice it is found closer to 0.8. The contracted wake is defined as Ac D  rc2 R2 . The inner area of the lower rotor encounters incoming stream-tubes with velocity V1 C .A=Ac /vu (V1 C 2vu in the ideal case). For beam sections lying inside the upper rotor contraction area, the inflow distribution is given as s



.r; 1 / D  

1 C .A=Ac / u Cl˛  16F 2

1 C .A=Ac / u Cl˛  16F 2

2  ;

C

Cl˛ l r 8F (49.46)

where l is the blade pitch distribution on the lower rotor. For points outside the contraction area, the inflow distribution is given as

1236

F. Wang et al.

s

.r; 1 / D



1 Cl˛  16F 2

2

C

Cl˛ l r  8F



1 Cl˛  16F 2

 :

(49.47)

When the inflow velocity distribution is obtained, the total thrust lift coefficient could be found by integrating Eq. 49.42 as Z CT D

rD1 rD0

dCT :

(49.48)

For the rotor torque coefficient and power coefficient, their incremental calculation is also provided by BEMT: dCQ D dCP D

.Cl C Cd /r 3 dr; 2

(49.49)

where Cd is the rotor drag coefficient. By knowing the fact that D r, the incremental power is defined as Cl r 3 dr C Cd r 3 dr 2 2 2 D Cl r dr C Cd r 3 dr 2 2 D dCPu C dCPo :

dCP D

(49.50)

The induced power coefficient can be obtained by integrating dCPu as follows: Z CPu D

rD1 rD0

Z dCPu D

rD1 rD0

dCTu ;

(49.51)

and the profile part of the rotor power is given by CPo D

2

Z

1 0

Cd r 3 dr:

(49.52)

49.3.5.2 Bare Rotor Flapping Dynamics Similar to that of the fixed-pitch coaxial case, the rotor flapping dynamics of the variable-pitch coaxial helicopter is also seen as a rigid disk which can tilt about its longitudinal and lateral axes. However, instead of directly linking the flapping angles’ dynamics to the servo inputs, the flapping angles (ai , bi ) are first related to the cyclic pitch angles (cyc;bi , cyc;ai ) as intermediate variables. The presence of cyc;bi and cyc;ai is a joint consequence of servo inputs and the stabilizer bar flapping angles. The detailed description of the rotor equations is extremely complicated. Here, a simplified formulation is adopted, where the rotor forces and moments are expressed as a polynomial function of the rotor state variables (Mettler 2002).

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1237

By removing the higher-order terms of the TPP equation, the remaining first-order rotor dynamics could be expressed as  bPi D bi  p C Ba ai C cyc; bi ;

(49.53)

 aP i D ai   q C Ab bi C cyc; ai ;

(49.54)

where 8Kˇ ;  2 Jˇ   8e 1 16 1 ; D r  3R

Ab D Ba D

D

cCl˛ R4 : Jˇ

(49.55) (49.56) (49.57)

ai and bi are the first-order TPP flapping angles in the longitudinal and lateral directions for either the upper rotors or the lower rotors.  and  are the flapping time constant and the lock number of the rotor blades, respectively. Jˇ is the blade moment of inertia. cyc; ai and cyc; bi are the longitudinal and lateral cyclic pitch of rotor blade. The other terms are defined by the same symbols as that of the fixedpitch coaxial helicopter analysis in Sect. 49.3.4. The approximate formulation in Eqs. 49.53 and 49.54 characterizes the crucial TPP responses with respect to rotor cyclic control inputs.

49.3.5.3 Stabilizer Bar Flapping Dynamics The stabilizer bar, which is attached to the upper main rotor shaft via a free-teetering hinge, can be regarded as a third disk. It consists of two paddles and a steel rod. The stabilizer bar is not designed to produce thrust or moment on the main hub, whereas its main function is to adjust the helicopter dynamics via the Bell-Hiller mixer by augmenting the cyclic pitch command of the upper rotor. It serves as a feedback system which increases the helicopter robustness against wind gust and turbulence (Cai et al. 2011). The flapping dynamics of stabilizer bar can be expressed as two first-order differential equations: 1 C Psb D q  sb C ıail ; sb sb

(49.58)

1 D Psb D p  sb C ıele ; sb sb

(49.59)

where sb is the stabilizer bar flapping time constant, and it can be calculated as sb D

16 ; sb 

(49.60)

1238

F. Wang et al.

where sb is the stabilizer bar lock number:  4  4  rsb csb Cl˛;sb Rsb : sb D Iˇ;sb

(49.61)

The free-teetering hinge does not constrain the flapping of the stabilizer bar; thus, there is no coupling between the longitudinal and lateral flapping motions. The augmented rotor cyclic pitch of upper rotor can be expressed as cyc; aup D Aa ıele C Ksb sb ;

(49.62)

cyc; bup D Bb ıail C Ksb sb ;

(49.63)

where Ksb is the ratio of rotor blade cyclic pitch to stabilizer bar flapping.

49.3.5.4 Lumped Flapping Dynamics In this variable-pitch coaxial configuration, the upper rotor and the lower rotor receive the same cyclic input (ıail , ıele ) since the top swashplate and bottom swashplate are always parallel. To minimize the overall complexity of the model, the two counterrotating rotor disks are treated as one equivalent rotor disk with respect to flapping motions. This assumption is only valid when the helicopter does not perform rapid maneuvering. By making this assumption, the model can be simplified to a large extent yet maintaining reasonable fidelity. The imaginary rotor has equivalent longitudinal and lateral angles expressed as as and bs . Combining Eqs. 49.53–49.63, the lumped flapping dynamics subsystem could be represented in the following state-space form: xP D A x C B u;

(49.64)

yP D C x;

(49.65)

where

2

0 6 0 6 6 AD6 0 6 4 1

0 1 p     BqC C ; u D ıail ; y D p ; xDB @as A ıele q bs 3 0 0 Lb 3 2 0 0 0 Ma 0 7   7 6 0 0 7 1000 1 Ab 7 7 6 ; ; C D ; B D 7 1  4 0 A0 5 0100 s s 7 lon 0 Ba 15 Blat 0 0  s s Lb D

mg Hmr C Kˇ mg Hmr C Kˇ ; Ma D : Jxx Jyy

(49.66)

(49.67)

(49.68)

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

49.4

1239

Model-Based Parameter Identification

When model formulation of a certain type of aerial platform has been derived, the next step is to identify unknown parameters of the derived model for a particular case. This is commonly known as the model-based parameter identification. In this section, several common parameter identification methods will be introduced based on the case study of ESky Big Lama, belonging to the fixed-pitch coaxial helicopter. It will be seen that some of the model parameters can be directly measured, while the remaining ones need special test bench experiments or flight tests to be carried out. The identification procedures for the case of variable-pitch coaxial helicopter will not be repeated since similar methods can be applied. It is also hoped that the readers will try the suggested experimental setups on other types of aerial platforms if they face similar parameter identification problems in their UAV-related projects.

49.4.1 Direct Measurement For the ESky Big Lama fixed-pitch coaxial helicopter, some of its model parameters, especially those inherently defined by the platform dimension, geometry, weight loading, and the operational environment, can be directly measured or algebraically calculated, for example, the total mass of the platform (m), the distance from the rotor hubs to CG (jlup j and jldw j), the air density (), the rotor diameter (R), the effective drag area of the fuselage (Sx , Sy , Sz ), and the moment of inertia of the upper and lower rotor system (Jup , Jdw ). Table 49.4 shows all the parameter values that can be identified through direct measurement and simple calculation.

49.4.2 Test Bench Experiment In most aerial vehicle modeling cases, the direct measurement method is only able to determine a small portion of the parameters. The remaining majority parameters have to be identified by additional test bench experiments or actual flight tests. In this subsection, some common test bench methods are introduced, and they are again illustrated based on the ESky Big Lama fixed-pitch coaxial helicopter case. First of all, the diagonal elements of the helicopter moment of inertia matrix Jxx , Jyy , Jzz can be measured by the so-called trifilar pendulum method proposed in Harris (1996). The experimental setup is shown in Fig. 49.7. In this experiment, the coaxial platform is suspended by three flexible strings with equal length l. The horizontal distances between the attached points and the CG are l1 , l2 , and l3 , respectively. One can slightly twist and release the platform around the z-axis and record the oscillation period tl . The moment of inertia is then given by Jzz D

mgl1 l2 l3 tl2 l1 sin ˛1 C l2 sin ˛2 C l3 sin ˛3 ;  4 2 l l2 l3 sin ˛1 C l1 l3 sin ˛2 C l1 l2 sin ˛3

(49.69)

1240

F. Wang et al.

Table 49.4 Parameters determined via direct measurement Parameter Physical meaning  D 1:204 kg m3 m D 0:977 kg R D 0:250 m g D 9:781 m s2 Sf x D 0:00835 m2 Sf y D 0:01310 m2 Sf z D 0:01700 m2 jlup j D 0:195 m jldw j D 0:120 m Jup D 6:8613  104 kg m2 Jdw D 3:2906  104 kg m2

Air density Total mass of the platform Rotor radius Earth gravity Effective longitudinal fuselage drag area Effective lateral fuselage drag area Effective vertical fuselage drag area Distance from platform CG to the upper rotor hub Distance from platform CG to the lower rotor hub Moment of inertia of the upper rotor with stabilizer bar Moment of inertia of the lower rotor

a3 l1

l2

a1 a2

l3

l

Fig. 49.7 The trifilar pendulum method

where ˛1 , ˛2 , and ˛3 are the angles denoted in Fig. 49.7. Similar experiments can be done to obtain the moment of inertia around the y and z axes. Figure 49.8 shows the experimental setups to carry out this trifilar pendulum method to obtain the moment of inertia of the ESky Big Lama in all three axes. Next, to identify the rotor thrust coefficient and torque coefficient (kT;i and kQ;i ), two self-designed test bench experiments were carried out

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1241

Fig. 49.8 The setups to test helicopter moment of inertia. (a) x-axis. (b) y-axis. (c) z-axis

(see Figs. 49.9 and 49.10). The main measurement sensors include a force meter and a tachometer. When different values of pulse width modulation (PWM) signals are given to the motors, the steady-state rotor rotational speed and the generated thrust/toque are recorded. For the thrust experiment, results are summarized in Fig. 49.11. There are four lines in the plot, in which two of them (solid lines) perfectly match. They represent the cases when only one rotor, either the upper rotor or the lower rotor, is rotating. The dashed line on the top is a numerical combination of the two solid lines, while the dash-dot line comes from the actual test with both rotors spinning at the same speed. The gap between the two lines shows a drop in thrust efficiency caused by aerodynamic interactions between the two rotors. According to Deng et al. (2003), for a coaxial helicopter operating in near-hover condition, the induced-velocity effect of the upper rotor to the lower rotor is significantly larger than that of the lower rotor to the upper rotor. Thus, the loss of thrust efficiency can be assumed to be fully absorbed by the lower rotor thrust coefficient. Hence, kT;up is the gradient of the solid line, and kT;dw is the gradient difference between the dash-dot line and the solid line. For the torque experiment, results are summarized in Fig. 49.12. The solid line represents the case when only

1242 Fig. 49.9 Setup to investigate relation between rotor thrust and rotor rotational speed

Fig. 49.10 Setup to investigate relation between rotor torque and rotor rotational speed

F. Wang et al.

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1243

18 Upper Rotor Measurement Lower Rotor Measurement Dual Rotor Measurement 1 Dual Rotor Measurement 2

16

Generated Thrust (N)

14 12 10 8 6 4 2 0

0

1

2

3

4

Square of Rotor Speed

Ω2

5 (rad/s)2

6

7 x 104

Fig. 49.11 Data plot of thrust against square of rotor speed

the stabilizer bar is rotating, while the dash-dot line is for a single rotating rotor. The dashed line is generated with the upper rotor and the stabilizer bar spinning together. Unsurprisingly, it matches the numerical combination of the lower two lines. Thus, the gradient of the dashed line is kQ;up , and the gradient of the dash-dot line is kQ;dw . The identification of parameters involved in the motor dynamics can be done via test bench experiments also. The method to determine motor time constant (mt ) is a bit tricky, as the transient response of the rotor speed with a step motor input is very difficult to be recorded in real time. As such, instead of examining the transient response of the rotor speed with motor step input, the transient response of the input voltage subject to the changes of the motor Back-EMF (voltage generated by the spinning motor) is recorded using an oscilloscope (see Fig. 49.13). In theory, the time constant of the two transient responses should be the same. On the other hand, mup and mdw can be identified by plotting the steady-state relationship between the rotor speed and the normalized motor input (see Fig. 49.14). mup and mdw are the gradients of the two fitted lines in the figure. The headlock gyro forms the most inner-loop control in the helicopter yaw channel. As mentioned in Sect. 49.3.4, it is a PI controller with three parameters (Ka , KP , KI ) to be identified. The identification of Ka can be done via a hovering

1244

F. Wang et al. 0.3 Stabilizer bar Rotor Rotor & Stabilizer Bar

Generated Torque (N m)

0.25

0.2

0.15

0.1

0.05

0

0

2

4

6

8

Square of Rotor Speed Ω2 (rad/s)2

Fig. 49.12 Data plot of torque against square of rotor speed

Fig. 49.13 Estimation of time constant of motor speed dynamics

10

12 x 104

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

260

1245

Upper Rotor Lower Rotor

240

Rotor Speed Ω (rad/s)

220 200 180 160 140 120 100 −0.4

−0.2

0 0.2 Normalized Input to Motor

0.4

0.6

0.8

2.60 0.40

3.50 0.55

Fig. 49.14 Data plot of rotor speed against motor input Table 49.5 Yaw rate against rudder input

r (rad/s) ırud (1, 1)

1.50 0.25

2.50 0.35

turn test of the coaxial helicopter with different steady-state yaw rates. This test belongs more to the test bench category instead of flight test because it can be done on a swivel table with minimal friction. Table 49.5 shows four sets of data recorded. By plotting the least-square-fit line (see Fig. 49.15) and calculating its gradient, Ka can be determined. For the identification of KP and KI , the helicopter is placed stationary on a table. KP and KI can be determined by observing the headlock gyro output signal (in pulse width modulation form) caused by a small known step inputs. The initial ratio between the output and the input is KP  Ka , while the climbing rate of the step response is KI  Ka . Last but not least, test bench experiments are also capable of determining some of the model parameters involved in the rotor flapping dynamics. First of all, by tilting the helicopter suddenly with rotor rotating at hovering speed and observing the transient step response of the stabilizer bar TPP (see Fig. 49.16) by a highspeed camera, the time constant (sb ) can be found to be about 0:2 s. In addition, the on-axis parameters that appear in the rotor flapping equations (Aa;up , Bb;up , Aa;dw , and Bb;dw ) can be roughly identified by measuring lengths and angles under

1246

F. Wang et al. −0.5 −1

Collected data Best fit line

−1.5 −2 r (rad/s)

−2.5 −3 −3.5 −4 −4.5 −5 −5.5 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

δrud (−1,1)

Fig. 49.15 Data plot of yaw rate against rudder input

extreme conditions (see Figs. 49.17 and 49.18) and assume a linear proportional relationship between each pair of them. It should be noted that the parameters related to rotor flapping dynamics are most critical to the whole coaxial helicopter model. The above rough measurements may not be good enough to finalized their values. However, they can be used as an initial guess and get fine-tuned later by a modelbased numerical search method via in-flight test data. In the next subsection, this method will be explained in detail. Here, Table 49.6 only lists those parameters that are already finalized at this test bench experiment stage.

49.4.3 Flight Test After the majority of parameters have been identified, the remaining ones and a few uncertain ones can be identified and refined by analyzing flight test data with input perturbations (frequency sweeping). The recommended software for this task is called “Comprehensive Identification from FrEquency Responses” (CIFER). It is a MATLAB-based software package developed by NASA Ames Research Center for military-based rotorcraft system identifications. In this ESky Big Lama case, since the remaining unidentified parameters are all about rotor flapping dynamics, only aileron and elevator perturbations need to be done to collect the relevant data for CIFER analysis. However, CIFER and most other parameter identification tools can only handle linear models. Hence, linearization needs to be done first to relate the aileron and elevator inputs to

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs Fig. 49.16 Step response of stabilizer bar TPP motion (top: t D 0; center, t  0:2 s; bottom, t D 1)

1247

1248

F. Wang et al.

Fig. 49.17 Left: maximum teetering angle of the lower rotor hub; right: maximum flapping angle of the lower rotor

Fig. 49.18 Left: maximum teetering angle of the stabilizer bar; right: maximum teetering angle of the upper rotor hub

the helicopter roll, pitch angular rates. By relating and linearizing all the model formulation related to roll-pitch rate dynamics and upper rotor flapping dynamics at the hovering condition, one can obtain the following fourth-order linear state-space approximation: 0

1

2 Xdw Bp;dw

pP 6 Bq C 6 0 B CD6 @aup A 6 4 Ab;up bup Bb;up Jxx

0 Xdw Aq;dw Jyy

Aa;up Ba;up

0

Xup Jxx

3

0

1

2X

dw Bb;dw

6 Jxx 7 p B q C 6 Xdw Ab;dw 0 7 7 B C C 6 Jyy 7@ A 6 0 5 aup 4 0 bup 0 0  1sb

Xup Jyy  1sb

3

Xdw Ba;dw Jxx 7  Xdw Aa;dw 7 7 ıail ; Jyy 7 0 5 ıele

0 (49.70)

where Xup D Tup lup CKˇ and Xdw D Tdw ldw CKˇ . By treating ıail , ıele as the inputs and p, q as the outputs (all can be recorded during flight tests) and giving known

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1249

Table 49.6 Parameters determined via test bench experiments Parameter Physical meaning Jxx D 0:0059 kg m2 Jyy D 0:0187 kg m2 Jzz D 0:0030 kg m2 kT;up D 1:23  104 N s2 rad2 kT;dw D 8:50  105 N s2 rad2 kQ;up D 4:23  106 N ms2 rad2 kQ;dw D 3:68  106 N ms2 rad2 mup D 106:90 rad s1 mdw D 106:45 rad s1 1  up D 203:38 rad s 1  dw D 217:88 rad s mt D 0:12 s sb D 0:2 s Ka D 6:4267 KP D 0:667=Ka KI D 0:713=Ka

Platform moment of inertia in the x-axis Platform moment of inertia in the y-axis Platform moment of inertia in the z-axis Effective thrust coefficient of the upper rotor Effective thrust coefficient of the lower rotor Effective torque coefficient of the upper rotor Effective torque coefficient of the lower rotor Steady-state ratio between change of upper rotor speed and change of motor input Steady-state ratio between change of lower rotor speed and change of motor input Trimming rotational speed of the upper rotor Trimming rotational speed of the lower rotor Motor time constant Stabilizer bar time constant Feed-forward gain of the headlock gyro system Proportional feedback gain of the headlock gyro system Integral feedback gain of the headlock gyro system

constraints and reasonable initial values (on-axis values from Sect. 49.4.2 and offaxis values as zeros), CIFER helps to search for optimal numerical solution based on frequency response matching. A stable result with good matching is obtained as follows: 30 1 2 3 1 2 p 17:19 0 0 934:1 102:48 38:08 pP   B C 6 7 Bq C 6 0 5:360 291:3 0 7 7 B q C C 6 11:73 31:95 7 ıail : B CD6 @aup A 4 0:2745 0:49 5 0 5 @aup A 4 0 0 5 ıele 0:49 0:2745 0 5 0 0 bup bup (49.71) 0

With this numerical result, Figs. 49.19–49.22 show the corresponding comparison of frequency response between the data collected via actual flight tests and the CIFER derived model fit. For both the on-axis and off-axis responses, the matching is very good, indicating a high-quality identification result. Next, by comparing Eqs. 49.70 and 49.71, all the parameters involved in angular rate and rotor flapping dynamics can be finalized. Table 49.7 shows the identification results that have been obtained via flight test. Till now, all unknown parameters in the fixed-pitch coaxial model have been identified.

1250

F. Wang et al. δail −− p

Magnitude (DB)

20

Actual data Derived model 0

−20

101

Phase (deg)

0 −200 −400 −600

101

Coherence

1

0.5

0

101 Frequency (Rad/Sec)

Fig. 49.19 Response comparison using frequency-sweep input (ılat  p) Magnitude (DB)

20

δail −− q Actual data

0

−20

Derived model

101

Phase (deg)

0 −200 −400 −600

101

Coherence

1

0.5

0

101 Frequency (Rad/Sec)

Fig. 49.20 Response comparison using frequency-sweep input (ılat  q)

Magnitude (DB)

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs 20

δele −− p Actual data Derived model

0

−20

1251

101

Phase (deg)

200 0 −200 −400

101

Coherence

1

0.5

0

101 Frequency (Rad/Sec)

Magnitude (DB)

Fig. 49.21 Response comparison using frequency-sweep input (ıele  q) δele −− q

20

0

−20

Actual data Derived model

101

Phase (deg)

200 0 −200 −400

101

Coherence

1

0.5

0

101 Frequency (Rad/Sec)

Fig. 49.22 Response comparison using frequency-sweep input (ıele  p)

1252

F. Wang et al.

Table 49.7 Parameters determined via flight test and CIFER Parameter Physical meaning Aa;up D 0:4900 rad s1 Bb;up D 0:4900 rad s1 Ab;up D 0:2745 rad s1 Ba;up D 0:2745 rad s1 Aa;dw D 0:1217 rad s1 Bb;dw D 0:1217 rad s1 Ab;dw D 0:0450 rad s1 Ba;dw D 0:0450 rad s1

49.5

DC gain from stabilizer bar pitching angle to upper rotor longitudinal flapping angle DC gain from stabilizer bar rolling angle to upper rotor lateral flapping angle DC gain from stabilizer bar rolling angle to upper rotor longitudinal flapping angle DC gain from stabilizer bar pitching angle to upper rotor lateral flapping angle DC gain from elevator input to lower rotor longitudinal flapping angle DC gain from aileron input to lower rotor lateral flapping angle DC gain from aileron input to lower rotor longitudinal flapping angle DC gain from elevator input to lower rotor lateral flapping angle

Model Verification

It is always a good practice to verify a derived system model with actual input and output data. In this section, a comprehensive evaluation on the fidelity of the obtained nonlinear model of the ESky Big Lama (fixed-pitch coaxial helicopter) is shown. Four manual flight tests were carried out, which include: 1. Aileron channel perturbation with platform rolling left and right 2. Elevator channel perturbation with platform pitching forward and backward 3. Throttle channel perturbation with platform flying up and down 4. Rudder channel perturbation with platform yawing clockwise and anticlockwise In each of these four flight tests, the pilot was asked to agitate only one of the four input channels. However, to make sure the helicopter position does not drift too much, minor off-axis inputs were also issued to lightly counter the cross-couplings between the channels. The time-domain results are shown from Figs. 49.23 to 49.26. Based on the recorded inputs, the transient response of the UAV attitudes, angular rates, and body-frame velocities are calculated by a MATLAB simulation program with the aforementioned nonlinear mathematical model (dashed lines in the figures). They are plotted together with the in-flight true data obtained by the onboard sensors (solid lines in the figures). The matching between the two is quite perfect. Note that for the roll and pitch angular rate dynamics, both the on-axis and the offaxis responses match very well, indicating a precise formulation of the coupling terms. Some minor mismatches are caused by the ignorance of high-frequency dynamics when formulating the model, especially for the motion of rotor flapping, which is theoretically highly complicated. Other discrepancies come from ground effect, wind disturbances, and measurement noises present in practical flight tests. In general, this is an accurate cross-coupled model for a fixed-pitch coaxial UAV with low maneuvering speed.

q (rad/s)

p (rad/s)

δele (−1, 1)

δail (−1, 1)

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

1253

1 0 −1

0

5

10

15

20

0

5

10

15

20

0

5

10

15

20

0

5

10

15

20

1 0 −1 2 0 −2

2 0 −2 time (s)

q (rad/s)

p (rad/s)

δele (−1, 1)

δail (−1, 1)

Fig. 49.23 Responses from aileron input perturbation

1 0 −1

0

2

4

6

8

10

12

14

0

2

4

6

8

10

12

14

0

2

4

6

8

10

12

14

0

2

4

6

8

10

12

14

1 0 −1 2 0 −2

2 0 −2 time (s)

Fig. 49.24 Responses from elevator input perturbation

1254

F. Wang et al.

δthr (−1, 1)

1

0

−1

0

5

10

15

20

0

5

10 time (s)

15

20

w (m/s)

2 0 −2

Fig. 49.25 Responses from throttle input perturbation

δrud (−1, 1)

1

0

−1

0

2

4

6

8

10

12

14

16

0

2

4

6

8 time (s)

10

12

14

16

r (rad/s)

4 2 0 −2 −4

Fig. 49.26 Responses from rudder input perturbation

49 Flight Dynamics Modeling of Coaxial Rotorcraft UAVs

49.6

1255

Conclusions

This book chapter has demonstrated the main procedures and challenges in modeling flight dynamics of aerial vehicles for the purpose of UAV development. By using the fixed-pitch and variable-pitch coaxial helicopters as examples, the model formulation, parameter identification, and model verification methods are explained in sequence with an adequate level of complexity for people with general engineering backgrounds. For the model formulation, all analysis starts from the general working principle and model structure of the targeted systems. Formulas to describe the 6DoF kinematics and rigid-body dynamics are listed, and they are universal to all other types of UAVs. The general composition of forces and torques exerted on the UAV fuselage is more or less the same for the fixed-pitch and variable-pitch coaxial helicopters. However, the detailed formulation of individual force and torque generation are different and thus explained separately. For the model-based parameter identification, the identification methods are categorized into three types, namely, direct measurement, test bench experiments, and flight tests. The ESky Big Lama, belonging to the fixed-pitch coaxial helicopter, is chosen to be a case study to illustrate some useful test bench and flight test setups to determine the model parameters. After all, a nonlinear model of the ESky Big Lama flight dynamics is fully derived with all parameters identified. Finally, model verification is done to prove the feasibility of the whole methodology. By comparing simulation data with the actual in-flight data, the fidelity of the derived nonlinear model is guaranteed. It is also worth noting that the derived model has been actually used in control law design for an indoor coaxial UAV, and good hovering performance has been achieved (Wang et al. 2012). It is believed that with the proposed systematic modeling methodology, the readers will get insightful knowledge about UAV modeling and parameter identification. Hopefully, these methods can be utilized and applied to other types of aerial platforms, too. It is also welcomed that more detailed aerodynamic formulations and other innovative parameter identification methods can be shared in future and thus complement this work.

References C. Bermes, Design and dynamic modeling of autonomous coaxial micro helicopters. Dissertation, Eidgenossische Technische Hochschule ETH Zurich, 2010, Nr. 18847 A.R.S. Bramwell, G. Done, D. Balmford, Bramwell’s Helicopter Dynamics, 2nd edn. (Butterworth-Heinemann, Oxford, 2001) G. Cai, B.M. Chen, T.H. Lee, Unmanned Rotorcraft Systems. Advances in Industrial Control Series (Springer, New York, 2011) G. Cai, B.M. Chen, T.H. Lee, K.Y. Lum, Comprehensive nonlinear modeling of a miniature unmanned helicopter. J. Am. Helicopter Soc. 57, 1–13 (2012) C.P. Coleman, A survey of theoretical and experimental coaxial rotor aerodynamic research, NASA technical paper 3675 (1997)

1256

F. Wang et al.

Y. Deng, R. Tao, J. Hu, Experimental investigation of the aerodynamic interaction between upper and lower rotors of a coaxial helicopter. ACTA Aeronaut. ET Astronaut. Sin. 24(1), 10–14 (2003) C.M. Harris, Shock and Vibration Handbook, 4th edn. (McGraw-Hill, New York, 1996) R.K. Heffley, M.A. Mnich, Minimum-complexity helicopter simulation math model, NASA technical report 177476 (1988) W. Johnson, Helicopter Theory (Dover, Mineola, 1994) H.W. Kim, R.E. Brown, Coaxial rotor performance and wake dynamics in steady and manoeuvring flight, in American Helicopter Society 62nd Annual Forum Proceedings, Phoenix, 2006, vol. 1, pp. 20–40 J.G. Leishman, Principles of Helicopter Aerodynamics (Cambridge University Press, Cambridge/ New York, 2006) J.W. Lim, K.W. McAlister, W. Johnson, Hover performance correlation for full-scale and modelscale coaxial rotors. J. Am. Helicopter Soc. 54(3):32005-1–32005-14 (2009) B. Mettler, Identification Modeling and Characteristics of Miniature Rotorcraft (Kluwer, Norwell, 2002) P. Mukherjee, S.L. Waslander, Modeling and multivariable control techniques for small coaxial helicopters, in AIAA Guidance, Navigation, and Control Conference, Portland, 2011. AIAA2011-6545 D. Neamtu, R. Deac, R.D. Keyser, C. Ionescu, I. Nascu, Identification and control of a miniature rotorcraft Unmanned Aerial Vehicle (UAV), in AQTR’10: Proceedings of the IEEE International Conference on Automation, Quality and Testing, Robotics, Cluj-Napoca, 2010, pp. 1–6 K. Nonami, F. Kendoul, S. Suzuki, W. Wang, D. Nakazawa, Fundamental modeling and control of small and miniature unmanned helicopters, in Autonomous Flying Robots, ed. by K. Nonami et al. (Springer, Tokyo, 2010), pp. 33–60 O. Rand, V. Khromov, Aerodynamic optimization of coaxial rotor in hover and axial flight, in 27th International Congress of the Aeronautical Sciences, Nice, 2010, pp. 1–13 A. Rosen, D. Seter, Vertical autorotation of a single-winged samara, in Proceedings of the ASME Joint Applied Mechanics/Bioengineering Conference, Ohio State University, Columbus, 16–19 June 1991 D. Schafroth, C. Bermes, S. Bouabdallah, R. Siegwart, Modeling and system identification of the muFly micro helicopter. J. Intell. Robot. Syst. 57(1–4), 27–47 (2010) M.K. Taylor, A balsa-dust technique for air-flow visualization and its application to flow through model helicopter rotors in static thrust, NACA technical note 2220 (1950) S. Tsach, A. Tatievsky, L. London, Unmanned Aerial Vehicles (UAVs), in Encyclopedia of Aerospace Engineering, ed. by R. Blockley, W. Shyy (Wiley, Hoboken, 2010) B. Wang, F. Wang, B.M. Chen, T.H. Lee, Robust flight control system design for an indoor miniature coaxial helicopter, in Proceedings of the 10th World Congress on Intelligent Control and Automation, Beijing, 2012, pp. 2918–2924

Modeling of a Micro UAV with Slung Payload

50

Ying Feng, Camille Alain Rabbath, and Chun-Yi Su

Contents 50.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1258 50.2 Modeling for Quadrotor UAV with Slung Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1259 50.2.1 Quadrotor UAV Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1259 50.2.2 Modeling of External Slung Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1261 50.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1266 50.3.1 Stabilization Analysis of UAV with Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1267 50.3.2 Estimation of Payload Position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1268 50.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1270 50.5 Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1271 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1271

Abstract

In this chapter, the mathematical modeling and simulation of the micro UAV with a payload is derived. When the load is slung underneath the UAVs by cable, the flight dynamics of the UAVs will be altered, which makes the stability of the UAVs disturbed. The unstable oscillation may occur to degrade the performance of the UAVs, and the accurate placement of the load will be affected. Unlike the external disturbance, the negative effects are related to the characteristics of the UAV and the payload. In order to make the UAVs have the ability to adopt the change of the system dynamics and reduce the effects caused by the swing of the load, one modeling method of micro UAVs with single slung payload is addressed. The slung payload is treated as a pendulum-like mass point, and the coupling factors between the UAVs and the payload are considered in the Lagrangian formulation. The conducted model can be used to estimate the

Y. Feng () • C.A. Rabbath • C.-Y. Su Department of Mechanical and Industrial Engineering, Concordia University, Montreal, QC, Canada e-mail: [email protected]; [email protected]; [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 108, © Springer Science+Business Media Dordrecht 2015

1257

1258

Y. Feng et al.

negative effects acting on the UAVs, and it also can be used to assist estimating the trajectory of the load, which provides the possibility to improve the accuracy of placement of the loads.

50.1

Introduction

In recent years, unmanned aerial vehicles (UAVs) have become useful mobile platforms in industrial and academic fields, and their potential applications in numerous areas and their scientific significance in academic research have attracted more attentions (MacKunis et al. 2010; Mahony et al. 2012). Among various UAVs, the micro UAVs can provide an added measure of safety, security, and convenience for different applications, specifically, the micro UAV helicopters are suitable for a limited area or application by using their specific characteristics such as hovering and vertical landing (Marconi et al. 2011; Ryll et al. 2012; Tayebi and McGilvray 2006). In this chapter, one typical micro UAV, quadrotor-type aircraft, is chosen as the platform for the research work. As a versatile aerial vehicle, one of the potential tasks of the UAV is to carry loads hanging in cable underneath the vehicles. However, carrying a slung load will take a new challenge since the flight characteristics of the UAVs is altered. Besides increasing the mass of UAVs, the slung load will induce oscillations that degrade the UAV’s performance. For example, unstable oscillations may occur at high speeds due to the different aerodynamics of the slung loads that results in the crash. Also, the stability of UAV will be disturbed by the slung load, which slows or prevents an accurate placement of the loads. For example, placing a suspended payload accurately on a desired location, the swing load will degrade the placement precision, also the swing load can couple with the vehicle motion leading to poor flight handling performance with a potential for instability. There is, therefore, an interest in the scientific community to develop technologies that can address the challenge in operating UAVs with slung loads (Cruz and Fierro 2012; Min et al. 2011). The modeling and control for vehicles carrying external suspended loads have been paid attention since the 1950s (Fusato et al. 2001; Hoh and Heffley 2006). Previous work addressed for single-helicopter or multi-helicopter with slung loads (Cicolani and Kanning 1992; Ronen 1985; Theron et al. 2005). Due to the coupling between the helicopter and the loads, it may not be suitable to deal with the oscillation or the instability caused by the slung loads as the external disturbance or the parameter uncertainty directly (Yang et al. 2011), and the related works have explored the specific or general dynamic models of the helicopter system coupled with the swing of the loads, and the control methods are designed for the stabilization of the helicopter systems (Cicolani and Kanning 1992; Fusato et al. 2001). With the development of the unmanned aerial vehicles over the last few years, the flight stabilization of the UAVs with slung loads poses novel problems for the research of UAVs (Bernard et al. 2008; Bisgaard et al. 2010). The characteristics of the UAVs are different with the conventional helicopters, which makes the available methods may not be shifted to the UAV systems directly (Thanapalan and Wong 2010; Zameroski et al. 2011). For the quadrotor used in this research platform, it is

50 Modeling of a Micro UAV with Slung Payload

1259

a four-rotor helicopter, which can be treated as an underactuated, dynamic vehicle with four inputs (four rotors) and six output coordinates (Castillo et al. 2005a; Peng et al. 2009). Different with the conventional helicopters changing the pitch and roll control torques by swashplate, in a quadrotor, rotors are paired to spin in opposite directions so that the quadrotor-type UAV has high payload and relatively easily balanced venter of mass. The quadrotor can be controlled by varying the angular speed of each rotor and moving forward by pitching. These characteristics of quadrotor make it necessary to consider the modeling method of the micro UAVs with payload, which can be used to estimate the performance of the micro UAV systems and to improve the precision of the payload location by choosing the proper cable length, UAV/load mass ratios, etc. In this chapter, the modeling of quadrotor flying with single slung load is addressed. The approach taken in this chapter is to develop a systematic analytical formulation for general micro UAV systems with slung payload. The micro UAV systems with payload are treated as a class of multibody dynamic systems consisting of two rigid bodies connected by massless cable, and the cable is considered as massless and inelastic link. Therein, the mathematical model is derived from the Newton-Euler equations for rigid bodies and the Langrange’s equations for general dynamic systems. Besides the stabilization analysis of the micro UAV system slung payload, the position of the payload can be obtained by calculating the suspension angles of the payload, which is useful to estimate the position of the payload.

50.2

Modeling for Quadrotor UAV with Slung Payload

In this section, the derivation of a mathematical model of the micro UAV with slung load is presented. As an illustration, one quadrotor-type UAV with slung load is addressed. The quadrotor system with payload is treated as the rigid body motion of slung-load systems, and the effects caused by the pendulum-like behavior of the slung load are considered in the general simulation model. The modeling method discussed in this chapter is to develop a systematic analytical formulation for micro UAVs with slung payload, which can be utilized to improve the control performance of the UAVs. Firstly, the dynamic model for quadrotor is introduced briefly.

50.2.1 Quadrotor UAV Dynamic Model Taking inspiration from the work addressing on the classical helicopter, a quadrotor UAV system is studied in this chapter (Castillo et al. 2005a; Min et al. 2011). As shown in Fig. 50.1, the fuselage dynamics and the coordinate system can be divided into an earth frame fEg and a body frame fBg, which are used to describe the relative motions between the two coordinate frames. In order to simplify the highly nonlinear factors in quadrotor UAV system (Min et al. 2011), the following assumptions can be considered in developing the mathematical model of the quadrotor UAV, which are made based on the slower speed and lower altitude:

1260

Y. Feng et al.

Fig. 50.1 Schematic of quadrotor UAV dynamics

Assumption 1. The center of mass and the body frame origin are assumed to coincide. Assumption 2. Interaction with ground or other surfaces is neglected. Assumption 3. The body is rigid and symmetrical. Utilizing the above assumptions and considering the hovering and flying with low-speed situation, the quadrotor UAV dynamic model (50.1) can be derived via a Lagrange approach to describe a six-degree-of-freedom rigid body model (Alexis et al. 2011), driven by forces and moments as follows:

xR D yR D zR D R D R D R D

U1 .cos  sin  cos C sin  sin / M U1 .sin  sin  cos  cos  sin / M U1 .cos  cos / g M   Jy  Jz l P P C U2  Jx Jx   l Jz  Jx C U3 P P Jy Jy   Jx  Jy 1 C U4 P P Jz Jz

(50.1)

50 Modeling of a Micro UAV with Slung Payload

1261

where .x; y; z/ are three positions; .; ; / three Euler angles; representing pitch, roll and yaw, respectively; g is the acceleration of gravity; M is the mass of the UAV; Jx , Jy , Jz are the inertia moments applied to the center of mass of UAV; and l is the distance between the propeller and the center of mass of UAV. The system inputs are posed U1 , U2 , U3 , and U4 , which can be defined as follows (Castillo et al. 2005a): U1 D F1 C F2 C F3 C F4 U2 D F3  F1 U3 D F4  F2 U4 D F1 C F3  F2  F4 where F1 , F2 , F3 , and F4 represent control forces generated by four rotors, which are proportional to the square of the angle speed. U1 is the sum of the individual thrusts of each rotor. U2 and U3 can change the roll torque and pitch torque, respectively. The yaw motion can be changed by U4 while keeping U1 as constant. The characteristics of the quadrotor make the quadrotor-type UAVs with good performance, such as the high payload and relatively easily balanced center of mass.

50.2.2 Modeling of External Slung Payload In order to obtain the general system equations of motion, the underslung payload is considered as a point mass that behaves like a spherical pendulum from a single point in this chapter. As shown in Fig. 50.2, the payload is approximated as a point mass with an aerodynamic drag force acting on it (Chen 2009). The equations that describe the load dynamics are obtained by considering motion with reference to the longitudinal suspension angles ˛x in the x-z plane and ˛y in the y-z plane. The investigation of the dynamics of the payload is changed to discuss the motion of the mass point, and the position of the payload is determined by the length of the cable and the suspension angles ˛x and ˛y .

50.2.2.1 Payload Dynamics Since the payload is modeled as a three-dimensional point mass pendulum, the position of the slung payload in the frame fBg can be defined as 3 0 rB D 4 0 5 L 2

(50.2)

where L is the distance between the mass point of the UAV and the payload. Remark 1. Based on Assumptions 1 and 3, the hook point is assumed under the mass point of the UAV, and the length of the hook with respect to the UAV Ld is considered as the distance between the mass point of the UAV and the UAV. Then, it

1262

Y. Feng et al.

Fig. 50.2 The point mass slung payload model

has L D Lt C Ld , where Lt is the length of the cable. When Lt  Ld , the length of the cable Lt can be treated as L for convenience. To make the payload independent of the UAV attitude change, the payload position will be derived using the earth-fixed coordinates, then the position of the payload with respect to the center of mass of UAV is 3 0 r D R4 0 5 L 2

(50.3)

where 2

cos ˛y R D Rot.˛y /Rot.˛x / D 4 0  sin ˛y 2 cos ˛y D4 0  sin ˛y

32 3 1 0 0 0 sin ˛y 1 0 5 40 cos ˛x  sin ˛x 5 0 cos ˛y 0 sin ˛x  cos ˛x 3 sin ˛x sin ˛y  cos ˛x sin ˛y 5 (50.4) cos ˛x  sin ˛x sin ˛x cos ˛y  cos ˛x cos ˛y

and Rot.˛x / and Rot.˛y / are rotational matrices.

50 Modeling of a Micro UAV with Slung Payload

1263

The position of the payload using earth-fixed coordinates is redefined as 3 2 3 2 3 2 x x  L cos ˛x sin ˛y L cos ˛x sin ˛y 5 D 4 y  L sin ˛x 5 rL D 4 y 5 C 4 L sin ˛x L cos ˛x cos ˛y z  L cos ˛x cos ˛y z

(50.5)

then the absolute velocity rPL of the payload is given by 2

3 xP C L.˛Py cos ˛x cos ˛y  ˛P x sin ˛x cos ˛y / 5 rPL D 4 yP  L˛P x cos ˛x zP C L.˛P x sin ˛x cos ˛y  ˛P x cos ˛x sin ˛y /

(50.6)

To derive the dynamic model of the slung payload, the energy conservation property by means of Lagrange’s equations is applied. The effects due to the motion of the quadrotor and the characteristics of the payload are involved in the investigation of the suspension angles. Define the Lagrangian function ƒ as ƒD V

(50.7)

where ƒ D   V,  is the kinetic energy, and V is the potential energy. The kinetic energy of the system  is the sum of the translational kinetic energy of the payload t and rotational kinetic energy of the payload r . Therein, the translational kinetic energy is defined as t D

1 ML rPLT rPL 2

(50.8)

and the rotational kinetic energy is defined as r D

1 1 ML L2 ˛P x2 C ML L2 ˛P y2 2 2

(50.9)

and the potential energy of the payload is V D ML L cos ˛x cos ˛y . The equations of motion for the payload (no direct force acting on the payload) can be derived from the general form of Lagrange’s equations as d dt d dt where

@ƒ @˛P x

and

@ƒ @˛P y





@ƒ @˛P x @ƒ @˛P y

 

@ƒ D0 @˛x

(50.10)



@ƒ D0 @˛y

(50.11)



are the momentum conjugate to ˛x and ˛y , respectively.

1264

d dt

Y. Feng et al.

  d @ƒ @ƒ Based on the above definition of Lagrangian function ƒ, each item dt @˛P x , @˛x ,   @ƒ @ƒ , and @˛ can be derived by using mathematical softwares, such as Maple. @˛P y y

Remark 2. Solving the equations defined in (50.10) and (50.11), the symbolic expressions for ˛R x and ˛R y can be calculated by software, which are the functions of ˛x ; ˛y ; ˛P x ; ˛P y ; x; P y; P zP; x; R y; R zR, and Lc . 

˛R x ˛R y

 D F.L; ˛x ; ˛y ; ˛P x ; ˛P y ; x; P y; P zP; x; R y; R zR/

(50.12)

However, the expressions of ˛R x and ˛R y are too complicated to be implemented for the real-time application. Usually, the small-angle assumption sin ˛x ! ˛x and sin ˛y ! ˛y are used, then calculation of the angle acceleration of payload ˛R x and ˛R y can be simplified. Remark 3. In some special cases, the high speed or high acceleration of the UAV may cause the high oscillation of the payload, the small-angle assumption is not applicable. For these cases, sin ˛x and sin ˛y can be expanded as Taylor ˛3

˛3

series sin ˛x D ˛x  3Šx and sin ˛y D ˛y  3Šy . Usually, the small-angle assumption is suitable for the regular operation of quadrotor flying with low speeds.

50.2.2.2 Dynamics of Micro UAV with Payload As a motivational example to show the effects acting on the micro UAV, the dynamics of the quadrotor carrying single payload by cable is discussed in this section. The quadrotor and the payload are considered two separate rigid bodies and the force acting on the axes of quadrotor, respectively. As shown in Fig. 50.3, the curvilinear motion of the payload will result in the force acting on the quadrotor, which can be separated in the tangential, normal directions. The formulation of the tangential acceleration a and the normal acceleration an can be obtained, respectively, using the theorem of motion of center of mass law of conservation of momentum, which are calculated through the suspension angles ˛x and ˛y . Then, the forces caused by the payload in respective axes are derived as follows: Fox D ML aox D ML a x cos ˛x  ML anx sin ˛x D ML ˛Rx L cos ˛y cos ˛x  ML ˛P x2 L cos ˛y sin ˛x Foy D ML aoy D ML ay cos ˛y  ML any sin ˛y D ML ˛Ry L cos ˛x cos ˛y  ML ˛P y2 L cos ˛x sin ˛y Foz D ML aoz  ML g

50 Modeling of a Micro UAV with Slung Payload

1265

Fig. 50.3 System dynamics of quadrotor with pendulum-like payload

Fig. 50.4 Mechanical model of quadrotor with single payload

D ML ˛R x L cos ˛y sin ˛x C ML ˛P x2 L cos ˛y cos ˛x CML ˛R y L cos ˛x sin ˛y C ML ˛P y2 L cos ˛x cos ˛y  ML g Using the Newton-Euler function, the mechanical model reflecting the force acting on the quadrotor is shown in Fig. 50.4. Therefore, the model of the quadrotor with payload can be expressed as

1266

Y. Feng et al.

.M C ML /xR D U1 .cos  sin  cos

C sin  sin /  Fox

.M C ML /yR D U1 .sin  sin  cos

 cos  sin /  Foy

.M C ML /Rz D U1 .cos  cos /  M g C Foz   l Jx  Jz P R P C U2 D Jx Jx   l Jz  Jx C U3 R D P P Jy Jy   R D P P Jx  Jy C 1 U4 Jz Jz

50.3

(50.13)

Simulation

In order to gain insight into the dynamics of the quadrotor with slung payload, the numerical simulation for the slung payload with quadrotor is presented in this section. The simulation parameters for the quadrotor are chosen as M D 1:52 kg, and its moments of inertia were estimated as 0:03, 0:03, and 0:04 kg  m2 for the x, y, and z body axes, respectively. The distance Ld between the center of the quadrotor and the anchor point is 20 cm. In order to show the effects caused by the slung payload, the linear quadratic regulator (LQR) control laws (Castillo et al. 2005b) are used in the position controller and orientation controller, and the control parameters are kept as the same for different mass of payload and length of cable. Remark 4. In this chapter, the main task is to discuss the modeling method of the micro UAV with payload, and the controller used in simulation has been discussed in the previous work for controller design of quadrotor- type UAV. As an illustration to show the system performance with slung payload, the controller used in this simulation can be replaced by other controllers for quadrotor UAV (Fig. 50.5).

Controller Reference +

Position Controller

Orientation Controller

Model of Micro UAV with Payload

Quadrotor Model

Payload

–

UAV system states

Fig. 50.5 The close-loop scheme of the quadrotor with payload

Payload Position

50 Modeling of a Micro UAV with Slung Payload 2

1267

1.4

2.2

1.2

2

1.8 1.6 1

1.4

1.8 z [m]

0.8 y [m]

x [m]

1.2 1

1.6

0.6

0.8

1.4

0.6

0.4

0.4 1.2

0.2

0.2

UAV Command

0 0

10

20

30

40

50

0

UAV Command 0

10

Time (s)

20

30

40

50

1

Time (s)

UAV Command 0

10

20

30

40

50

Time (s)

Fig. 50.6 The reference and UAV outputs without payload

Table 50.1 List of payload parameters Payload mass (g) 68 68

Cable length (cm) 10 70

Payload mass (g) 108 108

Cable length (cm) 10 70

50.3.1 Stabilization Analysis of UAV with Payload In this section, the modeling method for the quadrotor with payload derived in the previous section is utilized to estimate the unknown effects caused by the payload. The simulation performs the motion of the quadrotor from position (0 m, 0 m, 1 m) to (1.75 m, 1.2 m, 2 m). The quadrotor system simulation response without payload is shown in Fig. 50.6. In order to illustrate the coupling between the quadrotor and payload which is related to the length of the cable and the mass of the payload, the simulation is conducted with the same controller parameters working on the quadrotor without payload, then compare the simulation results by choosing different payload parameters listed in Table 50.1. Figures 50.7 and 50.8 show the trajectory oscillation caused by the slung payload ML D 68 g with different cables ML D 10 cm and ML D 70 cm. Based on the simulation results, the frequency of the oscillation caused by the slung load is related to the cable length, and the existence of the payload will cause the delay and the increase of the overshoot in z axis, when the controller keeps the same parameters. Figures 50.9 and 50.10 show the trajectory oscillation caused by the slung payload ML D 108 g with the cable length as Lt D 10 cm and Lt D 70 cm. Compared with the results in Figs. 50.7 and 50.8, the negative effects caused by the payload are increasing with the mass of the payload.

1268

Y. Feng et al.

2

2.5

1.4

1.8

1.2

1.6 1

1.4

2

1

0.8 z [m]

y [m]

x [m]

1.2

0.6

0.8

1.5

0.6

0.4

0.4 0.2

0.2

UAV Command

0 0

10

20

30

40

UAV Command

0 0

50

10

Time (s)

20

30

40

UAV Command

1

50

0

10

Time (s)

20

30

40

50

Time (s)

Fig. 50.7 The reference and UAV outputs with payload ML D 0:068 kg and length of cable Lt D 10 cm

2

1.4

2.5

1.8 1.2 1.6 1

1.4

2

1

0.8

z [m]

y [m]

x [m]

1.2

0.6

0.8

1.5

0.6

0.4

0.4 0.2 0.2

UAV Command

0 0

10

20

30

Time (s)

40

50

0

UAV Command 0

10

20

30

Time (s)

40

50

1

UAV Command 0

10

20

30

40

50

Time (s)

Fig. 50.8 The reference and UAV outputs with payload ML D 0:068 kg and length of cable Lt D 70 cm

By constructing the dynamic model for the UAV with payload, the coupling effects related to the length of cable and mass of the payload can be considered in the system dynamics, and these characteristics can be addressed in the control approaches, improving the effectiveness of the control performance.

50.3.2 Estimation of Payload Position Another purpose for modeling of UAV carrying payload is to estimate the position of the payload, improving the accuracy of placement of payload. Addressing this

50 Modeling of a Micro UAV with Slung Payload 2

1269 2.8

1.4

1.8

2.6

1.2 1.6

2.4 1

1.4

2.2

1

0.8

z [m]

y [m]

x [m]

1.2

0.6

0.8

2 1.8 1.6

0.6

0.4 1.4

0.4 0.2

0.2

UAV Command

0 0

10

20

30

40

0

50

1.2

UAV Command 0

10

20

30

40

UAV Command

1 0

50

10

Time (s)

Time (s)

20

30

40

50

Time (s)

Fig. 50.9 The reference and UAV outputs with payload ML D 0:108 kg and length of cable Lt D 10 cm

2

2.8

1.4

1.8

2.6

1.2 1.6

2.4 1

1.4

2.2

1

0.8

z [m]

y [m]

x [m]

1.2

0.6

0.8

2 1.8 1.6

0.6

0.4 1.4

0.4 0.2 0.2

UAV Command

0 0

10

20

30

Time (s)

40

50

UAV Command

0 0

10

20

30

Time (s)

40

50

1.2

UAV Command

1 0

10

20

30

40

50

Time (s)

Fig. 50.10 The reference and UAV outputs with payload ML D 0:108 kg and length of cable Lt D 70 cm

task, the dynamics modeling for the quadrotor can be used to describe the motion of the payload. Following the definition in the earth-fixed frame fEg, the position of the payload can be obtained when the suspension angles of the payload ˛x and ˛y are determined. As an illustration, the trajectories of the payload are shown in Figs. 50.11 and 50.12, when the desired trajectories are set as xd D 1:2 sin.0:1t/ and yd D sin.0:1t/.

1270

Y. Feng et al.

2

1 0.5

1

y [m]

z [m]

1.5

0

0.5 −0.5 UAV

0 2

2

Command

0 y[m]

Payload UAV Command

−1

Payload

−1.5

−1

−0.5

0 −2

x[m]

−2

0

0.5

1

1.5

x [m]

Fig. 50.11 The positions of the reference, quadrotor, and payload with ML D 0:068 kg and Lt D 37 cm

2

1 0.5

1

y [m]

z [m]

1.5

0

0.5 −0.5 Command UAV Payload

0 2

y [m]

2 −1.5

0

0 −2

−2

x [m]

Payload UAV Command

−1 −1

−0.5

0

0.5

1

1.5

x [m]

Fig. 50.12 The positions of the reference, quadrotor, and payload with ML D 0:108 kg and Lt D 37 cm

50.4

Conclusion

One of the UAVs’ tasks is to carry the payload by cable, and the negative effects caused by the slung load are coupled with the UAV systems, degrading the system performance and altering the flight dynamics of UAVs, which poses a new problem for the operating UAVs with slung payload. In this chapter, the model of single UAV with single payload is addressed. The slung payload is treated as a pendulum-like mass point, and the dynamics of the UAV with payload is obtained by formulating Lagrangian function. The derivation of the suspended angle of the payload makes it possible to estimate the position of the payload. The results discussed in this chapter can be utilized for further development of stabilizing and high-performance control laws.

50 Modeling of a Micro UAV with Slung Payload

50.5

1271

Symbols

E B x, y, z   g M l Jx , Jy , Jz ML Lt Ld L ˛x ˛y

Inertial world frame UAV body frame Positron along fEg frame x-, y-, and z-axes Pitch angle of UAV Roll angle of UAV Yaw angle of UAV Gravitational acceleration constant The mass of UAV The distance between the propeller and the center of mass of UAV The inertia moments applied to the center of mass of UAV The mass of payload The length of the cable The length of the hook with respect to the UAV The distance between the mass point of the UAV and the payload The suspension angle of the payload x-z plane The suspension angle of the payload y-z plane

[m] [rad] [rad] [rad] Œm=s2  [kg] [m] Œkg  m2  [kg] [m] [m] [m] [rad] [rad]

References K. Alexis, G. Nikolakopoulos, A. Tzes, Switching model predicitive attitude control for a quadrotor helicopter subject to atmospheric disturbances. Control Eng. Pract. 19, 1195–1207 (2011) M. Bernard, K. Kondak, G. Hommel, A Slung load transportation system based on small size helicopters, in Autonomous Systems – Self-organization, Management, and Control (2008), pp. 49–61. doi:10.1007/978-1-4020-8889-6-6 M. Bisgaard, A. La Cour-Harbo, J.D. Bendtsen, Adaptive control system for autonomous helicopter slung load operations. Control Eng. Pract. 18, 800–811 (2010) L.S. Cicolani, G. Kanning, Equations of motion of slung-load systems, including multilift systems. NASA report paper, 1992 P. Castillo, R. Lozano, A. Dzul, Stabilization of a mini rotorcraft with four rotors. IEEE Control Syst. Mag. 25(6), 45–55 (2005a) P. Castillo, R. Lozano, A. Dzul, Modelling and Control of Mini-Flying Machines. Springer-Verlag Series in Advances in Industrial Control (Springer, New York, 2005b) C.Y. Chen, Multiple degree of freedom inverted pendulum dynamics: modeling, computation and experimentation. Ph.D Dissertation of University of Southern California, 2009 P. Cruz, R. Fierro, Agile load transportation: safe and efficient load manipulation with aerial robots. IEEE Robot. Autom. Mag. 19(3), 69–79 (2012) D. Fusato, G. Guglieri, R. Celi, Flight dynamics of an articulated rotor helicopter with an external slung load. J. Am. Helicopter Soc. 45(1), 2–13 (2001) R.H. Hoh, R.K. Heffley, Development of handling qualities criteria for rotorcraft with externally slung loads. NASA report, 2006 W. MacKunis, Z.D. Wilcox, M.K. Kaiser, W.E. Dixon, Global adaptive output feedback tracking control of an unmanned aerial vehicle. IEEE Trans. Control Syst. Technol. 18(6), 1390–1397 (2010) R. Mahony, V. Kumar, P. Corke, Multirotor aerial vehicles: modeling, estimation, and control of quadrotor. IEEE Robot. Autom. Mag. 19(3), 20–32 (2012) L. Marconi, R. Naldi, L. Gentili, Modelling and control of a flying robot interacting with the environment. Automatica 47, 2571–2583 (2011)

1272

Y. Feng et al.

B.C. Min, J.H. Hong, E.T. Matson, Adaptive robust control (ARC) for an altitude control of a quadrotor type UAV carrying an unknown payloads. Paper presented at the 2011 international conference on control, automation and systems, Gyeonggi-do, Korea, 26–29 Oct 2011 K. Peng, G. Cai, B.M. Chen, M. Dong, K.Y. Lum, T.H. Lee, Design and implementation of an autonomous flight control law for a UAV helicopter. Automatica 45(10), 2333–2338 (2009) T. Ronen, A Helicopter with a sling load. Ph.D Dissertation of Stanford University, 1985 M. Ryll, H.H. Bulthoff, P.R. Giordano, Modeling and control of a quadrotor UAV with tilting propellers. Paper presented at 2012 IEEE international conference on robotics and automation, Minnesota, 14–18 May 2012 A. Tayebi, S. McGilvray, Attitude stabilization of a VTOL quadrotor aircraft. IEEE Trans. Control Syst. Technol. 14(3), 562–571 (2006) K. Thanapalan, T.M. Wong, Modeling of helicopter with an under-slung load system. Paper presented at the 29th Chinese control conference, Beijing, 29–31 July 2010 J.N. Theron, E.P.N. Duque, L. Cicolani, Three-dimensional computational fluid dynamics investigation of a spinning helicopter slung load. NASA report, Document ID: 20070017931, 2005 Y. Yang, J. Wu, W. Zheng, Variable structure attitude control for an UAV with parameter uncertainty and external disturbance. Procedia Eng. 15, 408–415 (2011) D. Zameroski, G. Starr, J. Wood, R. LumiaRapid, Swing-free transport of nonlinear payloads using dynamic programming. J. Dyn. Syst. Meas. Control 130, 041001 (2011)

Command and Control of Autonomous Unmanned Vehicles

51

David H. Scheidt

Contents 51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1274 51.2 Autonomous Unmanned Air Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1276 51.2.1 The Case for Autonomous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1278 51.2.2 C2 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1278 51.2.3 The Organic Persistent Intelligence Surveillance and Reconnaissance System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1286 51.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1297 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1297

Abstract

Motivated by Generals Rommel and Guderian’s innovative command and control techniques used in Europe in 1940, this chapter begins by using information theory to examine unmanned air vehicle (UAV) command and control (C2). The information-theoretic analysis provides a justification and uses cases for autonomous UAVs. An autonomous unmanned vehicle system “Organic Persistent Intelligence Surveillance and Reconnaissance” (OPISR) that is designed to duplicate Guderian’s innovations is introduced. OPISR is an autonomous unmanned vehicle system that combines the immediate response to tactical ISR needs provided by organic assets with the time-on-station, minimal logistics provided by persistent unmanned systems. OPISR autonomous vehicles collectively interpret real-time tactical intelligence surveillance and reconnaissance (ISR) objectives submitted by any number of disadvantaged users, gather the required ISR data, and return the needed intelligence directly to the affected user. OPISR is an ad hoc, decentralized system that requires no central base or authority and is capable of functioning in communications-denied environment. The chapter

D.H. Scheidt Johns Hopkins University Applied Physics Laboratory, Laurel, MD, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 110, © Springer Science+Business Media Dordrecht 2015

1273

1274

D.H. Scheidt

describes a series of experiments including 2011 experiments in which 16 fully autonomous unmanned vehicles, including 9 unmanned air vehicles, were used to simultaneously support mounted, dismounted and maritime users. During these experiments users provided abstract mission-level ISR needs to the “vehicle cloud.” These needs were interpreted by the vehicles, which self-organized and efficiently achieved the user’s objectives.

51.1

Introduction

In the spring of 1940, the combined French, British, Dutch, and Belgian forces outnumbered their German counterparts in troops, mechanized equipment, tanks, fighter planes, and bombers. The ME109E German fighter aircraft was roughly equivalent to the British Spitfire, and the French CharB1 tank was superior to the German Panzer III. In addition, the allies were fighting on their home soil which greatly simplified their logistics. Yet in less than 6 weeks, the Belgians, Dutch, and French surrendered to the Germans, and the British retreated across the English Channel. Even though the allies had superior equipment and larger forces, they were defeated by the Germans who employed Auftragstaktik, a command and control technique that enabled “edge” war fighters to directly coordinate on tactical decisions using modern communications equipment (in WWII this was radio). Allied forces were forbidden to use radio because it “was not secure,” and allied maneuver decisions were made by generals at headquarters and based upon hand-couriered reports. German decisions were made on the fly by Panzer III commanders and JU-87 (Stuka) pilots conversing over the radio. By the time the French commanders met to decide what to do about the German advance, General Erwin Rommel and General Heinz Guderian’s Panzers had travelled over 200 miles and reached the English Channel. As demonstrated repeatedly in military history, including the German advance in 1940, the speed at which battlefield decisions are made can be a deciding factor in the battle. A process model that describes military command and control is the Observe, Orient, Decide, Act (OODA) loop described by Boyd (Fig. 51.1). Boyd shows that, in military engagements, the side that can “get inside the opponent’s OODA loop” by more rapidly completing the OODA cycle has a distinct advantage. In their influential book Power to the Edge, Alberts and Hayes use the term agility to describe an organization’s ability to rapidly respond to changing battlefield conditions. Modern warfare case studies, such as the Chechens against the Russians, and not-so-modern warfare, such as Napoleon at Ulm, indicate that agile organizations enjoy a decisive military advantage. Alberts points out that a common feature of agile organizations is an empowerment of frontline forces, referred to as “edge” war fighters. Commanders facilitate organizational agility by exercising “command by intent,” in which commanders provide abstract goals and objectives to edge war fighters who then make independent decisions based upon these goals and their own battlefield awareness. This empowerment of edge war fighters reduces the OODA loop at the point of attack, providing the desired agility.

51 Command and Control of Autonomous Unmanned Vehicles Fig. 51.1 Boyd’s Observe, Orient, Decide, Act (OODA) cycle models the military decision-making process. Military organizations that perform their OODA cycle more rapidly than opponents gain a substantial competitive advantage

orient

observe

1275

decide

Natural World

act

A distinguishing characteristic of the conflicts in Afghanistan and Iraq is the explosive growth in the use of unmanned air vehicles. Between the first and second Gulf Wars, unmanned vehicles transitioned from a novelty item to an indispensable component of the U.S. military. Field deployable organic unmanned air vehicles such as the AeroVironment Raven are essential equipment for the modern war fighter. The agility provided by field deployable vehicles comes at a cost, as the use of field deployable units increases logistics and workload demands on frontline forces. When compared to larger unmanned vehicles, field deployable units such as the Raven (Fig. 51.2) offer limited sensing and time-on-target capabilities. Medium-sized vehicles, such as the Boeing-Insitu ScanEagle and AAI Shadow, offer longer time on station and more capable payloads. Medium-sized vehicles also do not make logistics or workload demands on the edge war fighter. Large unmanned vehicles such as the General Atomics Reaper and Northrup Grumman Triton offer still more capable payloads and increased time on station and also do not increase edge war fighter logistics or workload. However, providing timely edge war fighter access to intelligence products produced by medium and large vehicles is a challenge because medium- and large-sized unmanned air vehicles produce massive amounts of data that is difficult to process and disseminate from centralized command posts. In fact, as reported by Ariel Bleicher, “In 2009 alone, the U.S. Air Force shot 24 years’ worth of video over Iraq and Afghanistan using spy drones.” The trouble is there aren’t enough human eyes to watch it all. The deluge of video data from these unmanned aerial vehicles, or UAVs, is likely to get worse. A single Reaper drone can record 10 video feeds at once, and the Air Force plans to eventually upgrade that number to 65. John Rush, chief of the Intelligence, Surveillance and Reconnaissance Division of the U.S. National

1276

D.H. Scheidt

Fig. 51.2 An AeroVironment Raven being launched

Geospatial-Intelligence Agency, projects that it would take an untenable 16,000 analysts to study the video footage from UAVs and other airborne surveillance systems. The intelligence, surveillance, and reconnaissance (ISR) capability represented by medium- and large-scale unmanned vehicles represents a tremendous potential for the edge war fighter if only the information could be processed and distributed in time. For the edge war fighter to take advantage of the ISR capability represented by these assets, information relevant to that specific war fighter must be gleaned from the mass of information available and presented to the war fighter in a timely manner. This presents a challenge as crews analyzing UAV payload data (far fewer than Rush’s 16,000 analysts) are not appraised of the changing tactical needs of all war fighters, nor do the war fighters have access or the time required to select and access data from UAV sources. Currently, operation centers are used to gather and disseminate information from persistent ISR assets. This centralized information management process introduces a delay between the observation and transmission to the war fighter which reduces force agility and operational effectiveness. While U.S. soldiers are empowered to operate on “command by intent,” their ISR systems are all too frequently centralized systems reminiscent of the French command structure. For U.S. forces to become a fully agile force, the ISR systems supporting the U.S. soldier must be as agile as the soldier it supports. Agile unmanned vehicle systems require that some decisions are made at the edge nodes; therefore to become agile, unmanned vehicles must become autonomous.

51.2

Autonomous Unmanned Air Vehicles

UAVs currently in use are described as unmanned strictly because no human rides inside the air vehicle; however, the manual labor required to operate an air vehicle

51 Command and Control of Autonomous Unmanned Vehicles

1277

remains largely unchanged as ground-based UAV pilots perform similar tasking to airborne pilots. Because flight procedures have not evolved to match the removal of the human from the air vehicle, the manpower required to operate an UAV equals or exceeds the manpower required to operate a manned aircraft. Because UAV pilots, by definition, fly the vehicle remotely, there is no longer a requirement that a pilot be co-located with the UAV area of operations, and it is not uncommon particularly for large, expensive UAVs for the UAV pilot to operate the vehicle from an office-like environment thousands of miles from the operating area. As the size, range, and capability of a UAV diminish, its use becomes more tactical, and the UAV pilot is located closer to the operating area with medium-sized UAVs such as the Boeing ScanEagle and AAI Shadow being controlled from forward operating bases and the AeroVironment Raven being controlled by a tactical unit in the field. Clearly remote vehicle operations significantly reduce pilot risk when compared to manned aircraft, yet this risk reduction comes at the cost of tactical awareness and involvement. Remote UAV pilots, when emplaced in a hierarchical command structure, have a reduced ability to assist in agile operations; without being immersed in the environment and able to, when required, communicate with actors outside of the current command structure, UAV pilots cannot support tactical operations as well as a JU-87 pilot from 1940. Automated aircraft control techniques are emerging that allow UAV designers to design and field UAVs with varying degrees of autonomy. For the purpose of discussion, three general levels of autonomy are defined: tele-operation, which is essentially no automation at all; automatic, in which UAVs perform simple actions that can be fully enumerated and tested during design; and autonomy, in which the vehicle independently devises a course of action in response to complex operating conditions. The distinction between automatic and autonomy is subtle but important. Most UAVs being used today are automatic. Engineers of automatic UAVs have designed in action-based commands for pilot use (e.g., follow this path, loiter here, land there), the response matrix has been exhaustively enumerated and tested by development engineers, and the task of managing uncertainty, and complexity, lies with the pilot. Autonomous UAVs, which are not currently in service, are capable of devising a course of action in response to a complex, uncertain situation that was not, in detail, examined by the engineer during design or by the pilot. In other words, an automatic UAV is capable of following a path; an autonomous UAV is capable of finding a path. The technology required to build autonomous UAVs is available today. Prototype autonomous UAVs that exhibit a variety of complex autonomous behaviors have been developed at academic institutions and industry research centers. Demonstrated capabilities involve relatively simple tasks including search, interdiction, tracking, obstacle avoidance, path planning, logistics, communications, launch, and recovery. While vehicle tasking remains simple, the complex unpredictable nature of the operating environment represents a complex problem requiring an autonomous, not automatic, solution. The bulk of autonomous UAV efforts are conducted at a relatively low technology readiness level (TRL) using low-cost rotorcraft in a laboratory (Kumar and Michael 2012; Bethke et al. 2008). More mature demonstrations

1278

D.H. Scheidt

of autonomous tier 1 and tier 2 UAVs have been demonstrated outdoors at U.S. government ranges (Scheidt et al. 2004; Kwon and Pack 2011; Tisdale et al. 2008). Arguably the most mature autonomous UAV effort is the DARPA Heterogeneous Airborne Reconnaissance Team (HART) program, providing automatic unmanned air vehicles with an autonomous tactical decision aide (TDA). In these systems simple actions such as waypoint following are performed independently by each UAV, and complex decisions are performed by the TDA algorithm located on the pilot’s computer. Prior to execution, the pilot reviews and approves (or disapproves and modifies) the plan. Proponents of HART correctly argue that the pilot is allowed to “auto-approve” plans, effectively making HART an autonomous system; however, the requirement to continually enable the pilot to review and approve all flight plans profoundly impacts the UAV system architecture and performance by delaying the exchange of information between the sensor and the UAV controller. The U.S. Army planned on partial fielding of HART in 2012 (Defense Systems staff 2012).

51.2.1 The Case for Autonomous Systems Technological availability does not mean that autonomous UAVs will, or should, be used in the field. Autonomous UAV use requires that, when compared to manned aircraft or tele-operated/automatic UAVs, autonomous UAVs provide some tangible benefit to the organization deploying the autonomous UAV. Three general benefits are commonly offered that could justify the use of autonomous UAVs: first, by reducing the manpower required to fly/operate the UAV, autonomous vehicles are less expensive to fly; second, because autonomous UAVs do not require constant communications with a base, they are less vulnerable to electronic warfare attacks and capable of electromagnetic stealth; and third, by making better, more timely decisions, autonomous UAVs can, in certain circumstances, provide more effective performance. This third argument that autonomous unmanned vehicles can improve mission performance and that this performance can be understood and predicted by viewing autonomy as a command and control technique is the central theme of this chapter.

51.2.2 C2 Fundamentals In Power to the Edge, command and control (C2) is defined as the “common military term for management of personnel and resources” but also gives the formal definition of command as found in the Joint Chiefs of Staff Publication, which subsumes some portions of control in that definition. Viewed as a black box the purpose of the C2 system use observations to produce decisions. C2, including UAV C2, involves the production and execution of decisions that, when executed, change the world in which the UAV is operating in ways that benefit the operator. That world (X) is described as a set of states, X D fx1 ; x2 ; : : : xn g, each one of which represents a unique configuration of actors and attributes within the natural world in which

51 Command and Control of Autonomous Unmanned Vehicles

1279

the UAV operates. The command and control system can be viewed as a transfer function (f .xi / ! xj ) that produces a state change in the world. The “quality” of each state can be determined by applying a mission-based fitness criteria to elements within the state. For example, if a UAV mission is to track a target, then states in which the target is within the field of view of the UAV’s sensor are evaluated as of higher quality than those states in which the target is not seen by the UAV, and an effective C2 system would cause state transitions that are of high quality when compared to alternative transitions. C2 is a constant battle between chaos and order, with order being state transitions designed to achieve mission goals that are instigated by the C2 system and chaos being unanticipated state transitions produced by adversaries, poorly coordinated teammates or random acts. C2 can be best understood as an information-theoretic problem. This is apropos as both the situational awareness upon which decisions are based as well as the decision products can be viewed as information-theoretic messages and both the C2 process and the natural world can be viewed as information transfer functions. For an interesting, albeit somewhat off-topic, discussion on the information-theoretic nature of physics, see Wheeler (1990). The information content (S ) of a message (m) is defined by Shannon (1948) as: S.m/ D log2 .1=P.m//

(51.1)

The more improbable the message being received, the larger the information content. For example, baring a highly improbable change to celestial mechanics, a message stating “tonight it will be dark” has zero information content because the a priori probability that it would be dark is one. By comparison, “tonight it will rain” has positive information content because the a priori probability that it would rain during a given evening is less than one. Information content is measured in bits, which are real numbers. Note that the term “bit” is overloaded, and information theory bits are not the same as computer science bits, which are integers. The state space (p) of the world in which our UAVs operate is the amount of bits required to uniquely identify all possible states in the world, which is referred to as the state space of the world. p D log2 jX j (51.2) Completely describing the world requires a message of length p. A communication that completely describes the natural world would require a message whose length would be effectively infinite as the natural world includes each blade of grass, molecule of air, quantum states of ions within each atom, and so forth. Fortunately, effective C2 of UAVs does not require such detailed knowledge, and effective UAV control can be provided using artifacts that are abstract, limited, and while potentially quite large in number, expressible in messages that are small enough to be used within a modern computer network. In fact, military C2 systems routinely express tactical “worlds” in finite languages such as the protocols used by the Global Command and Control System and the Link16 network. In practice, the size of a UAV’s world as represented by the C2 system can be dynamic, with the state

1280

D.H. Scheidt

space of the tactical world changing as artifacts enter, or leave, the operational area. That the complexity of a UAV’s world, as represented by the state space of the current situation, is subject to change is important to understanding how to control UAVs. Knowledge of the UAV’s world is rarely complete, and UAVs are expected to operate in the presence of varying degrees of uncertainty. Uncertainty in tactical information can be produced by errors in sensor systems, gaps in sensor coverage, or approximation error, which is the difference between the actual ground truth and the coding scheme selected. For example, if the unit representation of a coding scheme used to represent linear position is 1 m, then an approximation error of 0.5 m is unavoidable. Consider the message (m) that encapsulates a C2 systems’ current situational awareness (SA). When S.m/ ¤ p, uncertainty exists and additional information is required to produce complete SA. Now consider a second message (m0 ) that contains all of the missing information required to reduce uncertainty to zero. The information content of m0 is information entropy (H) of m, shown as S.m0 / D H.m/ D

P

1 P .x/ log P .x/

where x is the assemblage of information contained in m

(51.3)

The information content of m is “negentropy” (N) or “order,” Command and control is a battle between the forces of order and chaos, in which the command and control system seeks to generate order by forcing the world into the most beneficial state for the commander and the uncontrollable forces within the world, which may include adversaries, continuously generates entropy that moves the world away from the commander’s ideal state. Together, the information content of m and m0 represents the total amount of information potential within the system s.t: S.m0 / D p  S.m/

(51.4)

For each bit of order produced by the C2 system, entropy is reduced by a bit. Likewise, each bit of entropy produced by unanticipated change reduces order by 1 bit. As demonstrated by Guderian and Rommel, C2 systems are temporally sensitive. As time elapses the information content of a message that describes a dynamic scene decreases in proportion to the unpredictable change in the scene. An example of this are unexpected maneuvers made by a target after a sensor observation was made but before the execution of an response to the observation. The loss of information over time is defined as entropic drag () which is expressed mathematically by Scheidt and Schultz (2011) as: .x; t; t0 / D

H. x.t/j x.t0 // t  t0

(51.5)

51 Command and Control of Autonomous Unmanned Vehicles

1281

Note that entropic drag is specific value for each state. Control of UAVs operating in an uncertain world in which multiple states are feasible requires a consideration of all admissible states. The measurement of entropic change that incorporates all feasible states is the normalized form of entropic drag (norm ) that is the average entropic drag for all states within the system: norm .t/ D

X H. xi .t/j xi .t0 // 8xi

.t  t0 /  jX j

(51.6)

UAVs acquire, process, and share information over a control infrastructure that includes onboard network and processing, radio downlinks and uplinks, and offboard processing. When processing information, two key characteristics of the control infrastructure are latency (ı0 ), which is the unavoidable delay in processing an information packet regardless of packet size, and bandwidth (ˇ), which is the rate in bits per second at which bits of information can be processed irrespective of the latency. The delay time (ı) required to process a message m of information can be viewed as   sizem (51.7) ım D ı0 C ˇ When communicating information from a sensor to a user over a communications link, large amounts of information take more time to transmit and process than small amounts of information. If the information observed by the sensor is a dynamic scene, which is often the case for UAVs, increases in information gathering cause an increase in processing delays that, in turn, cause an increase in entropy. This presents a paradox with respect to attempts to increase information content by increasing the quantity of data. This paradox is defined by the relationships between Eqs. (51.5) and (51.7) that was described by Scheidt and Pekala (2007) and is shown in Fig. 51.3. The figure shows uncertainty as a function of the unit resolution used to describe a dynamic scene using a constant cognitive bandwidth and also assuming that all sensor data is correct. In the plot the highest unit resolution (e.g., least precision) is shown on the right, while the smallest unit resolution (most precision) is shown on the left. Interpreting the plot from right to left shows, as one might expect, that initially increasing the information content of a message substantially reduces uncertainty. As precision increases, the time requires to communicate and process the information increases, which increases information loss due to the entropic drag. Eventually an uncertainty minima is reached at which the information gain from additional information content within the initial message equals the information loss due to entropic drag. Any attempt to use additional precision beyond the minima produces net loss in information content. If the function of the UAV is intelligence, surveillance, and reconnaissance (ISR), then mission goals are to provide information on targets of interest with minimal uncertainty. Counterintuitively, transmitting all possible information on a target may not be the best approach for an ISR system. The plot in Fig. 51.3 provides a guide as to the optimal amount of data that should be gathered, processed, and

1282

D.H. Scheidt 4.4 4.2 4 Uncertainty

Fig. 51.3 When information describing a dynamic (changing) scene is sent over, a network uncertainty can be minimized by using the optimal resolution (shown by the local minima). The local minima shown is the balance point where the rate of information gain equals the rate of entropy (entropic drag)

3.8 3.6 3.4 3.2 3 2.8 10−6

10−5

10−4 10−3 resolution

10−2

10

Table 51.1 The complexity and rate of unpredictable change vary by mission class Use case Description Complexity, P(x) Entropic drag, (t) Strategic

Operational

Focused tactical

Multiunit tactical

Strategic missions are dedicated to acquiring information on standing infrastructure Operational missions are dedicated to acquiring information on the general movement and condition of large units (e.g., company size or larger) Small units (individuals, squads, or platoons) performing tightly defined, focused missions in isolation Small units operating as part of a large engagement that involves multiple units

–

None

Low

Low

Low

High

High

High

provided from a UAV to the UAV operator. The local minima in the plot represents the optimal amount of information that should be transmitted. UAV system that transmit too much information, represented by high-resolution data on the left of the graph, reduces the net information provided due to the large loss of information across all data due to entropic drag. To further understand this relationship, let us examine bandwidth from Eq. (51.7) and its relationship to Eqs. (51.5) and (51.6) in more detail. This is accomplished by viewing canonical ISR missions and operating conditions through information theory and applying these views to differing forms of UAV control. Table 51.1 describes four general classes of ISR missions for UAVs and defines, in general terms, the complexity and entropic drag associated with those missions.

51 Command and Control of Autonomous Unmanned Vehicles

1283

Recall that UAVs may be controlled using three general methods: tele-operation, automatic control, and autonomous control. Descriptions of these methods are: • Tele-operation – The most common form of UAV control is tele-operation. Teleoperated UAVs use a ground-based pilot to fly the UAV using techniques that are identical those of human-piloted aircraft. Tele-operated UAVs utilize low-level control features found in an airplane cockpit as well and onboard sensor data on the pilot’s ground station that are duplicates of those found in the cockpit of a piloted aircraft. All decisions used to control tele-operated planes are made by the ground-based human pilot. • Automatic flight – UAVs that contain autopilots are capable of automatic flight. Automatic UAVs use autopilots to assure stable-controlled flight. The pilots of automatic UAVs provide waypoint locations that direct the path of the UAV. In addition to flying to pre-defined waypoints, automatic UAVs may be preprogrammed to handle simple changes in operating conditions; however, management of complex or unanticipated changes are handled by a ground-based pilot. • Autonomous flight – UAVs that contain an autopilot and an onboard, intelligent controller are capable of autonomous flight. Similar to automatic UAVs, autonomous UAVs provide stable flight between waypoints; however, unlike automatic UAVs, the intelligent controllers onboard an autonomous UAVs define new waypoints in response to unanticipated changes in operational conditions. Autonomous UAVs fundamentally change the relationship between the human and the UAV because autonomous UAVs, unlike tele-operated UAVs, automatic UAVs, and manned aircraft, do not require pilots. Autonomous UAVs do use human supervision; however, that supervision is performed at a higher level that the typical plane-pilot relationship. The relationship between an autonomous UAV and the supervising human resembles the relationship between a human pilot and an air traffic controller or, for Navy pilots, the Air Boss. Autonomous UAV operators supervise their UAVs by providing abstract, mission-level objectives as well as rules of engagements that are equivalent to the instructions provided to human pilots prior to a mission. During the mission autonomous UAVs devise a course of action aligned with these instructions in response to the current situation. As conditions change during the course of a mission, an autonomous UAV constantly modifies the course of action in accordance with the mission-level objectives. Unlike automatic UAVs, autonomous UAVs respond to complex situations that were not explicitly considered during UAV design. The different forms of UAV control demand different levels of communications. Direct control of UAV control surfaces requires that tele-operated UAV pilots use low-latency, high quality of service communications to command UAVs. For tele-operated UAVs even small perturbations in service run the risk of loss of control and catastrophic failure. Automatic UAVs are more forgiving, as the pilot is only required to provide guidance at the waypoint level. The communications requirement for automatic UAVs is determined by the rate of change of those operational elements that dictate the mission pace. For example, if the UAV is engaged in an ISR task to track a specific target, the UAV communications

1284

D.H. Scheidt

infrastructure used to control the UAV must be capable of providing target track data to the pilot prior to the target exiting the UAV’s field of view. Depending upon the nature of the target, the response time required for automatic UAVs can range from sub-second intervals to minutes. Autonomous vehicles are the most forgiving of communications outages and delays. In fact, autonomous vehicles are capable of performing without communication to human operators during an entire mission. The ability to function without continual human supervision changes operator and designer perspectives on UAV communications from being a requirement to being an opportunity. When autonomous UAVs can communicate, either to humans or other vehicles, mission performance is improved by the sharing of information and collaborating on decisions; however, when communications are not available, autonomous UAVs are still capable of fulfilling the mission. A synopsis on the communications availability, in terms of latency and bandwidth, for four different classes of conditions is provided in Table 51.2. When considering whether a UAV control decision should be made be a human operator or by a control processor onboard a UAV, three criteria should be considered: (1) what is the quality of decisions made by the human/machine, (2) what are the ethical and legal requirements for making decisions, and (3) what accessible information is available to the human and the machine? There are ethical and legal advantages and disadvantages for both human and machine decision-making. Arkin provides an excellent overview of the legal and ethical issues, concluding that no consensus exists that would favor human control over machine control or vice versa (Arkin 2009). Regarding the ability to make higher-quality decisions, anecdotal evidence suggests that there exist problem sets for which humans provide better quality decisions and problem sets for which intelligent control algorithms provide better decisions. For example, few would argue that the path planning algorithms provided by Mapquest and Google find superior paths over complex road networks in times unmatched by humans. On the other hand, even the most sophisticated pattern recognition, algorithms are incapable of matching small children in rapidly identifying and manipulating common household items in a cluttered environment. The neuroscience and psychology communities have longstudied human cognitive abilities, and computer science, particularly the subfield of complexity theory (Kolmogorov 1998), has been used to study the performance of cognitive algorithms as a function of the problem space being addressed; however, a comparative understanding between human and machine cognition (or even the tools necessary to achieve this understanding) does not exist at this time. In order to move the discussion of UAV C2 into a manageable space, two simplifying assumptions are asserted: first, decisions should be made using the maximum amount of information, and (51.2) given equivalent information, it is preferred that decisions be made by a human. These simplifying assumptions allow us to focus on the availability of information as the primary driver for command and control. While it is somewhat disconcerting to ignore the quality of the decision-maker and ethical issues, our focus on information as the driving factor in C2 is consistent with the lessons learned from Guderian and Rommel earlier.

51 Command and Control of Autonomous Unmanned Vehicles

1285

Table 51.2 Communication availability experienced by UAVs during a mission can vary greatly. Depending upon the operational conditions, latency and bandwidth can vary greatly Communications Bandwidth (ˇ) availability Description Latency (ı0 ) Dedicated Dedicated infrastructure whose Very low Extremely high communications access is tightly controlled and not contested by environmental or adversarial activities Uncontested Standing infrastructure that is Very low High broadband broadly used that is not contested by environmental or adversarial activities Contested Standing infrastructure that is Low Low communications broadly used and is not contested by environmental or adversarial activities that produce periodic outages and/or reduction in service Over the horizon Operations that involve periodic High Moderate movement in areas that are beyond communications range or involve periodic, planned communications blackouts

Having narrowed our focus into the information used to make a decision, three information-theoretic distinctions using between an operator and an onboard computer to determine a course of action are identified. These distinctions are as follows: (a) the complexity of the scene that must be described to make a decision, which dictates the size of the packets that must be communicated and the run time of decision processes; (b) the entropic drag of the system being represented by the data, which dictates the time for which the information is valid; and (c) the communications delay time provided in Eq. (51.7) which defines the earliest time at which the decision could be made. These values may be combined to form an information value g(x) that defines the UAV control problem. The information value is defined as the product of the complexity of the UAV’s world and the entropic drag of the world’s unpredictable change divided by the delay associated with communicating the world state to the decision-maker s.t: g.x/ D

P .x/.tm / ım .x/

(51.8)

The information value correlates to the utility of the autonomous, automatic, and tele-operated controls approaches. When the world the UAV operates in provides an information value that is high autonomous control dominates, when the world the UAV operates in provides an information value that is low teleoperated control dominates and automatic control is preferred in the midrange. Mapping this relationship to the UAV use cases and communications conditions

1286

D.H. Scheidt

Table 51.3 The appropriate conditions for using tele-operated, automatic, or autonomous UAV control are defined by the operational criteria and available communications Dominant control Dedicated Uncontested Contested technique communications broadband broadband Over the horizon Strategic Tele-operated Tele-operated Automatic Automatic Operational Tele-operated Tele-operated Automatic Autonomous Focused tactical Tele-operated Tele-operated Automatic Autonomous Multiunit tactical Autonomous Autonomous Autonomous Autonomous

defined earlier, operational scenarios that are appropriate for autonomous, automatic, and tele-operated UAV control are defined and enumerated in Table 51.3. As Table 51.3 indicates, there exist real-world circumstances in which UAV C2 should be tele-operated, automatic, or autonomous. Not surprisingly, those times which autonomous C2 is dominant are complex, dynamic situations. Exactly the sort of situation faced by Guderian and Rommel. So, having identified that complex, constantly changing scenarios is best supported by autonomous UAVs, how might we develop such as system.

51.2.3 The Organic Persistent Intelligence Surveillance and Reconnaissance System OPISR autonomous UAVs utilize a software and communications subsystem that is designed to support the rapid, autonomous movement of information across a tactical force. Commander and/or operators interact with OPISR as a system. When using OPISR, war fighters connect into the OPISR “cloud,” task OPISR with mission-level ISR needs and are subsequently provided with the intelligence they need (Fig. 51.4). This capability provides intelligence directly to the war fighter without requiring the war fighter to personally direct or, even know about, the OPISR assets gathering the information. OPISR is autonomous. As a system, OPISR seeks out relevant information, pushing key tactical information directly to impacted soldiers in real time. OPISR is capable of rapidly managing large complex, dynamic situations because it utilizes a decentralized, ad hoc organizational structure. Systems that use decentralized structures such as OPISR are known to be more effective at the timely coordination of complex systems (Scheidt and Schultz 2011). OPISR tracks the location and ISR needs of all blue forces, maintaining a contextual awareness of the war fighter’s current tactical needs. As relevant tactical information becomes available, OPISR presents it directly to the war fighter through an intuitive handheld device. The information requirements that are used to determine information relevance are defined by the war fighter through the same handheld interface. This interface supports abstract queries such as (1) patrol these roads, (2) search this area, (3) provide imagery of a specific location, (4) track all targets of a specific class on a specific route of location, or (5) alert me whenever a threat is identified within a certain distance of my location. Information

51 Command and Control of Autonomous Unmanned Vehicles

1287

Fig. 51.4 OPISR’s concept of operation allows UAVs of various sizes to communicate with each other, with users, and with commanders through an ad hoc, asynchronous cloud. The cloud communicates goals from users to vehicles and sensor observations from users to vehicles

that matches these queries is sent by the system to the handheld device. The handheld interface provides a map of the surrounding area that displays real-time tracks and detections and imagery metadata. The imagery metadata describes, at a glance, the imagery available from the surrounding area. OPISR-enabled vehicles are autonomous; if the information required by the war fighter is not available at the time the query is made, OPISR unmanned vehicles autonomously relocate so that their sensors can obtain the required information. OPISR-enabled unmanned vehicles support multiple war fighters simultaneously, with vehicles self-organizing to define joint courses of action that satisfy the information requirements of all war fighters. Because war fighters are required to operate in harsh, failure-prone conditions, OPISR was designed to be extremely robust and fault tolerant. OPISR’s designers viewed communications opportunistically, designing the system to take advantage of communications channels when available but making sure to avoid any/all dependencies on continual high quality of service communications. Accordingly, all OPISR devices are capable of operating independently as standalone systems or as ad hoc coalitions of devices. When an OPISR device is capable of communicating with other devices, it will exchange information through networked communications and thereby improve the effectiveness of the system as a whole. However, if communications are unavailable each device will continue to perform previously

1288

D.H. Scheidt

Fig. 51.5 OPISR’s hardware architecture is based on a modular payload that can be fitted onto different types of unmanned vehicles

identified tasks. When multiple devices are operating in the same area, they will self-organize to efficiently perform whatever tasks have war fighters have requested.

51.2.3.1 OPISR Hardware Off-the-shelf unmanned vehicles and unattended sensors can be incorporated into the OPISR system by adding the OPISR payload. As shown in Fig. 51.5, the OPISR payload consists of three hardware components: an OPISR processor that executes the OPISR software, an OPISR radio that provides communications to other OPISR nodes including OPISR’s handheld interface devices, and an analog to digital converter that is used to convert payload sensor signals into digital form. Unmanned vehicles that have an onboard autopilot capable of providing stable flight can be modified to become autonomous vehicles by connecting the autopilot to the OPISR processor. When the vehicle is operating autonomously, the autopilot sends guidance and control (GNC) telemetry to the OPISR processor. The processor using the GNC data to devise a continual stream of waypoints are sent to the autopilot to follow. The OPISR processor also uses the GNC telemetry to produce metadata that is associated with the sensor data. The combined sensor data and metadata is then used by the OPISR system as a whole. In service unmanned vehicles frequently use separate communications channels for control and imagery. Since OPISR devices perform both image processing and control onboard the device, these communications channels, and the traditional pilot-ground station, are no longer required. Effectively OPISR devices are capable of operating fully independent of direct human supervision. Note that OPISR devices are still responding to war fighter requests; however, these devices accomplish this without requiring continual communications with the war fighter

51 Command and Control of Autonomous Unmanned Vehicles

1289

being serviced. While OPISR does not require traditional control and payload communications, OPISR devices do support these legacy capabilities. Because the OPISR nodes communicate over a separate channel, OPISR functionality may be provided in tandem with traditional control. This is in keeping with the OPISR dictum that OPISR is an entirely additive capability; unmanned vehicle owners lose no functionality by adding OPISR. However, OPISR vehicles are responsive to commands from human operators and will, at any time, allow an authorized human operator to override OPISR processor decisions. Likewise, legacy consumers of information will still receive their analog data streams. Note that even when the OPISR processor is denied control by the UAV pilot, the OPISR system will continue to share information directly with edge war fighters as appropriate.

51.2.3.2 OPISR Software OPISR is based upon a distributed multi-agent software architecture. Each software agent serves as a proxy for the device on which it is located, and all devices within OPISR have their own agents including unmanned vehicles, unattended sensors, and user interfaces. Each agent is composed of four major software components (Fig. 51.6): a distributed blackboard, which serves as a repository for the shared situational awareness within the agent system; an agent communications manager, which manages the flow of information between agents; a cSwarm controller, which determines a course of action for those devices that are capable of autonomous movement; and a payload manager, which manages the sensor information from the device’s organic sensors. All devices within the system, including the war fighter’s handheld device, are peers within OPISR. 51.2.3.3 Distributed Blackboard In the 1980s, Nii (1986) described a method for multi-agent systems to communicate between each other in an asynchronous manner called a blackboard system. Like its namesake in the physical world, blackboard systems allow agents to post messages for peer agent consumption at an indeterminate time. Each OPISR agent contains a personal blackboard system that maintains a model of the agent’s environment. Three types of information are stored on each agent’s blackboard: beliefs, metadata, and raw data. Raw data is unprocessed sensor data from a sensor within the OPISR system. Metadata is information that provides context to a set of raw data including sensor position, pose, and time of collection. Beliefs are abstract “facts” about the current situation. Beliefs include geo-spatial artifacts such as targets, blue force locations, or search areas. Beliefs can be developed autonomously from onboard pattern recognition software and data fusion algorithms or asserted by humans. Mission-level objectives, the goals that drive OPISR, are a special class of belief that must be produced by a human. The storing and retrieval of information to and from agent blackboards is performed by the blackboard manager. The blackboard manager accepts stores and retrievals from sensors onboard the agent’s device, other agents, or pattern recognition/data fusion software contained within the agent. The integrity of the data stored on the blackboard is maintained by a truth maintenance system (TMS). The TMS performs two functions. First, the TMS resolves conflicts between beliefs. The simplest form of conflict resolution is

Fig. 51.6 Each OPISR node contains a four major software components that manages information flow and decision-making

1290 D.H. Scheidt

51 Command and Control of Autonomous Unmanned Vehicles

1291

accomplished by storing the belief with the more recent time stamp. For example, one belief might posit that there is a target at grid [x, y] at time t0 , and a second belief might posit that there is no target at grid [x, y] at time t1 . More sophisticated conflict resolution algorithms are scheduled to be integrated into OPISR in 2012. The second TMS function is the efficient storage of information within the blackboard. When performing this task, the TMS caches the most relevant timely information for rapid access, and when long-lived systems generate more data than can be managed within the system, the TMS removes less important information from the blackboard. For caching and removal, the importance of information is defined by the age, proximity, uniqueness, and operational relevance. Coordination between agents is asynchronous, unscheduled, and completely decentralized, as it has to be, for any centralized arbiter, or scheduled communications introduce dependencies that reduce the robustness and fault tolerance that is paramount in the OPISR design. Because agent communication is asynchronous and unscheduled, there is no guarantee that any two agents will have matching beliefs at an instance of time. Fortunately, the control algorithms used by OPISR are robust to belief inconsistencies. Cross agent truth maintenance is designed to the same criteria as agent–agent communications: Information exchanges between agents seek to maximize the consistency of the most important information but does not require absolute consistency between agent belief systems. Information exchange between agents is performed by the agent communications manager. When communications are established between agents, the respective agent communications managers (ACM) facilitate an exchange of information between their respective blackboards. When limited bandwidth and/or brief exchanges limit the amount of information exchanged between agents, each ACM uses an interface control component to prioritize the information to be transmitted. Information is transmitted in priority order with priority being determined by information class (beliefs being the most important, followed by metadata), goal association (e.g., if a war fighter has requested specific information that information is given priority), timeliness, and uniqueness.

51.2.3.4 Autonomous Control OPISR’s autonomous unmanned vehicles use dynamic co-fields (DCF), also known as stigmergic potential fields, to generate movement and control actions. DCF is a form of potential field control. Potential field control techniques generate movement or trigger actions by associating an artificial field function with geo-spatial objects. In OPISR, the objects that are used to derive fields are beliefs. Fields represent some combination of attraction and/or repulsion. By evaluating the fields for all known beliefs at a vehicle’s current location, a gradient vector is produced. This gradient vector is then used to dictate a movement decision. Developed in 2003 (Scheidt et al. 2005), DCF extends an earlier potential field approach called co-fields (Mamei et al. 2002) by making the potential fields used dynamic with respect to time and also making vehicle fields self-referential. Self-referential fields are fields that induce vehicle decisions that are generated by the vehicle’s own presence. Adding these dynamic qualities is key to managing two well-known problems with potential fields

1292

D.H. Scheidt

approaches: namely, the tendency of vehicles to become stuck in local minima and the propensity to exhibit undesired oscillatory behavior. As implemented in OPISR, DCF is used to effect specific behaviors such as search, transit, or track, as well as behavioral selection. The DCF algorithm is encoded in the cSwarm software module. All unmanned vehicles in OPISR execute cSwarm. DCF behaviors specific to unique classes of vehicle are produced by tailoring the field formula which is stored in a database within cSwarm. OPISR autonomous unmanned vehicles is a variety of behaviors including: • Searching contiguous areas defined by war fighters. • Searching linear networks such as roads. • Transiting to a waypoint. • Blue-force over-watch. • Target tracking. • Perimeter patrol. • Information exchange infrastructure, in which unmanned vehicles maneuver to form a network connection between an information source, such as an unattended sensor, and war fighters that require information on the source. Note that the war fighter is not required to specify this behavior; the war fighter need only specify the information need, and the vehicle(s) utilizes this behavior as a means to satisfy the need. • Active diagnosis, in which vehicles reduce uncertain or incomplete observations through their organic sensing capabilities. For example, a UAV with a sensing capability capable of classifying targets will automatically move to and classify unclassified targets being tracked by a cooperating radar. In addition to the mission-level behaviors enumerated above, OPISR vehicles exhibit certain attributes within all behaviors. These universal attributes are: • Avoiding obstacles or user-defined out-of-bounds areas. • Responding to direct human commands. OPISR unmanned vehicles are designed to function autonomously in response to mission-level objectives; however, when operators provide explicit flight instructions, OPISR vehicles always respond to the human commands in preference to the autonomous commands.

51.2.3.5 Experimentation The current OPISR system is the culmination of a decade-long exploration in autonomous unmanned vehicles. Experimentation with DCF began in 2002 as part of an effort to investigate agent-based control of unmanned vehicles to support the U.S. Army’s Future Combat System. These early efforts focused predominantly on cooperative search, the results of which are described by Chalmers (Scheidt et al. 2004). Since 2002 thousands of simulated engagements have been conducted with DCF. These simulations have shown that DCF’s computational load is independent of the number of vehicles cooperating to solve a mission. Two-hundred vehicle realtime simulations have been run on a single-Pentium class processor. Simulations from a variety of ISR missions have repeatedly shown that vehicle behavior is robust to perturbations in the number of vehicles or the lay-down of those vehicles.

51 Command and Control of Autonomous Unmanned Vehicles

1293

Hardware in-the-loop experimentation with DCF began in 2003 under a joint effort between the Johns Hopkins University, the Army Research Laboratory and Altarum, Inc. In the summer of 2004, this effort conducted a series unmanned air and unmanned ground vehicles experiments at the Aberdeen Proving Grounds (APG). Low-level control was provided by MicroPilot autopilots (air vehicles) and iRobot Mobility (ground vehicles). High-level control of these vehicles was provided by DCF and Altarum’s pheromone-based swarming algorithm (Parunak 1997). The distributed blackboard system was used to facilitate sharing between the ground vehicles, although centralized data sharing was used to support UAV control. High-level ground vehicle control software was located onboard the vehicles while high-level air vehicle control software was located onboard a ground station that communicated with the onboard autopilot over a 900 MHz communications link. The air vehicles used in these experiments were Army Mig-27 target drones (Fig. 51.7). These drones have a 6-foot wing span and are capable of air speeds of 60 knots. The ground vehicles used were iRobot ATRV, ATRV-JR, and mini robots. This effort concluded in an October 2004 demonstration in which two air vehicles and four ground vehicles conducted a multi-objective mission at APG. The air vehicles were equipped with GPS for localization and notional EO sensors for target detection and tracking. Mission objectives were provided by three independent human users. Objectives included are as follows: (1) patrol the base, (2) protect the moving convoy, (3) search operator-defined areas, (4) track unclassified targets, (5) classify unclassified targets, and (6) interdict targets classified as threats. AUVs supported objectives 2, 3, and 4. The demonstration started with UGVs patrolling the base and UAVs searching the engagement area. A convoy entered the engagement area with the intention of transiting to the base. The officer leading the convoy requested protection, causing the UAVs to change mode to cover the convoy en route. While patrolling the area infront of the convoy, the UAVs detected a number of dismounts loitering at a road intersection in the convoy’s path. This information was relayed to the convoy, causing the human driver to stop prior to the intersection. This same information was relayed to the UGVs, causing the UGV with the acoustic sensor payload to approach the intersection. Once in range of the intersection, the acoustic UGV classified a subset of the dismounts as hostile targets. This information caused the UGVs to pursue the hostile targets, which then fled the area. Once the intersection was cleared, the convoy completed its transit to the base. This demonstration, and variants of it, was successfully performed a number of times. During one demonstration a UGV suffered a hardware fault and went off-line. The other vehicles recognized the sudden absence of their peer, adapting their actions to maintain overall operational effectiveness. A similar set of experiments were conducted at the Department of Energy’s Nevada Test Site in the summer of 2005. These experiments used three Procerus Unicorn UAVs and infrared unmanned ground sensors (UGS). In these experiments all of the vehicles used DCF as their high-level control policy. The UAVs performed area search, road search, target tracking, and, for the first time, airborne information exchange. The engagement consisted of two mounted blue-force patrols,

1294

D.H. Scheidt

Fig. 51.7 The first UAVs to fly using OPISR’s DCF were modified U.S. Army Mig-27 drones

a single-dismounted aggressor, and two-mounted aggressors. The UGS, using onboard automated target recognition algorithms, detected and identified the dismounted and mounted adversaries. These detections caused the UAVs to track the adversaries and relay the contact information to the blue-force patrols. These experiments were the first in-flight uses of the previously described active-metadata framework. Also in 2005, DCF was successfully used in experiments to control UAVs detecting, taking samples from and tracking atmospheric plumes. These experiments used three AeroVironment Dragon-eye UAVs. The plume detection experiments were motivated by a desire to provide first-responders with an ability to rapidly identify aerosol contaminants emanating from an industrial accident and to understand and predict the location of those contaminants. These experiments were conducted at a Department of Homeland Security facility in Michigan. From 2006 to February 2008, DCF has been used regularly in the Naval Postgraduate School’s TNT experiments held at Camp Roberts, CA. These experiments have deployed as many as six Procerus Unicorns a fully mission-based swarm. The UAVs used for TNT are fully independent, as the high-level control, a variety of sensors, and the automated target recognition algorithms have been moved onboard the UAVs. Additional behaviors have also been demonstrated at TNT including automated obstacle avoidance (fixed and airborne) and communications. In a 2007 experiment six vehicles cooperated to provide streaming video of arbitrarily defined objects from over the horizon to human users. DCF has been used as a control metaphor in sea-based experiments conducted predominantly in littoral environments. Starting in 2005 DCF has been used to control several types of unmanned sea surface vehicles conducting search and track n’trail missions at speeds of up to 40 knots with no man in the loop. In 2007 and 2008 OPISR software was used to control unmanned undersea vehicles on search and track missions. Between 2002 and 2010 twenty hardware experiments were conducted using elements of the OPISR system, including DCF (Scheidt et al. 2005), the distributed blackboard (Nii 1986; Hawthorne et al. 2004), delay-tolerant communications

51 Command and Control of Autonomous Unmanned Vehicles

1295

Fig. 51.8 OPISR vehicles from the 2011 Webster field demonstration including one (of four) ScanEagles, two custom surface vehicles, one (of six) Procerus Unicorns, and an OceanServer Iver2 undersea vehicle

(Bamberger et al. 2004), and simultaneous support for multiple end users (Stipes et al. 2007). As successful as these experiments have been prior to 2011, the full suite of OPISR capabilities described in this chapter had not been demonstrated on a large disparate set of vehicles. In September 2011 a multi-vehicle system consisting of 16 OPISR-enabled nodes, including 9 UAVs in support of 3 users, was conducted at Webster Air Field in St. Inigoes, MD, and the surrounding Chesapeake Bay. The 2011 demonstration mixed air, ground, and sea ISR needs with surveillance being conducted under the water, on the water, and on and over land. The autonomous unmanned vehicles included four Boeing ScanEagles, six Procerus Unicorns, a Segway RMP ground vehicles, custom surface vehicles, and an OceanServer Iver2 undersea vehicle. The air, surface, and undersea vehicles are shown in Fig. 51.8. These vehicles used a wide range of payload sensors to detect, classify, and track waterborne vehicles, land vehicles, dismounts, and minelike objects, including EO, IR, radar, AIS, passive acoustic, side-scan sonar, and LIDAR. ISR tasking was generated by three proxy operators, two of which were on land (one mounted and one dismounted) and one of which was on the water. ISR tasks requested required the use of all of the vehicle behaviors previously described. The OPISR capabilities demonstrated by OPISR UAVs at St. Inigoes included a set of six enabling autonomous submission capabilities, which are:

1296

D.H. Scheidt

1. Area search – When user(s) requests that one or more contiguous areas should be searched, the autonomous vehicles respond by searching those areas. 2. Road network search – When user(s) request that one or more roads should be searched, the autonomous vehicles respond by searching those roads. This is functionally identical to performing an area search over an area that is confined to (or focused on) roads of interest. 3. Overwatch – A convoy protection mode that when a user requests “overwatch” protection, one or more vehicles circle the protected user. This is functionally identical to track a moving target, except that the U  V sensors will be directed at the region surrounding the convoy, rather than directly at a target. 4. Communication relay – When an area to be searched (see behavior #1) is farther from the user interested in the area than the vehicle-user communications range, one or more vehicles autonomously form a communications chain to relay data from the searched area to the user. 5. Obstacle avoidance – Obstacles that are known by a vehicle are avoided. Note that obstacles may include physical obstacles detected by vehicle sensors (e.g., trees) and/or abstract obstacles provided by users such as no-fly zones. 6. Behavior switching – Vehicles are capable of exhibiting multiple behaviors depending upon current user goals and circumstances. One highlight of the experiment was the indirect access to time-sensitive data from a remote camera sensor – located well outside the direct communication ranges of the C2 ground stations and their radios – to OPISR users by autonomous UAV communications chains. This communications link was formed fully autonomously, even to the extent that there was no specific command given to the ScanEagle UAV that it forms a communications chain. The ScanEagle was tasked to patrol the region, and as it became aware of sensor data, it relayed that data to the user, who immediately had the sensor data available on his display. Another key OPISR feature that was demonstrated in St. Inigoes was the OPISR UAV ability to coordinate on tasks that it (the UAV) cannot satisfy without recruiting other vehicle types. Very importantly in this experiment, command was shown not only from one ground station to multiple heterogeneous vehicle platforms but also commands from multiple users at multiple ground stations could set the goals of any and all OPISR components. During experimentation sensor and mission information was delivered to both C2 user stations both directly and via UAV communications chaining. This delivery occurred both automatically and in response to specific user requests. All relevant mission data was made available and reported to the C2 user, including the status of assigned search missions, represented as “fog of war” over the map geo display; detections from the various UGS, the two USVs, and the UUV; and camera imagery from the various platforms. Within the flight time windows available, over 65,000 images were collected by the various camera sensors, including onboard UAV payload sensors and relayed to the C2 stations, available to any of the notional war fighter users as their ground node was connected to cognizant portions of the mesh network.

51 Command and Control of Autonomous Unmanned Vehicles

51.3

1297

Conclusion

The OPISR system is a framework that provides a capability through which numerous unmanned platforms simultaneously provide real-time actionable intelligence to tactical units, provide abstract manageable situation awareness to theater commanders, and provide high-quality forensic data to analysts. OPISR is a demonstrated system that includes a distributed self-localizing camera payload that provides imagery and positional metadata necessary to stitch information from multiple sources, a distributed collaboration system that is based upon robust ad hoc wireless communications and agent-based data management, and a user interface that allows users to receive real-time stitched imagery from unmanned vehicles that does not require users to directly control (or even expressly be aware of) the unmanned vehicles producing the imagery. OPISR is a bold vision that presents an innovative approach to ISR, an important enabler emphasized in the Quadrennial Defense Review and other key policy documents, and gives the laboratory an enhanced ability to help sponsors address future capability gaps in this critical area.

References R.C. Arkin, Governing Lethal Behavior in Autonomous Robots (CRC, Boca Raton, 2009) R. Bamberger, R.C. Hawthorne, O. Farrag, A communications architecture for a swarm of small unmanned, autonomous air vehicles, in AUVSI’s Unmanned Systems North America Symposium, Anaheim, 3 Aug 2004 B. Bethke, M. Valenti, J. How, Experimental demonstration of UAV task assignment with integrated health monitoring. IEEE Robot. Autom. Mag., Mar 2008 Defense Systems Staff (2012) Army readies on-demand imagery tool for battlefield use. Defense Systems, 1 June R.C. Hawthorne, T. Neighoff, D. Patrone, D. Scheidt, Dynamic world modeling in a swarm of heterogeneous autonomous vehicle, in AUVSI Unmanned System North America, Aug 2004 A.N. Kolmogorov, On tables of random numbers. Theor. Comput. Sci. 207(2), 387–395 (1998) V. Kumar, N. Michael, Opportunities and challenges with autonomous micro aerial vehicles. Int. J. Robot. Res. 31(11), 1279–1291 (2012) H. Kwon, D. Pack, Cooperative target localization by multiple unmanned aircraft systems using sensor fusion quality. Optim. Lett. Spl. Issue Dyn. Inf. Syst. (Springer-Verlag) (2011) M. Mamei, F. Zambonelli, L. Leonardi, Co-fields: a unifying approach to swarm intelligence, in 3rd International Workshop on Engineering Societies in the Agents’ World, Madrid (E), LNAI, Sept 2002 H. Nii, Blackboard systems. AI Mag 7(2), 38–53 (1986); 3, 82–106 V.D. Parunak, ‘Go to the Ant’: engineering principles from natural multi-agent systems. Ann Oper Res 76, 69–101 (1997) D. Scheidt, M. Pekala, The impact of entropic drag on command and control, in Proceedings of 12th International Command and Control Research and Technology Symposium (ICCRTS), Newport, 19–21 June 2007 D. Scheidt, K. Schultz, On optimizing command and control, in International Command and Control Research Technology Symposium, Quebec City, June 2011 D. Scheidt, T. Neighoff, R. Bamberger, R. Chalmers, Cooperating unmanned vehicles, in AIAA 3rd “Unmanned Unlimited” Technical Conference, Chicago, 20 Sept 2004

1298

D.H. Scheidt

D. Scheidt, T. Neighoff, J. Stipes, Cooperating unmanned vehicles, in IEEE International Conference on Networking, Sensing and Control, Tuscon, 19–22 Mar 2005 C.E. Shannon, A mathematical theory of communications. Bell Syst. Tech. J. 27, 379–423, 623–656 (1948) J. Stipes, D. Scheidt, R.C. Hawthorne, Cooperating unmanned vehicles, in International Conference on Robotics and Automation, Rome, 10 Apr 2007 J. Tisdale, Z. Kim, K. Hedrick, An autonomous system for cooperative search and localization using unmanned vehicles, in Proceedings of the AIAA Guidance, Navigation and Control Conference, Honolulu, Aug 2008 J.A. Wheeler, Information, physics, quantum: the search for links, complexity, entropy and the physics of information, in A Proceedings Volume in the Sante Fe Institute Studies in the Sciences of Complexity, ed. by W.H. Zurek (Westview Press, 1990)

Section XI MAVs and Bio-Inspired UAVs Robert C. Michelson

52

MAVs and Bio-inspired UAVs: Introduction Kimon P. Valavanis and George J. Vachtsevanos

MAVs and Bio-inspired UAVs addresses the emerging UAV area of Micro Aerial Vehicles (MAVs) and bio-inspired MAVs and UAVs. Research and development activities in MAVs have accelerated significantly over the past years, driven by a need for small autonomous vehicles that can execute a variety of tasks, such as Intelligence, Surveillance, and Reconnaissance (ISR) in complex urban environments, search and rescue operations, and security and border patrol, among other applications. A substantial component of the MAV research is inspired by the exceptional flying behaviors of biological species, i.e., birds and insects. MAVs can be operated by a single person offering low weight and cost, extreme maneuvering capabilities, and rapid response times to requests for visual observations. Innovative MAV concepts are motivating advanced research and development sponsored by the government and industry. Micro Air Vehicles by R.C. Michelson addresses challenges related to the design of MAVs. Such challenges span across aerospace, electrical, mechanical, and computer engineering because of flight regime in which these tiny aircrafts operate. Aerospace designers must contend with issues surrounding low Reynolds number flight, while electrical and mechanical designers are concerned with issues of energy storage, behavior of materials at small scales, and non-scaling items. The missions at which MAVs excel demand increased levels of autonomy, forcing computer engineers to create innate onboard intelligence exhibiting high bandwidth and

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 142, © Springer Science+Business Media Dordrecht 2015

1301

1302

K.P. Valavanis and G.J. Vachtsevanos

superior abilities to interpret obstacle-rich environments not usually encountered by larger flying machines. Survey of the Human-Centered Approach to Micro Air Vehicles by S. Michelson presents a detailed overview of some of the Human Systems Integration (HSI) and Human Factors Engineering (HFE) issues involved with MAVs. The importance of a total systems engineering approach to MAV design, how MAVs fit into commonly accepted Human Systems Integration domains, and an exposure of some emerging issues with MAVs that require further research are discussed. The unique attributes of MAVs in terms of their size and control methods, combined with the challenges of the dynamic operational environments where they are deployed, represent HFE issues exclusive to the MAV platform that require special consideration. The importance of designing for the human operator is paramount for successful outcomes with MAV platforms. Specifically highlighted are some areas where currently researched HFE issues are particularly applicable to MAVs as opposed to large-scale systems. Development of Insect-Sized MAVs by Sunada, Liu, Tokutake, and Kubo describes a prototype bio-inspired flapping MAV with flexible wings, with a specific focus on the flexible-wing aerodynamics. The flapping-wing MAV has a weight of 2.4–3.0 g and a wingspan of 10–12 cm, which is comparable to hawk moths and hummingbirds. The MAV’s flexible-wing aerodynamics is analyzed by combining an in-house computational fluid dynamics (CFD) method and wind tunnel experiments (EXP). In addition, fixed-wing and rotary-wing MAVs with elements that enable the miniaturization of an aerial vehicle are introduced. Flapping-Wing Propelled Micro Air Vehicles by Jones and Platzer presents a brief history of the major discoveries in the scientific exploration of flappingwing flight. This is followed by a short review of the basic concepts of lift generation on wings in low-speed, steady flight, which leads into a discussion of the generation of thrust due to the flapping of wings. The aerodynamics of single flapping wings in forward and hovering flight, of flapping tandem and biplane wings, and of dual wings using the clap-and-fling effects are discussed. The chapter concludes with an overview of the major characteristics of five representative flapping-wing propelled MAVs developed to date, including models developed at the AeroVironment Company, Naval Postgraduate School, Wright State University, and Delft University. Inventing a Biologically Inspired, Energy Efficient Micro Aerial Vehicle by Ratti and Vachtsevanos introduces a novel framework for the design and control of a MAV, where the conceptual design is based on biologically inspired principles and emulates a dragonfly. The chapter addresses the design and control features of the proposed design and gives an overview on the developmental efforts towards the prototyping of the flyer. The potential applications for such a high endurance vehicle are numerous, including air deployable mass surveillance in cluster and swarm formations. The disposability of the vehicle helps in battlefield deployment as well, where such an MAV is made available to soldiers for proximity sensing and threat-level assessment. Other applications include search and rescue operations and civilian law enforcement.

52 MAVs and Bio-inspired UAVs: Introduction

1303

Issues Surrounding Communications with Micro Aerial Vehicles by C. Michelson seeks to answer many of the communications-linked questions that MAV designers have and gives a high-level overview of the factors that affect MAV data and control links. Challenges related to communications with MAVs are because of their size and very limited payload capabilities. Limited payload capacity leads to considerable constraints on power sources, sensors, and communication systems. Power sources are by far the most weight-inefficient components on an MAV. MAV designers are forced to look elsewhere to optimize their designs. The best way to do so in lieu of focusing on improving battery technology is to optimize the systems that draw power, thereby increasing endurance. Motors, onboard processing, and communications transceivers are the largest three power consumers on MAVs today. While motors and embedded processing are important to optimize, the sheer number of available communications options may leave MAV designers unsure how to proceed. By building an MAV around its onboard communications system, designers increase reliability, endurance, and capability with little or no added cost. Care must be taken to ensure that the end result meets the power, aerodynamic, and electromagnetic requirements for the particular MAV and its particular mission. Collectively, the chapters in this section refer to all aspects of MAV design, control, communications, operation, and applications.

Micro Air Vehicles

53

Robert C. Michelson

Contents 53.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1306 53.2 Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1306 53.3 The Energy Barrier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1307 53.4 Biological Inspiration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1308 53.5 Mission Effectiveness and Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1309 53.6 The Man and the Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1309 53.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1309 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1310

Abstract

The design of micro air vehicles (MAV) presents one of the most formidable engineering challenges, not only to aerospace, but electrical, mechanical, and computer engineers because of flight regime in which these tiny aircraft operate. Aerospace designers must contend with issues surrounding low Reynolds number flight, while electrical and mechanical designers will be concerned with issues of energy storage, behavior of materials at small scales, and non-scaling items. The missions at which MAVs will excel demand increased levels of autonomy, forcing computer engineers to create innate onboard intelligence exhibiting high bandwidth and superior abilities to interpret obstacle-rich environments not usually encountered by larger flying machines. This section deals with MAVs that conform to the original Defense Advanced Research Projects Agency (DARPA) definition (15 cm and smaller), rather than the larger UAVs which many are prone to label as “micro air vehicles,” but in reality are just small-scale UAVs. Topics covered in this section are somewhat unique relative to typical MAV discussions

R.C. Michelson Aerospace, Transportation, and Advanced Systems Laboratory, Georgia Tech Research Institute, Smyrna, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 104, © Springer Science+Business Media Dordrecht 2015

1305

1306

R.C. Michelson

in other texts dealing with MAVs that focus primarily on the air vehicle rather than the system. Not only is the air vehicle covered here but also issues of MAV deployment which are all too often neglected in discussions of MAV operations (e.g., communications and operational issues).

53.1

Introduction

“No other air vehicle design space has presented the mix of challenges as that of miniature flight platforms. By definition these tiny platforms are unmanned and endeavor to invade the flight regime of birds and insects. In order to do so, the creators of these aerial robots must address the same physical design constraints which have already been mastered by the world of airborne biology, including low Reynolds number aerodynamics, high energy density, and extreme miniaturization” (Michelson 2004). Critical issues arise when working at this scale, energy storage being one of the most significant. Nontraditional aerodynamics in low Reynolds number regimes complicate flight efficiency, driving designs toward the biologically inspired and away from the conventional. Due to the inability to scale all systems as a result of basic physics-based constraints, normally simple telemetry systems are constrained to operate at short ranges and high frequencies due to limitations on antenna structures. Further, teleoperated operation becomes impractical as the tiny MAV fades from view over fairly short distances. This leads to designs requiring greater or full autonomy – but an entirely new set of issues arise with fully autonomous MAVs and their potential for emergent behaviors that are difficult to predict and test. The following sections of this introductory MAV chapter will put these issues into perspective and show their interrelation.

53.2

Historical Perspective

Interest in tiny flying machines had its origins with the notion that a small insect-like flying platform could be devised for covert operations. The U.S. Central Intelligence Agency (CIA) experimented with the creation of a remotely-controlled pneumatic “dragon fly.” In 1993 a RAND Corporation study discussed the potential use of sensor-carrying insects and the concept of truly “micro” air vehicles (Hundley and Gritton 1994). Japanese researchers attempted to stimulate motor neurons in cockroaches to control the insect trajectory by a radio link (BBCi Sci/Tech 1998). Tiny airplanes powered by biological motors (large flies glued to a fixedwing fuselage) have been demonstrated recently but was actually a subject of experimentation at the turn of the twentieth century by Nikola Tesla (Cheney 1981). “Micro Air Vehicle (MAV) is a most unfortunate name given to this class of air vehicles because none are truly ‘micro’ and the original (ca. 1995) official DARPA vehicle definition requiring a maximum 15 cm (6 in.) dimension confirmed the name to be a total misnomer” (Michelson 2006). DARPA defined MAVs as “being less

53 Micro Air Vehicles

1307

than 15 cm” because this represented the juncture at which low Reynolds number effects begin to dominate and beyond which, integration of energy, propulsion, aerodynamic structures, and intelligence is a necessity (Michelson 2004). Initially (ca. 1997), DARPA’s vision for MAVs was that the individual soldiers at the platoon, company, or brigade level would use such vehicles for reconnaissance and surveillance, battle damage assessment, targeting, emplacing sensors, communications relays, or for sensing chemical, nuclear, or biological substances. The 15-cm vehicles would be able to conduct real-time imaging, have ranges of up to 10 km, and speeds of up to 30 miles per hour for missions that are 20 min– 2 h long. By 2012, this vision has been realized, but not without many operational impediments.

53.3

The Energy Barrier

In order to realize the endurance envisioned by DARPA, energy sources of significant energy density had to be developed. Various MAV concepts using high energy density battery technology, fuels (both combustion and catalyst-based), as well as energy harvesting techniques have been explored. Battery technologies exist that exhibit amazing energy densities; however, most of these are primary cells (nonrechargeable) and because of the greater danger of explosion from rapid overheating due to internal or external shorting of the cells, these battery types have not been fielded in MAVs. Presently, the most common high energy density battery chemistry in common use is the lithium polymer (LiPo) battery. LiPo technology allows for the creation of relatively high energy density secondary (rechargeable) cells and is the most widely used energy source for electric flight propulsion at present. Electric propulsion is almost exclusively by means of electric motors driving propellers/rotors or flapping wing mechanisms. Synthetic muscle technologies such as piezoelectric elements, shape-memory alloys, rheological fluids, and electropolymers offer an array of electrically controlled actuators which have been touted as a potential source of power for biologically inspired aerial robots, but all have yet to demonstrate sufficient efficiency to lift even the best high energy density battery (Michelson 2004). Researchers have also considered the use of chemical power sources due to the higher energy density contained in fuels of various types. Stored energy becomes a significant impediment as MAV mission duration increases. The present state of the art in battery technology does not allow for long endurance MAV missions, though it is hoped that someday improved electrical storage media (carbon-air, fuel cells, etc.) will result in the energy densities required for useful long endurance missions in MAV-sized vehicles. Near-term solutions to onboard energy storage will come from chemical or fossil fuels because of their superior energy density. As a point of comparison, consider the amount of releasable energy stored in a drop of gasoline compared to that which can be stored in a battery the size of a drop of gasoline (Michelson 1998). Further, as the mission progresses, a fueled propulsion system

1308

R.C. Michelson

gets lighter, whereas a battery-operated propulsion system remains just as heavy at the end of the mission as the beginning. One example of a chemically fueled MAV concept is the Entomopter, a flapping wing MAV that was under development by DARPA and NASA. The “Entomopter” (entomo as in entomology + pteron meaning wing or a “winged insect machine”) uses a liquid monopropellant such as highconcentration hydrogen peroxide or hydrazine in the presence of a catalyst to create gas for reciprocating actuation of flapping wings. A third approach to powering MAVs for longer endurance missions is energy harvesting. Small birds and insects are consumed with the task of energy harvesting: the search for food. Hummingbirds, the smallest of all avians, feeding on dilute nectar can ingest nearly three times their body mass in nectar per day to sustain life and mobility (Powers and Nagy 1988; McWhorther and Martinez del Rio 1999). Their small bodies cannot carry large amounts of food, so to improve efficiency they choose high energy foods that provide immediate energy access (sugars) as do many insects. Tiny aerial robots suffer from the same need for readily available energy. The energy density of the best battery technologies currently available still cannot match that which is locked chemically in various compounds such as sugars. Some have advocated energy harvesting through the use of solar panels on MAVs. Unfortunately, the efficiency of current solar cells (roughly 5 % for common cells, ranging up to 28 % for some of the best triple-junction gallium arsenide spacequalified cells) in sizes that could be carried by a MAV is insufficient for sustained fight (Dornheim 2003). The extra weight of such cells negates their use as an endurance extender, and their low voltage output is incompatible with many of the electronic actuators proposed (e.g., piezoelectric, electro polymers). Finally, night operation or fight through shadows is precluded. Research has been conducted in leaching power from power lines by UAVs, and a ground vehicle known as “slugbot” was developed to consume organic matter and digest it to produce energy for locomotion. In both cases, the amount of energy harvested and the time taken to process the energy made the application impractical.

53.4

Biological Inspiration

MAVs have been demonstrated as fixed, rotary, and flapping wing vehicles. As the size of the MAV decreases and the vehicle operates in lower and lower Reynolds number regimes, the efficiency for flight moves away from classical fixed-wing aerodynamics toward flapping wing solutions. It is no accident that all insects use flapping wings to fly. A number of MAVs such as the Entomopter mentioned above and various other unconventional flappers have been demonstrated. Some of these designs are “biomimetic” (copying nature), while others are “biologically inspired” (taking principals from nature but implementing them in a way that goes beyond what is found in nature). Examples of both approaches are discussed in detail in the subsequent chapters Development of Insect-sized MAVs and Flapping-Wing Propelled Micro Air Vehicles.

53 Micro Air Vehicles

53.5

1309

Mission Effectiveness and Payload

One area that is usually glossed over in most texts dealing with MAVs involves issues surrounding communications with micro air vehicles. Because of their small size, MAVs are restricted not only in the power that they can carry to support telemetry, but they are also hampered by the physical constraints placed on radiating elements. Operation at greater ranges is enhanced by the use of lower frequencies. The same is true when attempting to transmit through obstacles such as buildings or foliage. Unfortunately, lower frequency operation becomes inefficient as the antenna apertures shrink in size. Increasing transmitted power is not an effective solution due to the limited energy carried onboard the MAV as well as the mismatch between a low-frequency transmitter and a miniature antenna. As a result, higher-frequency telemetry is desirable from an emissions standpoint, but not from an attenuation standpoint. Specific consideration must be given to transmitter power, data and command link design, amplifier types, modulation techniques, simultaneous communications, and antenna aerodynamics. Various techniques to deal with these conflicting MAV telemetry design parameters are discussed in chapter Issues Surrounding Communications with Micro Aerial Vehicles.

53.6

The Man and the Machine

Certainly one of the least discussed topics relating to MAVs is the human interaction with the MAV system. A prime example of an MAV-specific human interaction problem is the fact that as the MAV flies away from the operator, even at only short ranges, useful visual feedback is lost due to its tiny size. Further, if the MAV mission involves the penetration of a building, visual feedback is lost. As briefly discussed in the prior section, teleoperation via a data link is problematic for any number of reasons, none the least of which is the cost in onboard stored energy. Another significant issue arises when autonomy is used to relieve the operator from having to control the MAV remotely. A fully autonomous MAV, while very desirable for many reasons, is highly difficult to test because of the unpredictable potential for emergent behaviors (Michelson 2008). Chapter Survey of the Human-Centered Approach to Micro Air Vehicles discusses these issues as well as how to effectively incorporate a human-systems integration approach to the design of MAVs in light of a valid MAV concept of operation. Automation, situation awareness, and workload issues are discussed, as are the test, evaluation, and training considerations particular to MAV operators.

53.7

Conclusion

The genesis of the micro air vehicle concept and the initial funding by DARPA to push the technology forward have resulted in many MAV configurations using

1310

R.C. Michelson

fixed, rotary, and flapping wing propulsion. The critical element preventing greater performance over the past 20 years has been high energy density propulsive power, and this remains true to this day. Battery and chemical fuel power sources, with few exceptions, have not progressed at a rate which allows MAVs to have the endurance of birds of similar size. Biomimetic control systems have been developed in laboratory settings, but the ability to control a MAV with the agility of a small bird, much less than that of a dragonfly, still eludes designers. When considering the best communications architecture for MAVs, designers must remember that everything is a trade-off when balancing miniaturization with high bandwidth remote non-line-of-sight telemetry. RF characteristics are largely bounded by the universal constraints of physics and are therefore not controllable by the designer. Remaining mindful of the distinction between human systems integration and human factors engineering with regard to MAV development is pivotal to correctly accounting for the human in the design process. The uniqueness of the MAV and its deployment is creating the need for a new understanding of man-machine systems in the unique operational environment of the MAV. MAVs are still in their infancy 20 years after the DARPA vision for such a vehicle, largely due to technical challenges. It is not that one does not understand the design issues; rather, one lacks the component efficiency and level of mechanical miniaturization necessary to compete with biological systems. Since fully 50 % of a MAV’s gross takeoff weight is consumed in stored energy, one of the areas in which future MAV designers will reap the greatest gains will come from advances in propulsive energy density.

References BBCi Sci/Tech, Cyberbugs are go. Wednesday, March 4, 1998 – Published at 00:21 GMT, at, http:// news.bbc.co.uk/2/hi/science/nature/61822.stm. Accessed 18 Aug 2012 M. Cheney, Tesla: Man Out of Time (Dorset, New York, 1981), p. 7. ISBN 0-88029-419-1 M. Dornheim, Get me through the night, in Aviation Week & Space Technology (McGraw-Hill, New York, 2003), pp. 66–68 R. Hundley, E. Gritton, Future Technology-Driven Revolutions in Military Operations. Document No. DB-110-ARPA (RAND Corporation, Santa Monica, 1994) T. McWhorther, C. Martinez del Rio, Food ingestion and water turnover in hummingbirds: how much dietary water is absorbed? J. Exp. Biol. 202, 2851–285812 (1999) R. Michelson, Update on flapping wing micro air vehicle research- ongoing work to develop a flapping wing, crawling entomopter, in 13th Bristol International RPV/UAV Systems Conference Proceedings, Bristol, 1998, pp. 30.1–30.12 R. Michelson, Novel approaches to miniature flight platforms. Proc. Inst. Mech. Eng. Part G J. Aerosp. Eng. 218(Special Issue Paper), 363–373 (2004) R. Michelson, Very small flying machines, in 2006 Yearbook of Science & Technology (McGrawHill, New York, 2006), pp. 341–344. ISBN 0-07-146205-8 R. Michelson, Test and evaluation of fully autonomous micro air vehicles. ITEA J. 29(4), 367–374 (2008) D. Powers, K. Nagy, Field metabolic rate and food consumption by free-living Anna’s hummingbirds (Calypte anna). Physiol. Zool. 61, 500–50611 (1988)

Survey of the Human-Centered Approach to Micro Air Vehicles

54

Stuart Michelson

Contents 54.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1312 54.2 Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1313 54.3 Incorporating an HSI-Centric Design Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1314 54.4 MAV Concept of Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1315 54.5 MAVs and the Man-Machine System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1315 54.6 Automation, Situation Awareness, and Workload. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1318 54.7 Test, Evaluation, and Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1321 54.8 MAV Ground Control Station Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1323 54.9 Ruggedization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1325 54.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1326 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1327

Abstract

A detailed overview of some of the Human Systems Integration (HSI) and Human Factors Engineering (HFE) issues involved with the newest and perhaps fastest growing research area in unmanned systems, micro air vehicles (MAVs), will be presented. This work will be useful to those studying MAV system concepts and designs, managers of HSI programs, users of MAV systems, and those who design MAVs and the resources to support them. The importance of a total systems engineering approach to MAV design, how MAVs fit into commonly accepted Human Systems Integration domains, and an exposure of some emerging issues with MAVs that require further research are discussed.

S. Michelson Human Systems Integration Division, Georgia Tech Research Institute, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 90, © Springer Science+Business Media Dordrecht 2015

1311

1312

S. Michelson

The unique attributes of MAVs in terms of their size and control methods, combined with the challenges of the dynamic operational environments where they are deployed (such as the battlefield), represent HFE issues exclusive to the MAV platform that require special consideration. The importance of designing for the human operator is paramount for successful outcomes with MAV platforms. Literature currently addressing HFE issues with unmanned platforms generally lump all flying systems together, making no distinction between the large high-altitude platforms and smaller ones, despite there being a unique set of challenges that are specific to smaller platforms. Specifically highlighted are some areas where currently researched HFE issues are particularly applicable to MAVs as opposed to large-scale systems.

54.1

Introduction

Successfully implementing micro air vehicle (MAV) programs requires one to maintain a total systems approach, and Human Systems Integration (HSI) must be an integral part of that approach. These are presuppositions to the remainder of the content in this work. Acquiring a human-centered approach to analyzing the observable problems with unmanned systems is essential, but one might feel inclined to question what the human has to do with such systems at all, especially when the hallmark of the domain is the absence of humans. While it is important to comprehend that the human is an integral part of the design of any unmanned system, humans are particularly important to the MAV platform where operators are thought to have more active roles in commanding the craft, and systems have a wider spread across the levels of automation. This reality can complicate system design and function allocation. A failed systems approach that deemphasizes the human in the loop will result in human-related challenges such as unmanageable operator workload or poor situational awareness (SA) culminating in compromised safety for warfighters. The Defense Acquisition Guidebook indicates that a total system approach “includes not only the prime mission equipment, but also the people who operate, maintain, and support the system; the training devices; and the operational and support infrastructure” (United States Department of Defense 2012). It would be a grave error for designers to assume that just because a human is not physically present onboard the vehicle that they do not need to be knowledgeable about human attributes and limitations. Testing and evaluating these attributes and limitations is a major aspect of Human Factors Engineering (HFE), and given the uniqueness of the manner in which MAVs are deployed, special design considerations must be given to elements such as human anthropometry and cognitive, physical, and sensory abilities. While not every engineer needs to be an expert in these areas, they should at least be able to discuss them intelligently with HFE/HSI practitioners and understand their extreme importance to successful system design, as it is their job to guarantee that the human is considered throughout every portion of the design process.

54 Survey of the Human-Centered Approach to Micro Air Vehicles

54.2

1313

Terms

Introducing system designers and operators to HSI and HFE considerations associated with MAVs requires one to recognize the finer distinctions between commonly used terms within the systems engineering domain. A brief discussion of terms used in this work is necessary for clarity since many use some of the common terms interchangeably while the author does not. Note that while there are many definitions of Human Systems Integration, the author ascribes to the United States Air Force’s definition of Human Systems Integration which is “the integrated and comprehensive analysis, design, and assessment of requirements, concepts and resources for system manpower, personnel, training, environment, safety, occupational health, habitability, survivability, and human factors engineering, with the aim to reduce total ownership cost, while optimizing total mission performance” (United States Department of Defense 2009). While different branches of the armed forces emphasize varying arrangements of domains (see MANPRINT, for example), the underlying theme is that HSI is a vital element that optimizes system design to enhance system performance while maximizing the abilities and mitigating the limitations of humans. In so doing, the abilities of the warfighter are enhanced and total ownership costs of a system are reduced (Air Force Human Systems Integration Office 2009). It is important to emphasize that HSI from an engineering standpoint not only deals with requirements and concepts but people and equipment. In this way, HSI should be understood to be a broader concept than HFE although many incorrectly use the two terms interchangeably. The major motivation for incorporating HSI into MAV programs is to control total ownership costs while maximizing system effectiveness. Note that Human Factors Engineering is included as one of the domains of HSI. HFE deals more exclusively with the realization of design criteria, psychological elements of human behavior, as well as physical and mental limitations of humans. By understanding human attributes (anthropometry, for example), HFE analysis seeks to enhance performance (to required levels), increase safety, and increase user satisfaction (Wickens et al. 2003). The science of human factors can be said to rest upon generalization and prediction. With respect to unmanned systems, researchers should be interested in generalizing common problems and predicting solutions – all while accounting for the humans in the loop and reducing designinduced failures. The importance of incorporating HSI and HFE analyses into the design process early cannot be stressed enough. The Department of Defense specifies that design changes can cost 1,000 and 10,000 times more in initial production phases than the same change would cost during a product’s earlier design phases (United States Department of Defense 2005). While the human element is a vital consideration throughout the entirety of the design process, it has been shown time and time again that design changes made early are less costly because they do not involve the modification of existing hardware and software. A human-centered approach to MAV development should be early and iterative but also should span throughout the entirety of the design process.

1314

54.3

S. Michelson

Incorporating an HSI-Centric Design Approach

A proper approach to system design must include HSI as a consideration of prime importance. So much so that the Department of Defense’s acquisition policy indicates that HSI should be central to the formation of integrated product teams (IPTs), which are a vital part of the integrated product and process development method. The Defense Acquisition Guidebook states that the product and process development method is a “management technique that integrates all acquisition activities starting with capabilities definition through systems engineering, production, fielding/deployment and operational support in order to optimize the design, manufacturing, business, and supportability processes. At the core of the IPDD technique are IPTs. Human Systems Integration should be a key consideration during the formation of IPTs” (United States Department of Defense 2012). It is important for HSI experts who are considering the human element for each HSI domain to be included as members of the various IPTs so that the human element is considered throughout the entire design process. This design concept places those charged with including the human element in positions to comprehensively impact the system’s design (United States Department of Defense 2012). HSI professionals should be included in the development of a MAV platform’s concept of operations rather than being invited into the design later to comment. In reality this rarely occurs, but doing so allows HSI considerations to impact the concept of operations and in turn create systems where total ownership costs are reduced and system effectiveness is maximized. It is not to be unexpected for engineers to resist the involvement of HSI professionals. Those that do are likely victims of a bad definition of HSI and do not understand how it will positively impact their work, or they think it is an unnecessary cost not realizing that without it their program will be more expensive. Unfortunately, many engineers treat HSI and HFE as checkboxes for after their work is considered complete. A tendency may exist among some engineers to place less emphasis on designing for the human when managing unmanned systems programs such as MAVs. The development of an HSI plan is one way to make sure that human considerations are not deemphasized. Managers must develop an HSI strategy for their program. This takes the form of an HSI plan that should be established early in the design process before any manufacturing takes place (United States Department of Defense 2012). Each phase of the program should include how the HSI plan will address issues associated with each HSI domain. In systems where new technologies are expected to replace or supplement human activities, HSI plans should anticipate the possibility that delays in development might still prevent such measures from being implemented, and therefore, program managers should have alternatives that still account for appropriate levels of human operator workload. While the finer points of the accepted defense acquisition strategy are beyond the scope of this work, a brief understanding of the context in which MAVs are being developed is valuable. Within the unmanned systems category, incorporating good HSI plans is particularly important for MAVs where typically more involvement is expected of the human operator in terms of direct control of the craft.

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1315

This distinction of the MAV platform requires a close look at all of the HSI domains but specifically those that deal with the cognitive, physical, and sensory abilities of humans and their knowledge, skills, abilities, and experience levels as system operators (the personnel, training, and human factors engineering domains).

54.4

MAV Concept of Operations

In contrast to larger, more traditional unmanned systems being widely deployed around the world, MAVs are deployed in a distinctive manner. DARPA’s definition of a MAV specifies that the vehicle be smaller than 15 cm (about 6 in.) in any dimension (McMichael and Francis 1997). The MAV was originally thought to be a military asset that could be included in every warfighter’s loadout, enabling him/her to engage in intelligence, surveillance, and reconnaissance (ISR) missions while in the field (Michelson 2008). This methodology stands in contrast to larger-scale systems with expansive ground control stations used with popular unmanned aerial vehicles (UAVs) such as Northrop Grumman’s Global Hawk or General Atomics’ Predator. In such systems, multiple human operators are involved in the command and control of the system from either homeland-based command centers or mobile trailer style units. From a technical standpoint, MAVs, being craft measuring in mere centimeters, represent a number of issues that have been written about in detail elsewhere. Some of these issues include robustness in inclement weather, endurance, antenna aperture, and strains on command and control due to limited storable energy on board the craft. While these are very real challenges deserving consideration, one cannot look to solve these problems while ignoring or merely paying lip service to the importance of a human-centered approach to design of MAV platforms. The aforementioned issues should be handled considering their effects on humans, but there are other specific issues that are perhaps not often thought of regarding the human’s interaction with MAV systems. The remainder of this chapter is dedicated to addressing some common considerations to MAV design and implementation from a human-centered perspective while comparing and contrasting their uniqueness against larger high-altitude UAV systems. HSI professionals involved in the development of a MAV’s concept of operations should emphasize that a system is driven by humans. Unmanned systems should be just as concerned with designing for the human in the loop as manned systems. The main difference between the two is that a human is not physically onboard the craft, but UAVs such as Predator have hundreds of humans involved in their operation, maintenance, and support.

54.5

MAVs and the Man-Machine System

As the popularity of UAVs has increased, so has the interest in shrinking their size. This reduction in size has expanded the range of applications for which humans can utilize them. No longer are UAVs limited to outdoor flight at altitude.

1316

S. Michelson

Recent developments in research and technology have afforded MAVs the prospect of flying into buildings, down hallways, through air ducts, sewer systems, or even acting as micro spies no bigger than a housefly on the wall. All of these abilities present new challenges not only to the MAV platform’s hardware and software but also to the human system designers and operators. The challenges associated with indoor flight such as obstacle avoidance and navigating environments without the assistance of GPS are difficult to overcome and are only enhanced by difficulties such as transmitting valuable location information back to human operators from inside of a structure. Engineers may feel inclined to meet these challenges with a reductionist approach to systems design. This traditionally popular approach is characterized by a focus on each physical and technical component of a system alone while not emphasizing the behavioral components of systems. This approach has been associated with catastrophic system failures on the order of magnitude of oil spills and commercial aircraft crashes because it does not consider the interaction among all of a system’s parts and does not view system components in terms of their relation to accomplishing overall system goals (Czaja and Nair 2012). The best way to avoid this pitfall is to incorporate a good HSI plan which dictates that all the various domains of HSI are considered together. Since HFE is concerned with enhancing the interactions between humans and all other components of a system, good systems theory sees the understanding of the entirety of a system’s components being in concert as implicit to good systems design (Czaja and Nair 2012). As the levels of autonomy continue to increase with MAVs, the complexities of automation and the system components required to support them also increase. As the emphasis continues to be on advancing the state of MAV technology, equal consideration in the realm of system design must be given to the human element or systems will underperform. This is an accepted tenant within the HFE field (Czaja and Nair 2012). While there are many ways to classify a system, classifying MAV systems in terms of levels of feedback mechanisms is particularly useful to the characterization of the platform. A closed-loop system provides feedback to an operator that is continuous for error correction. This type of system provides human operators with real-time information so that they can determine the difference between actual and desired system states. An open-loop system does not afford an operator the feedback necessary for continuous control (Czaja and Nair 2012). MAVs exhibit levels of autonomy ranging across the spectrum from 100 % remotely piloted robots to entirely autonomous robots making their own decisions in real time based on their environment apart from a human operator. For this reason, MAVs as a category can be closed-loop systems, open-loop systems, or both. This can complicate the design strategy one should employ since open- and closed-loop systems require varying design strategies. Designers should pay close attention to the consequences of dynamic function allocations to ensure that realistic expectations are being placed on human operators in certain circumstances where the human has more control. Extensive experimentation and testing using realistic scenarios that cover the breadth of possible human and computer workload should be undertaken to mitigate potential workload-related errors. For a more full

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1317

discussion of scenario development strategies and workload analysis techniques that extend beyond the scope of this work, see Mental Workload and Situation Awareness (Vidulich and Tsang 2012). When considering the capabilities and limitations of humans for MAV systems, it is important to recognize the way in which advancing vehicle autonomy has changed the nature of man-machine systems. Not only have new technologies enhanced the manner in which human operators can control their vehicles (spoken commands, for instance), but also the capacity for intelligence of the vehicles themselves has changed the relationship between the human and the system. This trend will continue as flying robots are increasingly able to perform more tasks previously restricted to humans. Historically the model for human machine interfaces has been based around control. Humans were said to interact with a system by controlling it. Under such circumstances, the system was subservient to the human. The current state of affairs is that intelligent MAVs are evolving into entities that can extend the capabilities of their human partners (Czaja and Nair 2012). With respect to MAVs, the robot should be viewed and treated as a member of the team working to accomplish the stated goals of the system. This reality of the human’s relationship with MAVs will only be enhanced further as each human operator becomes responsible for swarms of MAVs in the future. Perhaps the most prominent example of MAVs advancing the state of the art in robotic flight behavior is the International Aerial Robotics Competition. It has consistently highlighted the importance of the interdependent relationship of human operators and their aerial robots to accomplish tasks. Starting in 1991, the International Aerial Robotics Competition is the longest ongoing universitybased robotics competition in the world. In addition to exposing the next generation of aerial robotics engineers to the process of system design from concept to reality, the competition has always been driven by pushing the advancement of autonomous robotic technology to levels previously unattainable by governments or industry. While earlier competition challenges involved outdoor flight tasks including mapping hazardous environments, searching for disaster survivors, collecting and moving objects, locating specific buildings and their attributes, and transmitting information back to command and control stations, the most recent challenges have involved smaller flying robots that are tasked with penetrating structures and negotiating confined spaces full of obstacles with no GPS guidance. Small air vehicles are designed to penetrate openings in structures, deactivate security systems, read signs (in languages foreign to the human operators), avoid obstacles (furniture and debris), negotiate disturbances to the airspace (such as fans), and successfully retrieve and replace an object that must be located in a limited amount of time. These advanced robots systems are fully autonomous, making decisions on their own with no human operator control inputs, and are typical of the relationship between humans and their MAVs because they assist the human to accomplish a task that cannot be achieved without both entities working together. An attribute typical of all International Aerial Robotics Competition missions has been using unmanned technology to enhance the capabilities of humans by going places and doing things that are either not possible or too dangerous for humans to attempt. In this way,

1318

S. Michelson

the missions when accomplished become excellent case studies exemplifying the relationships of man-machine systems that can extend the capabilities of humans. For a detailed examination of some considerations that can assist one to create, test, and evaluate autonomous systems that are designed to extend human capabilities and not merely mimic human attributes, the author recommends Ten Challenges for Making Automation a “Team Player” in Joint Human Activity (Klein and Bradshaw 2004). This essay highlights the importance of focusing research objectives on promoting healthy human robot teamwork, not merely on how to make systems more autonomous.

54.6

Automation, Situation Awareness, and Workload

It has been established that human operators and MAVs should be thought of as team members working toward the satisfaction of mission goals. But how does one determine what should be an automated function and what should not with MAV systems? Traditionally, designers might reference a Fitts List (an exhaustive list outlining the abilities of humans against the abilities of machines) to attain an overall understanding of the types of activities humans excel at versus the types of activities at which machines excel. However, one should do so with caution, making sure that by examining individual activities their interrelation as a whole within the system is not overlooked. There are several primary reasons why a function might be considered a good candidate for automation. The first is when a task is considered too dangerous for a human to attempt without resulting in bodily harm or death. Another is when the nature of a task is something at which humans do not excel. A final example is when the safe and successful operation of a system requires a reduction in the workload of the human operator. To achieve optimal system performance, the capacity of the human to take on tasks successfully as well as the system’s limitations must ultimately direct which functions should be allocated to which team member and when (Hopcroft et al. 2006). Recent studies in the area of workload and adaptive automation have revealed that automation should be thought to redistribute workload, not reduce it (Vidulich and Tsang 2012). Too much automation on the other hand could lead to operator boredom, which could separate the human operator from the current state of the system resulting in difficulty for the human if they had to take over control to mitigate unusual crisis events. The concept behind adaptive automation is to implement automation that negotiates the fine balance between appropriate workload levels for human operators (the author advocates 75–80 % workload as a rule of thumb) without inhibiting situational awareness (SA) (Vidulich and Tsang 2012). Particularly when human operators are monitoring MAVs in flight, it is essential that ground control stations provide solid feedback in real time lest the operator lose track of actions taken by the vehicle resulting in incorrect mental models of the system state. Such a situation enhances the likelihood of errors and reduces human

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1319

trust in automation. Conversely, human trust in automation might be too high due to poor feedback resulting in less attentive monitoring and the overlooking of error (Hopcroft et al. 2006). Advocates of adaptive automation prescribe that predetermined fixed (or static) allocations do not compliment complex dynamic systems because individual factors can rapidly and unexpectedly change confounding human operators (Vidulich and Tsang 2012). This is a particularly important observation for MAVs deployed on the battlefield where the operational environment can quickly change not just in terms of threats from enemy combatants but also from environmental changes such as sandstorms or blinding rainfall. Since MAV operators on the battlefield may be involved in direct and supervisory control of aerial robots during a firefight, one should expect that the cognitive workload demands are enhanced greatly as a variety of stressors have the opportunity to overwhelm the operator (Parasuraman et al. 2009). In such cases, automation should be employed effectively to support the human’s performance with the system. Limitations such as mistrust in automation, a poor balance of workload between the human and the system, overreliance, and SA reductions are just a few of the challenges that can prevent automation from successfully aiding the warfighter in such high-stress operational environments (Parasuraman et al. 2009). Dynamic environments such as battlefields are excellent case studies for why static function allocations are not ideal. The solutions to the limitations the warfighter faces in the operational environment with MAVs can be addressed with proper adaptive automation where conditions in which the system intervenes are not fixed but rather avail themselves at the appropriate time depending on the context of the battlefield environment. This sort of context-sensitive system assistance can be triggered by mission events, human operator performance levels, or even the human operator’s physiological status (Parasuraman et al. 2009). For a further discussion of adaptive automation, reference Adaptive Automation for Human Supervision of Multiple Uninhabited Vehicles: Effects on Change Detection, Situation Awareness, and Mental Workload (Parasuraman et al. 2009). It encompasses a more thorough examination of the means of deploying effective adaptive automation with unmanned systems in general for warfighters on the battlefield. To fully understand the relationship between SA and workload for warfighters remotely piloting vehicles on the battlefield, human operators who have sent MAVs inside of structures that they cannot see, or human operators who are piloting MAVs that have left their line of sight outdoors, a brief discussion of what SA entails is warranted. The author’s description of SA offered below is designed to familiarize the reader with the basic concepts of SA, which is a topic that has been widely researched and written about. For a more full discussion of SA and the common challenges associated with it, reference Situation Awareness (Endsley 2012). SA is a human attribute derived from cognitive processes and is therefore not an attribute of computers. Endsley’s widely accepted definition of situational awareness is “the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the

1320

S. Michelson

near future” (Endsley 1995). Using this definition, one must accept that situational awareness has three vital elements (perception, understanding, and prediction) and that each of them must be applied to exact circumstances. Each of the three elements of situational awareness (perception, understanding, and prediction) requires different levels of human operator skill. Perception relies on selective attention, understanding relies on working and long-term memory, and prediction encompasses plan generation based on processed cues. Operator expertise is a major factor in SA (Endsley 2012). Applying the elements of situational awareness to the task of MAV flight monitoring, evaluators should expect to consider the following human tasks: 1. The assimilation of new information into preexisting knowledge bases within the human operator’s working memory 2. The perception of relationships and their significance 3. The projection of what will occur in the future given current information so that appropriate actions can be taken It is being suggested that evaluations studying ground control station effectiveness address each of the three tasks in order by representing information so that it can be comprehended, depicting relationships among stimuli and integrating with heuristics to aid the human in making predictions about future system states. Human cognition naturally attempts to organize and make sense of visual and auditory stimuli. Human cognition is attuned to spatial and temporal relationships. Specifically with regard to perception, the human mind organizes stimuli into sources and streams that have identity and continuity over time (Endsley 2012). Harnessing these tendencies to create appropriate system feedback and data presentations to the human operator is important. In assessing successful user outcomes, it is important to highlight that good SA is not the same as good performance. Good SA is measured as appropriate and timely responses to unexpected events taking place within the system being studied. An appropriate framework that considers the team members within a manmachine system working toward the accomplishment of mission goals must be adopted if one is to effectively manage the issue of automation with varying degrees of SA with a MAV system. One such framework has been proposed, which underscores the importance of considering which team members need awareness of other team members’ states. Using the warfighter as an example, human operators need information about the MAV’s state, humans need to know about each other (other soldiers involved with the mission), the MAV needs to know about the human’s state, and MAVs need to know about each other (Drury and Scott 2008). The human-UAV awareness framework should “accommodate the asymmetrical information needs of people and UAVs, be independent of any particular instantiation of a UAV, and be specific to the types of information needed in the UAV domain” (Drury and Scott 2008). While geared toward UAVs at large, the framework is well suited to MAVs and even refers to specific MAV platforms. The framework does well to address a common pitfall of SA analyses which is to paint an entire system’s human interface as either having good or bad SA because it considers a deeper level analysis of the

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1321

types of awareness various agents exhibit and require within a system in the context of levels of automation (Drury and Scott 2008). As stated earlier, HFE is concerned with generalization and prediction. One challenge worth noting with regard to conducting workload and SA analyses on MAV systems is that generalizing can be challenging. In the example of the warfighter, while one can simulate weather effects to some degree, it is not possible to simulate the stresses of live fire from a battlefield into one’s HFE evaluation protocol when studying a MAV system. Therefore, the actual levels of workload and stress observed during system evaluation cannot be considered to mirror what actual conditions would be. This reality is furthered by individual differences in human operator’s skill and SA needs in a given situation. Another generalization issue associated with MAVs is that one should be careful not to generalize workload assessments conducted on large-scale UAV platforms to MAV platforms because the concept of operations and the manning concepts are different. For instance, researchers at the Georgia Tech Research Institute have conducted workload analyses of large-scale high-altitude UAV platforms to determine the optimal crew size and workload levels under varying conditions for realistic operational scenarios (Georgia Tech Research Institute 2008). It would not be appropriate to generalize their findings about crew size and workload to a MAV because MAV flights are shorter in duration and typically do not involve handoffs of control as one would expect to see with a crew shift in order to support a large system for days at a time. Human operators within the large control centers serving UAVs such as Global Hawk find themselves in a more routine work cycle than the warfighters launching ISR missions with backpack MAVs from dynamic battlefields to see what is over the next hill. Depending on the MAV platform, soldiers may find themselves handing control off to one another to accomplish a mission, but this would not be because of operator fatigue like in large command and control centers. Two elements of workload that are often overlooked are what other routine tasks an operator was already involved in prior to the introduction of a new set of tasks for the MAV system, and how the introduction of the MAV system effects other agents (maintainers for example). Conducting a detailed task/job analysis as part of a good HSI plan should mitigate such an error.

54.7

Test, Evaluation, and Training

A unique attribute of MAV platforms distinguishing them from their larger unmanned counterparts is their low risk in terms of testing and evaluation. The testing of MAV platforms begins in wind tunnels but progresses to outdoor test flights much more rapidly than larger platforms that can ill afford costly crashes (Michelson 2008). AeroVironment’s Wasp MAV with a radius of just 5 nm was selected for Disruptive Technology Fund by the Navy which, among others things, specified a cost goal of 5,000 USD per vehicle (United States Department of Defense 2005). Using the Wasp as an example, testing a vehicle with this price point is a very

1322

S. Michelson

affordable prospect. For perspective, the popular large platforms being deployed today cost well over 10 million USD. At such a nominal cost in comparison, one could crash thousands of Wasps before equaling the cost of one of those systems. The affordability of test flights with MAVs is an advantage to the platform that is uniquely suited to its needs. Assessments of MAV performance during flight tests is generally completely reliant on stored information from testing instrumentation because the vehicle’s responses to control prompts become unobservable to operators on the ground. Fine-tuning stability can often be a trial and error process (Michelson 2008), and therefore the ability to crash early test flights with fewer financial consequences is an advantage of the platform. The often-repeated mantra of “crash early, crash often” can be adopted with MAVs, allowing designers and researchers to conduct a higher number of test flights early on in the design process than with larger-scale unmanned systems. Although there are advantages to testing with the MAV platform in comparison to larger craft, there are drawbacks. Perhaps one of the most challenging drawbacks is how one handles test instrumentation. As mentioned previously, MAVs generally have to rely on test instrumentation onboard to store flight data. It is often difficult to outfit MAVs with this special test instrumentation, and given that the majority of them operate with 50 % of the gross takeoff weight dedicated to propulsion and energy and the other 50 % to airframe and payload (Michelson 2008), some MAVs have very little payload to spare. The test instrumentation needs to be unique in size and weight, which has numerous implications on its cost, power, and versatility. In this way, MAV test instrumentation exists at the cost of fuel and endurance. Because these platform-specific challenges to test and evaluation are unique, program managers need to have tasked HSI practitioners to address human concerns for each of the domains of HSI. Testing and evaluation falls within the bounds of the HFE domain because it deals not only with the satisfaction of design criteria but also with making sure that systems operation, maintenance, and support are appropriately suited to the capabilities and limitations of human operators and maintainers of the system (Air Force Human Systems Integration Office 2009). Due to the technical and manning requirements specific to MAVs, one should never attempt to reuse or adjust an existing test plan for a larger system and apply it to a MAV system. Doing so could result in major deficiencies in the test plan resulting in key attributes not being tested appropriately or at all. When considering the training and selection for UAV operators, it is important to note that there are no common standards across the branches of the U.S. military. The Air Force restricts UAV pilots to military pilots only, while the Navy, Marines, and Army require a pilot to have only a private pilot’s license. While human factors research has shown that appropriate levels of positive knowledge transfer is possible from manned flight to UAV control, further research is required to ascertain whether manned flight experience should be required (McCarley and Wickens 2004). These questions surrounding training standards and qualifications become more complex for MAVs where the desired model may be to place the common foot soldier in command of one or many vehicles. Further analysis, and ethnographic research,

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1323

for instance, should be required to determine what levels of training is appropriate for the piloting of MAVs from the battlefield. Manned flight experience may prove to be unnecessary as novel approaches to the command and control of MAVs are developed. Ideally, ground control station interface designs should incorporate good usability attributes that enhance the operator’s ability through good affordances to remotely pilot vehicles without extensive prior flight experience.

54.8

MAV Ground Control Station Considerations

If the design of a new system or product is to be successful, a considerable amount of time must be spent trying to understand who the users will be and what their needs are. Failure to do so will result in products that do not accommodate their intended users or their environments. One of the greatest pitfalls for designers is to design for themselves and fail to consider their actual users. Ground control stations for MAVs are more mobile in nature compared to other UAV platforms. Many of them are carried in backpack rigs by soldiers or deployed from trunks of vehicles. Further, a real possibility exists that MAV ground control stations will be used in harsh outdoor battlefield environments. Although incorporating solid user interface design principles into ground control station concepts, such as element recognition over recall, clear affordances, and continuous feedback, is crucial for successful outcomes in general, this section is concerned with addressing a few of the unique challenges that MAV ground control stations have to overcome in order to be successful. The most popular types of ground control stations serving MAVs are laptop-style designs. These come in many forms, and some of them are packed into ruggedized plastic boxes to prevent damage to system components. An advantage to the laptop style as a design is that training can be reduced because operators most likely already have familiarity with laptops in general. This style system is also advantageous because it can be rapidly deployed if necessary. Recently, there has been an interest within industry to develop universal ground control stations that promote interoperability. Achieving interoperability increases the efficiency of MAV systems and capabilities of military units to share information. Designers should pay attention to preexisting standards and requirements for interoperability established by various branches of the armed forces as well NATO’s standardization agreement (STANAG) when developing their own ground control station requirements. Even the best interoperable ground control station designs are subject to degradation under certain conditions. In bright urban environments or desert battlefields, direct sunlight can render a ground control station’s display useless. While hoods and shields can be helpful, they come at the cost of safety to the operator. If a MAV operator has his/her head buried in a display to interpret it or to maintain direct control over the vehicle, they can no longer observe their immediate physical surroundings. On battlefields or other environments where threats to the operator may abound, flying a MAV can be a high-risk activity. Minimally, the operator’s

1324

S. Michelson

colleagues will have to guard the operator. In other situations, the operator may be able to operate the MAV from within an armored vehicle, offering protection as well as reduction in direct sunlight on the display. Designers should consider how obvious their ground control station designs make an operator to potential enemies in order to avoid increasing their likelihood of becoming a special target. Therefore one should not discount the manner in which the introduction of MAVs at the platoon level impacts the operators’ preexisting job descriptions. Accounting for HSI early in a program (during the development of the concept of operations) can deter design induced manning and safety issues. There are other environmental concerns apart from sunlight. Ground control stations can fall victim to excessive heat, cold, moisture, sand, and dust to name a few. These sorts of design considerations are often addressed in standards and guidelines. Designers should reference these when making sure that their hardware is appropriately engineered to handle these sorts of threats to functionality. What might be overlooked however is that the human operator interacting with the ground control station is also subject to the adverse effects of those elements. Soldiers, for instance, may not have the luxury of choosing their operational environment, and their performance can be limited due to these elements degrading their sensory and physical abilities. As mentioned previously, proper human interface design principles can help to reduce these negative effects by ensuring that displays and controls are simple to use. It is important for designers to remember that the human operator while remotely piloting a MAV from a ground control station is separated from the sensory inputs that pilots receive in manned aircraft. This means that displays showing flight information need to be easy to read and operators should be able to trust them because their physical bodies are not actually feeling the effects of the forces that the vehicle is. For this reason, one should never assume that findings from studies about workload and cockpit control layout are scalable from manned aircraft to MAV ground control stations. The MAV ground control station should be considered as a unique entity with its own challenges in terms of workload and design layout. While a pilot’s inability to feel the forces of flight can be seen as a disadvantage, it can also be seen to enable the human operator, who may be able to remain calmer in riskier situations where his/her physical well being is not tied to the fate of the MAV. Designers might consider implementing haptic feedback to augment the physical effects that the MAV may be encountering. This can greatly enhance flight condition feedback to the operator. The display of ground control stations should not overburden a human operator’s mental resources. Selective attention and working memory are required because these two processes are particularly important to enhancing SA (Vidulich and Tsang 2012). An important consideration for MAVs that will be remotely piloted is the viewpoint that the display will show the operator. There are advantages and disadvantages to choosing an egocentric (one where the viewer sees what they would see as a pilot physically sitting in the vehicle) versus exocentric (a view from directly behind the vehicle) display, and one should consider the tradeoffs carefully. From one point of view, the environment is moving and the vehicle is

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1325

seen to be stationary, and from another, the environment is fixed and the vehicle is moving. An egocentric display may be preferable for obstacle avoidance, while an exocentric display may be better for acquiring an overall awareness of the terrain and environment (Vidulich and Tsang 2012). When designing for certain applications, it may be possible and advisable to create a display that affords the user both views simultaneously or the ability to switch easily. Whichever display type is selected, the design should support the human by accounting for his/her cognitive resources. Another consideration for designers of ground control stations is the design of controls. Designing for gloved operation is normally a good idea for joysticks, buttons, and latches. Any testing and evaluation of a system should reflect as closely as possible the actual operational environment in which the system will be deployed, and therefore test participants should be outfitted with all their equipment for an appropriate context including gloves or other elements which could inhibit mobility or fine motor control required for delicate articulation. For some applications, the adherence to certain MILSPEC guidelines and standards might be required with regard to design features such as gloved operation, button size, and spacing. In certain environments where the warfighter will be launching and commanding MAVs, light and noise discipline may hinder ground control station effectiveness. Systems that rely on light and audio to function are poor designs for this application. One should not design command and control to be limited to a single type of control input such as speech recognition. Alerts and alarms should not be audible in such conditions and displays should be designed to consider the potential for immediate muting. The necessity of night vision integration may also prove useful in certain applications. Another element to consider for the battlefield is that if the recognition of audible alerts and alarms is vital to proper system use, they can easily be muffled by the sounds produced by a firefight.

54.9

Ruggedization

The primary attribute of MAVs that distinguish them from other man-made flying things is their small size. This means that MAVs are more fragile than other UAVs, and how they transported and deployed in unforgiving environments such as the battlefield is worthy of consideration. Many MAVs are designed to fit into protective bags and cases that can be attached to the warfighter’s loadout. Many MAVs are designed to be packable within the standard Modular Lightweight, Load-carrying Equipment (MOLLE) system. Providing bags for compatibility with common equipment attachment methods is desirable but does not guarantee the level of protection that may be necessary to protect the smallest and most fragile MAVs. In the future, MAVs are going to continue to shrink in size, and it stands to reason that their fragility will continue to increase. This reality presents a challenge for the future as manufacturers of military components have had difficulty in the past keeping up with the speed of the evolution of commercial off-the-shelf components. Components get smaller, faster, and more rugged, while military systems can get bogged down in expensive

1326

S. Michelson

and time-consuming tests to make sure they meet military standards (Military and Aerospace Electronics 2004). Applied Research Associates’ Tactical Mini-Unmanned Air Vehicle (TACMAV) was a novel approach to storage and carrying. TACMAV had a 20.9-in. wingspan, and its flexible wings folded up around its carbon fiber fuselage. The entire system was storable in a 22  5 in. tube that attached to a soldier’s backpack and weighed 0.8 pounds (United States Department of Defense 2009). Designers should consider not only the vehicle itself but its ground control station’s protection for harsh environments. The ruggedness of a MAV system should be driven by its intended use. Systems deployed inside of structures should be seen in contrast to those deployed outdoors. Standards and guidelines can be useful to designers as they determine to what degree a system’s packaging needs to repel water, heat, cold, or other elements. When examining ruggedization from an HSI perspective, one should consider the survivability domain and it’s interdependence on the other HSI domains.

54.10 Conclusion Remaining mindful of the distinction between Human Systems Integration and Human Factors Engineering with regard to MAV development is pivotal to correctly accounting for the human in the design process. The uniqueness of MAVs and how they are deployed is forging a new understanding of man-machine systems as their intelligence expands the capabilities of humans in diverse operational environments. Teamwork in the future among humans and MAVs will depend on systems appropriately accounting for operator-specific elements such as workload, situational awareness, and training. Developing and implementing a Human Systems Integration Plan throughout acquisition is part of a total systems approach and will ensure that the human is considered throughout the entire design process. While specific design choices regarding ground control station interfaces or ruggedization will vary from system to system, one must remember that Human Factors Engineering is comprised of a series of tradeoffs. The individual choices made should support the operational performance objectives set forth for a specific MAV system to accomplish its mission. The relative merits of the issues discussed in this work are driven by how a particular system is intended to be deployed. For example, MAVs intended for urban combat environments will require a different set of design criteria than ones intended for agricultural surveying. As civilian and military operators continue to appreciate the valuable differences MAVs can make in protecting human life and accomplishing formidable missions, much of the focus moving forward will be on meeting the technical and political challenges of achieving interoperability, enhancing communications, managing system affordability, furthering autonomous behavior, increasing endurance, and integrating into national airspace. MAVs will no doubt continue to be an

54 Survey of the Human-Centered Approach to Micro Air Vehicles

1327

increasingly popular solution for warfighters as they operate in dynamically challenging environments, and therefore the means of further integrating MAVs at the platoon level will be a concern. In a time when so much of the media attention surrounding MAVs is focused on advances in technical capability and policy considerations, one must not forget to pay attention to the needs, wants, and desires of human operators within the design space. How well designers implement a human-centered approach to MAVs in the future will largely dictate the platform’s success.

References Air Force Human Systems Integration Office, United States Air Force FY09 Human Systems Integration Management Plan (U.S. Government Printing Office, Washington D.C., 2009) S. Czaja, S. Nair, Human Factors Engineering and Systems Design in Handbook of Human Factors and Ergonomics, 4th edn. (Wiley, New Jersey, 2012), pp. 38–56 J. Drury, S. Scott, Awareness in unmanned aerial vehicle operations. Int. C2 J. 2(1) (2008), pp. 1–28 M. Endsley, Toward a theory of situational awareness in dynamic systems. Hum. Factors 37(1), 32–64 (1995) M. Endsley, Situation Awareness in Handbook of Human Factors and Ergonomics, 4th edn. (Wiley, Hoboken, 2012), pp. 553–568 Georgia Tech Research Institute, Human systems integration division (2008), www.gtri.gatech.edu R. Hopcroft, E. Burchat, J. Vince, Unmanned Aerial Vehicles for Maritime Patrol: Human Factors Issues (DSTO Defence Science and Technology Organisation, Victoria, 2006) G. Klein, J. Bradshaw, Ten challenges for making automation a “Team Player” in joint humanagent activity. IEEE Intell. Syst. 19(6), 91–95 (2004) J. McCarley, C. Wickens, Human Factors Concerns in UAV Flight (Institute of Aviation, University of Illinois, Urbana-Champaign, 2004) J. McMichael, S. Francis, Micro air vehicles – toward a new dimension in flight, federation of American Scientists (1997), http://www.fas.org/irp/program/collect/docs/mav auvsi.htm. Accessed 1 May 2012 R. Michelson, Test and evaluation of fully autonomous micro air vehicles. ITEA J. 29(4), 367–374 (2008) Military and Aerospace Electronics, Rugged Computers Become Everyday Battlefield Equipment 15(1), 22 (2004) R. Parasuraman, K. Cosenzo, E. De Visser, Adaptive automation for human supervision of multiple uninhabited vehicles: effects on change detection, situation awareness, and mental workload. Mil. Psychol. 21(2), 270–297 (2009) United States Air Force, Air Force Human Systems Integration Handbook (U.S. Government Printing Office, Washington D.C., 2009) United States Department of Defense, Unmanned Aircraft Systems Roadmap 2005–2030 (U.S. Government Printing Office, Washington D.C., 2005) United States Department of Defense, Unmanned Systems Integrated Roadmap 2009–2034 (U.S. Government Printing Office, Washington D.C., 2009) United States Department of Defense, Defense Acquisition Guidebook (U.S. Government Printing Office, Washington D.C., 2012) M. Vidulich, P. Tsang, Mental Workload and Situation Awareness in Handbook of Human Factors and Ergonomics, 4th edn. (Wiley, Hoboken, 2012), pp. 243–273 C. Wickens, J. Lee, Y. Liu, S. Gordon-Becker, Introduction to Human Factors Engineering (Prentice Hall, Upper Saddle River, 2003)

Development of Insect-Sized MAVs

55

Shigeru Sunada, Hao Liu, Hiroshi Tokutake, and Daisuke Kubo

Contents 55.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1330 55.2 Flapping Wing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1332 55.2.1 A Hummingbird-Sized Flapping-Wing MAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1333 55.2.2 The MAV’s Flexible-Wing Kinematics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1335 55.2.3 The MAV’s Flexible-Wing Aerodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1336 55.3 Rotary Wing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1344 55.3.1 Ducted-Fan UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1344 55.3.2 Quadro-Rotor UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1345 55.4 Tail-Sitter MAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1346 55.4.1 Sato’s Spherical Air Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1347 55.5 Flow Sensor on the Wing Surface Mimicking Hairs on an Insect’s Body . . . . . . . . . . . . . . . 1348 55.6 Effects of Flexibility and Corrugation on Wing Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1353 55.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1356 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1356

S. Sunada () Department of Aerospace Engineering, Osaka Prefecture University, Sakai, Osaka, Japan e-mail: [email protected] H. Liu Graduate School of Engineering, Chiba University, Chiba, Japan Shanghai Jiao Tong University and Chiba University International Cooperative Research Center (SJTU-CU ICRC), Shanghai, China e-mail: [email protected] H. Tokutake Faculty of Mechanical Engineering, Kanazawa University, Kanazawa, Ishikawa, Japan e-mail: [email protected] D. Kubo Aviatios Program Group, Japan Aerospace Exploration Agency, Tokyo, Japan e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 9, © Springer Science+Business Media Dordrecht 2015

1329

1330

S. Sunada et al.

Abstract

MAVs (micro air vehicles), with a maximal dimension of 15 cm and nominal flight speeds around 10 m/s, operate in a Reynolds number regime of 105 or lower, the same as that of most natural flyers, including insects, bats, and birds. Due to the light weight and low flight speed, the MAV’s flight characteristics are substantially affected by environmental factors such as wind gust and demonstrate distinguished features with fixed, rotary, and flapping wings. Like natural flyers, the wing structures of MAVs are often flexible and tend to deform during flight. Consequently, the aero/fluid and structural dynamics of these flyers are closely linked to each other, making the entire flight vehicle difficult to analyze. In this chapter, a prototype bio-inspired flapping MAV with flexible wings is described with a specific focus on the flexible-wing aerodynamics. The flapping-wing MAV has a weight of 2.4–3.0 g and a wingspan of 10–12 cm, which is comparable to hawk moths and hummingbirds. The MAV’s flexiblewing aerodynamics is analyzed by combining an in-house computational fluid dynamics (CFD) method and wind tunnel experiments (EXP). In addition, fixedwing and rotary-wing MAVs with elements that enable the miniaturization of an aerial vehicle will be introduced.

55.1

Introduction

A large earthquake occurred in the Tohoku area of Japan on March 11, 2011. A devastating tsunami caused by the earthquake then hit the area, destroying all the power supplies at the nuclear power plant at Fukushima. The plant subsequently lost the ability to cool down the reactor core. The building covering the reactor vessel was destroyed by the explosion of H2 , and a great amount of radioactive materials were released from the power plant. Various methods have been undertaken to clean up after this accident. Acquiring visual information of the power plant from a high altitude using a small aerial vehicle is instrumental in this regard. The Global Hawk (Northrop Grumman Corporation), an airplane of the Fuji IMVAC Corporation, and the T-Hawk (Honeywell Corporation), a ducted-fan air vehicle, were used for this purpose. A photograph taken by Air Photo Service, Inc. by using the Fuji IMVAC Corporation airplane is shown in Fig. 55.1. The importance of unmanned aerial vehicles was confirmed by the activities of these meter-sized aerial vehicles. Smaller aerial vehicles, in particular, the micro air vehicle (MAV), have also been studied and developed. MAVs are now the subjects of an active and wellintegrated area of research, attracting participation from a wide range of talent. With a maximal dimension of 15 cm and nominal flight speeds of around 10 m/s, MAVs are desired for performing missions such as environmental monitoring, surveillance, and assessment in hostile situations. MAVs normally operate in a Reynolds number regime of 104 –105 or lower, the same as that of most natural flyers, including insects, bats, and birds. The prominent feature of the MAVs’ aerodynamics, in general, is characterized by a large-scale vortex flow structure and, hence, is highly unsteady

55 Development of Insect-Sized MAVs

1331

Fig. 55.1 Overview of the nuclear power plant at Fukushima after the earthquake on March 11, 2011 (Courtesy of Air Photo Service, Inc.)

(Shyy et al. 2007). Further, with the flexible wing structures and wing deformation of MAVs, the aero/fluid and structural dynamics of these flyers are closely linked to each other, making the entire flight vehicle difficult to analyze and design. In the past decade, there has been a remarkable increase in the research and development of MAVs, and numerous vehicle concepts, including the fixed wing, rotary wing, and flapping wing, have been proposed (Mueller 2001; Platzer et al. 2008; Stanford et al. 2008; Liu et al. 2010). AeroVironment, Inc. has developed a very tiny aerial vehicle with flapping wings, the NANO hummingbird. Keennon et al. (2012) at AeroViroment state about the NANO hummingbird as follows: The NANO hummingbird differs from other ornithopters in its ability to control attitude in roll, pitch, and yaw during hover through the actuation of the same flapping wings it uses for propulsion, whereas other hovering ornithopters have used large tail surfaces to deflect the wake of the main wings. By performing flight and control with its wings alone, the NANO hummingbird moves one step closer to nature’s flyers, which commonly demonstrate the ability to hover without large tail areas or large tail deflections. According to online documents about the NANO hummingbird (http://www. avinc.com/nano), this flapper can: 1. Demonstrate precision hover flight 2. Demonstrate hover stability in a wind gust flight that requires the aircraft to hover and tolerate a 2-m/s (5 miles per hour) wind gust from the side without drifting downwind more than 1 m

1332

S. Sunada et al.

3. Demonstrate a continuous hover endurance of 8 min with no external power source 4. Fly and demonstrate a controlled transition flight from hover to 11 miles per hour fast forward flight and back to hover flight 5. Demonstrate flying from outdoors to indoors and back outdoors through a normal-size doorway 6. Demonstrate flying indoors “heads-down” where the pilot operates the aircraft looking only at a live video image stream from the aircraft, without looking at or hearing the aircraft directly 7. Fly in hover and fast forward flight with a bird-shaped body and bird-shaped wings The NANO Hummingbird has high flight capabilities, as stated above. When its flight abilities are increased, it will be used to accomplish a true mission. Is this high performance of the NANO Hummingbird caused by the flapping wings? At the present stage, one cannot deny the possibility that a tiny aerial vehicle with a configuration other than an entomopter/ornithopter has high performance flight capabilities. The following flights are required for a tiny aerial vehicle to accomplish a real mission: 1. Flights under a wide range of forward velocities including hovering 2. Flights with high acceleration and angular acceleration 3. Flights that are very robust to wind gusts Independent of configuration, when an aerial vehicle becomes smaller, it is strongly affected by wind gusts, and the frequencies of some oscillated motions are larger, while the time constants of some converged/diverged motions are smaller. By considering these characteristics of reduced-size aerial vehicles, insect-sized aerial vehicles, which can accomplish the flights stated above, must have the following abilities: 1. Large and quick control force/moment 2. Small variations of aerodynamic force/moment when wind gusts are encountered 3. Sensors and actuators with short time-delay and high accuracy In this chapter, recent studies and developments of the tiny entomopter/ ornithopter and an aerial vehicle with other configurations that are thought to be suitable for downsizing will be introduced.

55.2

Flapping Wing

An airplane, which flies with a fixed wing, and a helicopter, which flies with a rotary wing, are the primary human-passenger aircrafts. Flying living creatures, such as birds and insects, fly with flapping wings. What type of aircraft is suitable for an insect-sized airplane? One answer might be an ornithopter, which is a vehicle with flapping wings. This is because birds and insects with very high flight capabilities have flapping wings. Sigthorsson et al. (2012) showed by their analysis that the flapping-wing vehicle is less sensitive to velocity perturbations in all aerodynamic forces and moments than the rotorcraft, except for the forward/backward

55 Development of Insect-Sized MAVs

1333

direction force. The following are the presumed advantages over an airplane with a propeller(s): 1. An ornithopter does not require a thruster because flapping wings can generate thrust as well as lift. This contributes to reducing the total weight of the vehicle. 2. A propeller generates pitching and yawing moments when it encounters side and vertical winds, respectively (Ribner 1943). This disturbs the attitude of an airplane with propellers. This does not happen when an ornithopter encounters side and vertical wind gusts. Moreover, Iida pointed out at 2011 Annual Meeting, Japan Society of Fluid Mechanics that the lift generated by an ornithopter is independent of turbulence. On the other hand, lift on a fixed wing at a small angle of attack of 6 deg is decreased as turbulence is increased. Though more research is required, this result raises the possibility that an ornithopter is more robust to turbulence than a fixed-wing airplane. All of the successful flapping-wing MAVs developed up to this point have flexible and light wings such as those observed in natural biological flyers (Wootton 1981), which indicates that wing flexibility is likely to have a significant influence on the resulting aerodynamics, as well as the flight stability (Young et al. 2009; Mountcastle and Daniel 2009; Shyy et al. 2010; Liu et al. 2010). In a sense, flapping flexible-wing aerodynamics is of great importance not only in uncovering the novel mechanisms of insect and bird flight but also in designing efficient flapping flight vehicles. Recently, a biologically inspired flapping-wing MAV is developed, which has four flexible wings. In this prototype MAV, the clap and fling mechanism is applied and achieved by a prototype crank system, not only because such a mechanism is observed in insect flight and thought to be capable of enhancing the aerodynamic force generation (Weis-Fogh 1973) but also because such physical interaction can affect the in-flight deformation of flexible flapping wings and, hence, aerodynamic performance.

55.2.1 A Hummingbird-Sized Flapping-Wing MAV A prototype flapping MAV developed at Chiba University (Nakata and Liu 2011; Nakata et al. 2011) is illustrated in Fig. 55.2a. The wingspan is designed to be around 12 cm – the size observed in hawk moths and hummingbirds. The wing has a semielliptic planform that is made of a polyethylene film with a thickness of 0.03 mm and a carbon rod with a diameter of 0.3 mm at the leading edge. The wing length is 60 mm, and the wing chord length is 30 mm at the wing base. The mean chord length is calculated as 23.6 mm. Two pairs of wings are attached to a crank system that is designed to achieve the clap and fling mechanism (Weis-Fogh 1973). As illustrated in Fig. 55.2b, the wings touch thrice in a wing beat on the top by upper wings and on the side by upper and lower wings. The gearbox is fabricated by cutting the acrylonitrile-butadiene-styrene (ABS) resin so as to ensure a nice match with the motor (MK04S-10, DIDEL), the gears, and the wing hinges. A 12-tooth pinion gear made from polyacetal with a module of 0.3 is attached to the motor.

1334

S. Sunada et al.

a

b

Upper wings

Lower wings

Fling

Clap

Clap

Fling

Fig. 55.2 (a) A prototype flapping micro air vehicle (MAV): a bio-inspired MAV. (b) Schematic of wing kinematics in an X-type wing MAV (viewed from the leading edge)

With a speed-reduction ratio of 60/12 teeth of the idler gear, the crank is mounted to link and actuate the two pairs of wings on the 60-tooth final gear. The gearbox system, the crank, and the wings are connected by a carbon rod with a diameter of 0.5 mm with the tail, the rudder, the receiver, and the remote control. The rudder is controlled by a magnetic actuator (hinge act, Plantraco) that moves laterally, weighs 0.23 g, and can provide sufficient control power. The remote control with infrared ray offers two channels to control both the motor frequency and the rudder angle. A rechargeable lithium polymer battery (FR30SC, Fullriver) is utilized as the power source. With all the parts mounted together, their flapping MAV weighs less than 3 g in toto and is able to fly with a time duration of up to 6 min, a maximum height over 10 m, and a region of 20  20 m.

55 Development of Insect-Sized MAVs

1335

55.2.2 The MAV’s Flexible-Wing Kinematics The flexible-wing aerodynamics is measured with a high-speed camera filming system as depicted in Fig. 55.3a–d by tracking the body-wing motions and deformations during tethered flapping flights. Three high-speed cameras (Miro, Vision Research Ltd.) with an image of 800  600 pixels at a frequency of 1,000 Hz are synchronized and operated for 2 s. Given that the MAV’s flapping frequency normally varies over a range of 20–35 Hz, the recorded image sequences are able to provide sufficient temporal resolution for the current flexible-wing kinematics. Calibration is done in a global three-dimensional system by a right-angled portable calibration frame, as shown in Fig. 55.3b, c. The markers (the white dots) are arranged in an array with an interval of approximately 100 mm, which is sufficient to reasonably resolve the flexible-wing kinematics during a complete wing beat.

Fig. 55.3 (a) High-speed camera filming system for wing kinematics measurement. (b) A rightangled portable calibration frame. (c) An image of the flapping MAV captured by a high-speed camera. (d) Reconstructed coordinates on the wing surface based on a sequence of images in a complete beat cycle

1336

S. Sunada et al.

The recorded image sequences are then downloaded to a computer, and the three-dimensional coordinates of these marked points are reconstructed utilizing commercial software, DippMotion (Ditect) (Fig. 55.3d). A realistic kinematic model of the MAV’s flexible wing is constructed by interpolating the reconstructed coordinates of the markers on the flapping wings. The displacements u(t, x, y/ at some point of the wing (x, y/ (Fig. 55.4d) are interpolated by using a function of Fourier series, such as u .t; x; y/ D

ny nx X n X X   a .l; m; n/ x l y m cos .n!t/ C b .l; m; n/ x l y m sin .n!t/ ; lD0 mD0 nD0

(55.1)

where terms a and b are derived by the least square method. A realistic morphological model of the MAV’s wing for computational fluid dynamics (CFD) analysis is constructed by tracing the outline of the wing planform. A uniform thickness is defined, but with elliptic smoothing at the leading and trailing edge, as well as at the tip. Figure 55.4b illustrates the computational geometric models and grid systems of the MAV. The CFD analysis is performed by introducing a symmetric plane as depicted in Fig. 55.4b under the assumption that the left and right wings move and deform symmetrically. Figure 55.5b shows the time courses of both the upper and lower wings at a wing cross section of 0.7 R from the wing base (Fig. 55.5a), in which the cross sections at the stroke reversal are shown by a dotted line. Note that those markers where the coordinates cannot be reconstructed due to the overlapping of upper and lower wings are excluded (Fig. 55.5b). Since the crank system can only generate the one-axial flapping motion, the wing rotation or the feathering motion is apparently induced passively due to the wing flexibility. The angles of attacks of both the upper and lower wings are, however, observed to somehow be in good timing, contributing to the lift force production during each half stroke. The upper wing is nearly vertical to the horizontal axis at both stroke reversals, which implies a “symmetric” rotational phase of the upper wing (Dickinson et al. 1999), and hence is likely a result of the clap and fling at each stroke reversal. The lower wing keeps the attitude throughout the half stroke while showing some “phase delay” at the stroke reversal, which may lead to lowering the aerodynamic performance of flapping wings.

55.2.3 The MAV’s Flexible-Wing Aerodynamics 55.2.3.1 CFD Modeling of Flexible Flapping Wing Evaluation of the aerodynamic performance of the bio-inspired flapping MAV is conducted by using an insect dynamic flight simulator (Liu 2009; Liu et al. 2010; Nakata and Liu 2011) that is designed to integrate the modeling of realistic wingbody morphology, realistic flapping wing and body kinematics, and unsteady aerodynamics in biological flight. The CFD study is performed under the assumption of

55 Development of Insect-Sized MAVs

a

1337 Symmetrical plane

b

c

d

Flapping wing with deformation

ui(t, xi, yi)

F

lb (xi, yi)

lc

la

Fz

x

Initial wing

y

Fx

Fig. 55.4 (a) A mechanical flapping-wing MAV model. (b) A computational fluid dynamic model of MAV wings and a multiblock grid system (single wing grid: 33  37  19, background grid: 89  129  81). (c/ Force measurement system for the mechanical flapping-wing MAV model. (d) Definition of displacement u at point (xi , yi / on the wing surface at time t

a hovering flight condition. Given the mean chord length cm as the reference length Lref and the mean wing tip velocity in hovering flight as the reference velocity Uref , which is proportional to Uref = !R, where R is the wing length and ! is the mean angular velocity of the wing (! = 2 ˚f , where ˚ is the wing beat amplitude and f is the flapping frequency), the Reynolds number in hovering flight can be reformed as 2˚fRcm ˚fR2 Uref Lref D D Re D   



 4 ; AR

(55.2)

where the aspect ratio AR is in the form of AR = .2R/2 =S with a wing area of S D 2Rcm . Note that the Reynolds number here is proportional to the wing beat amplitude, ˚, the flapping frequency, f; and a square of the wing length, R2 , but proportional inversely to the aspect ratio of the wing, AR.

1338

S. Sunada et al.

Fig. 55.5 (a) Definition of angle position f . (b) Time courses of both upper and lower wings at a wing cross section of 0.7 R from the wing base

The reduced frequency that normally characterizes rotational versus translational speeds is defined in the case of hovering flights as kD

 cm f Lref D : D Uref 2˚R ˚AR

(55.3)

Note that the reduced frequency k is proportional inversely to the beat amplitude ˚ and the aspect ratio AR of the wing. According to the measured data of the MAV’s mechanical model (cm D 23:6 mm, R D 60 mm, F = 1 rad, f D 18:5 s1 , n D 1:5  105 m2 /s), Re and k are calculated to be about 3,400 and 0.59, respectively. The CFD modeling of unsteady flows around the MAV’s flexible wings undergoing flapping is performed for a single flexible-wing model in which realistic geometric and kinematic models are utilized, as described in the preceding sections. The CFD-based results show that a leading-edge vortex (LEV) and, hence, a strong negative pressure region are generated on the upper and lower wings during both half strokes (Fig. 55.6a). As observed in insect flapping flight (Ellington et al. 1996),

55 Development of Insect-Sized MAVs

Fig. 55.6 (continued)

1339

1340

S. Sunada et al.

Fig. 55.6 (a) Instantaneous streamlines, iso-vorticity surface, and pressure contours on the upper surface of flapping wings at each half stroke. (b) Pressure contours on a virtual cylindrical surface at 2.0 cm from the wing base. (c) Wake topologies at the end of each half stroke. Velocity vectors and contours are visualized at a virtual cylindrical surface at 2.0 cm from the wing base

55 Development of Insect-Sized MAVs

1341

this LEV apparently plays a crucial role in the lift and/or thrust force production in the MAV flight. The vortex rings that are formed from the LEV, the tip vortex (TV), and the trailing-edge vortex (TEV) are also observed, showing a similar pattern to those of insect flight (Liu and Aono 2009). Obviously, strong negative pressure regions are detected between the upper right and left wings, as illustrated in Fig. 55.6b, which are likely induced by the clap and fling mechanism. Figure 55.6c illustrates the wake topology on a virtual cylindrical surface at the end of each half stroke. An intense downwash is generated behind the path of the wings identical to the center of the vortex rings (Liu 2009). The downwash disks are connected by the clap and fling with a pitching up motion, resulting in a pronounced downwash in the far wake below. In contrast, it is interesting to see that the downwash generated by the lower wings without the clap and fling is rather weak and there is almost no high-speed zone visible downward in the far wake. Figure 55.7 shows the time courses of aerodynamic forces generated by the upper and lower wings in a wing beat. Clearly, the upper wing generates very few aerodynamic forces until 0.15 T after the clap and fling (the vertical dashed line in Fig. 55.7), while the aerodynamic forces generated by the lower wing show a stable increase after the stroke reversal. However, the horizontal force Fx by the upper wing shows a subsequent rapid rise having a higher peak than the lower wing. Then, force Fx continues to decrease to a low level until 0.6 T after the next stroke reversal (the vertical dashed line in Fig.55.7), while the upper wing shows a larger peak again at the half stroke. Since the present flapping-wing mechanism generates large forces at each half stroke, one peak in the total Fx is observed at each half stroke. Furthermore, the mean aerodynamic force is calculated to be 23.3 mN, which is in good agreement with the measurement of a value of 26.46 mN acting on the electronic balance. The mean force components of Fx and Fz generated by the upper wing are 4.2 and 0.2 mN and those by the lower wing are 3.8 and 2.0 mN, respectively. The mean aerodynamic power consumed by the upper and lower wings is estimated to be 14.2 and 14 mW, respectively.

55.2.3.2 Wind Tunnel Experiments: Lift, Drag, and Flexible-Wing Kinematics The aerodynamic performance of the present prototype MAV (Fig. 55.3b, c) and a mechanical flapping-wing MAV model (Fig. 55.4a, c) that can measure forces is evaluated by means of wind tunnel experiments. The forces acting upon the MAV undergoing the tethered flight are measured by utilizing the load cells. Figure 55.8 shows the mean lift and drag forces plotted against the body inclination angle of qb , which varies from 0 to 70 deg. The lift force is seen to increase with increasing the body angle, the flapping frequency (with two frequencies of 10 and 20 Hz), and the wind velocity (when qb is greater than 40 deg), which reaches a value of 25 mN at a flapping frequency of 20 Hz and a wind velocity of 1.6 m/s, very close to the MAV’s weight. The drag force also shows an increase with increasing the body angle, but a negative drag force, i.e., the thrust force, is observed when the body angle is less than 50 deg. Note that the thrust force is generated merely in the case of no wind velocity and turns to the drag force at the larger wind velocity of 3.0 m/s.

1342 Fig. 55.7 Time courses of force components (a) Fx and (b) Fz on the right wings of the four-winged MAV

S. Sunada et al.

a : Total : Upper : Lower

10

Fx/mN

0

−10

−20

b

−30 20

Fz/mN

10

0

−10 0.15 −20

0

0.6 0.25

0.5 0.75 Stroke cycle, t/T

1

One key reason that the present MAV utilizes the X-type wing (Fig. 55.2) with four wings is because it can generate more lift and/or thrust forces than a twowinged MAV. Comparison of the lift forces between the two MAV types indicates that the present four-winged MAV does generate more lift force at almost twice the lift forces of the two-winged MAV when qb is greater than 40 deg. However, two times the lift force by the two-winged MAV with the thin Mylar wing shows a larger value than that of the present four-winged MAV.

55.2.3.3 The Clap and Fling Mechanism in Four-Winged MAVs The flapping-wing mechanism utilizing an X-type wing with four wings is employed for the present four-winged MAV, which is inspired by the smart two-winged natural flyer, the hummingbird, but obviously has crucial differences. Though there is a great variety of biological flyers, such as insects, bats, and birds, most fly with one pair of wings. The present flapping-wing mechanism is quite unique

55 Development of Insect-Sized MAVs

a

1343

b

40 : f = 20 Hz : f = 10 Hz : f = 0 Hz

30 20

20 10 0

10

−10

0 U = 1.6 m/s

−20

−10 40

20 Drag/mN

Lift/mN

30

U = : 3.0 m/s U = : 1.6 m/s U = : 0.0 m/s

30 20 10 0

10 0 −10

f = 20 Hz

−20

−10 40

30

: 4 wing : 2 wing : 2 wing (soft)

30

20 10

20

0

10 U = 1.6 m/s f = 20 Hz (4 wing) f = 22 Hz (2 wing)

0 −10

30

0

20

40

60

Body angle/deg

80

−10 −20

0

20 40 60 Body angle/deg

80

Fig. 55.8 (a) Lift and (b) drag forces plotted against body angles for four-winged and two-winged MAVs. The red line represents two times of lift and drag forces generated by the two-winged MAV

because of the multiple physical interactions in a wing beat cycle. Here, it is found that the four-winged MAV is able to generate two times greater lift force than that of a two-winged MAV at body angles larger than 40 deg where the MAV actually flies (Fig. 55.8). This implies that the present flapping-wing mechanism very likely utilizes the clap and fling mechanism effectively and hence enhances force production more than the two-winged flapping mechanism. On the other hand, it is seen that a two-winged MAV with a softer Mylar wing generates even more lift force than the four-winged MAV in the case of body angles larger than 30 deg (Fig. 55.8), which further points to the importance of the choice of an appropriate wing structure in the MAV design (Heathcote and Gursul 2007; Heathcote et al. 2008).

1344

55.3

S. Sunada et al.

Rotary Wing

A lot of meter-sized rotary-wing unmanned aerial vehicles (UAVs) are helicopters with attitudes controlled by cyclic pitch. The following two types of rotary-wing UAVs without cyclic pitch control seem to be successful.

55.3.1 Ducted-Fan UAV A ducted-fan MAV is a vehicle having a lift-augmented ducted fan driven by a reciprocating engine or an electrical motor at the center of the vehicle and aerodynamic surfaces (vanes) close to the outlet of the duct for flight control. This entails less danger to the operator and has an acoustic signature rather than a rotarywing-type MAV because of the shrouded rotator. In 2001, the Defense Advanced Research Projects Agency (DARPA) of the United States Department of Defense started the Organic Air Vehicle (OAV) program following the DARPA MAV program (1996–2000) (DARPA 2003). The OAV program is one of the technologies supporting the DARPA/ARMY Future Combat Systems (FCS) program. The smallest category of the FCS is a backpacksized vertical takeoff and landing (VTOL) MAV for platoon-level operations (FCS class I). One of the contractors of the OAV program is Honeywell, which developed the “T-Hawk” (originally called the “Honeywell MAV”) ducted-fan MAV (Daly 2011). It was selected for the FCS I platform vehicle, and further development included a high performance heavy fuel engine. The U.S. Army and also the U.S. Navy awarded T-Hawk contracts to Honeywell for deployment to Iraq. The U.S. military used the T-Hawk in search missions for roadside bombs in Iraq. The T-Hawk was also used to conduct surveillance of the damaged Fukushima Dai-ichi nuclear power station in Japan in 2011. Ducted-fan vehicles are unique in many respects: They can hover but are unlike helicopters, and they can dash at high speeds but are unlike airplanes. Much of the design intuition within the aerospace community is no longer applicable when venturing into the realm of VTOL ducted-fan vehicle design (Ohanian et al. 2010). Therefore, a wide variety of research and development has been conducted since the year 2000. Even just focusing on “micro” size, there is considerable research and development of ducted-fan MAVs all over the world (Table 55.1). Advanced methods for improving the aerodynamic characteristics of ducted fans are being investigated. Barrett (2004) applied piezoceramic actuator technology that enables higher bandwidth control than the conventional servomotor actuator for XQ-138 ducted-fan MAVs. Bilgen et al. (2010) applied a variable camber airfoil via smart materials for the control vanes in wind tunnel tests. Ohanian et al. (2011) applied synthetic jet actuators to the duct lip and the trailing edge to control the force and moment of the duct. Colman et al. (2011) experimentally investigated a cyclic and collective pitch control of a proprotor shrouded with an asymmetric duct.

55 Development of Insect-Sized MAVs Table 55.1 Developed ducted-fan MAVs

1345

MAV name XQ-138-600 (Barrett 2004) iMAV (Ohanian et al. 2010) SEMi (Kubo et al. 2009) RMIT Univ. (Zhao and Bil 2009) XQ-138-400 (Barrett 2004)

Rotor1

Duct dia. (in.) 6

Gross weight (g) D.ˆ   Pˆ/5 „ ƒ‚ … A BRM

D

.ˆ   Pˆ/> DR „ ƒ‚ …

(60.20)

bBRM

A 1 BRM bBRM :

Note that the inverse of A BRM always exists (Scherrer 2010). The closed-form solution for LSTD is given by 31

2

7 6  D 4ˆ > D.ˆ   Pˆ/5 ƒ‚ … „ A LSTD

> ˆ DR … „ ƒ‚

(60.21)

bLSTD

D A1 LSTD bLSTD : The LSTD solutions can be approximated using L1 sampled transitions in the form of .si ; ai ; ri ; si0 /, where si0 D si C1 : 2 6 e 6 ˆD 6 4

—–  > .s1 / —– —–  > .s2 / —– :: :

—–  > .sL1 / —–

3

2

6 7 7 g 6 7 ;PˆD 6 4 5

—–  > .s10 / —– —–  > .s20 / —– :: :

—–  > .sL0 1 / —–

3

2

6 7 7 e 6 7 ;RD 6 4 5

3 r1 r2 7 7 :: 7 ; : 5

(60.22)

rL1

e LSTD D 1 ˆ e   Pˆ/; g e > .ˆ A L1

(60.23)

1 e> e e ˆ R; bLSTD D L1

(60.24)

1464

S.S. Ponda et al.

e 1 e DA LSTD bLSTD :

(60.25)

In the limit of infinite samples, approximations become exact provided that the sampling policy is ergodic (Lagoudakis and Parr 2003). However, extending the BRM solution to the sample-based case is problematic. In particular, for calculating an unbiased estimate of A BRM , for each pair of si and ai , two next states should be sampled independently. This observation makes the BRM approach hard for domains with limited available samples. Once the new value function is calculated, a new policy can be obtained using Eq. 60.17. Least-squares policy iteration (LSPI) (Lagoudakis and Parr 2003) is an algorithm based on LSTD evaluation and policy improvement. A key property of the LSPI algorithm is that it does not require access to the model of the MDP in order to solve it. Moreover, samples are gathered through a trajectory of experience. Such a framework is known as reinforcement learning (RL) and is widely used throughout the literature. A simple RL solver algorithm is SARSA (state, action, reward, state, action) (Rummery and Niranjan 1994), in which, given a tuple of experience at time t C 1 in the form of .st ; at ; rr ; st C1 ; at C1 /, the weights are updated as D C ˛ Œrt C Q.st C1 ; at C1 /  Q.st ; at / .st ; at /;

(60.26)

where ˛ is the learning rate. The policy at each step is calculated as  .s/ D

argmaxa Q.s; a/ with probability 1   Random.A/ with probability 

(60.27)

to satisfy ergodicity (i.e., all states will be visited infinitely often in the limit). It has been shown that, for a fixed policy, right step size parameters, and any linear function approximation, applying Eq. 60.26 will converge to the LSTD solution asymptotically (Tsitsiklis and Roy 1997). There are a vast set of advanced DP and RL methods in the literature (interested readers are referred to Sutton and Barto (1998), Bus¸oniu et al. (2010), and Bellman (2003)). As expected, solving POMDPs is much more difficult as compared to solving MDPs. As a result, the area for solving real-world POMDP problems has been barely touched upon (Spaan and Vlassis 2005; Zhou and Hansen 2001; Paquet et al. 2005).

60.2.3 Game Theory The field of game theory presents an alternate way of addressing the multi-agent planning problem by treating the interaction between agents as a game. The basic idea behind game theory is that agents are individual decision-making entities that perform actions to maximize their own local utility based on knowledge of other agents and the environment. As such, game-theoretic frameworks lend themselves naturally to solving autonomous task allocation problems in a decentralized fashion,

60 Cooperative Mission Planning for Multi-UAV Teams

1465

motivating significant work in this area (Arslan et al. 2007; Marden et al. 2009; Chapman et al. 2010; Marden and Wierman 2009). Since in game theory agents make individual decisions about their own actions, these frameworks are useful for modeling noncooperative environments; however, enforcing cooperation is difficult because it involves ensuring that individual agent utility functions and incentives are aligned with the global mission goals. Therefore, the main challenge associated with game-theoretic cooperative planning strategies involves designing proper utility functions and negotiation strategies to ensure appropriate levels of coordination and collaboration between agents in order to maximize global mission performance. This section describes current game-theoretic approaches to cooperative multiagent planning problems, including the design and selection of utility functions and negotiation strategies, and highlights the major associated challenges and limitations.

60.2.3.1 Problem Statement In the game-theoretic literature, a commonly studied cooperative multi-agent problem is the vehicle-target assignment problem (Murphey 1999), which is very similar in nature to the task allocation problem described in Sect. 60.2.1. The vehicle-target assignment problem consists of assigning a set of Na vehicles to a set of Nt targets with the object of maximizing the global utility. The vehicles may have varying characteristics and the targets different reward values. Using a gametheoretic formulation leads to a decentralized implementation framework where the vehicles are considered self-interested decision-making agents intent on maximizing their own local utility (Arslan et al. 2007). It is desirable to find an “agreeable assignment” (Arslan et al. 2007) where no agent has an incentive to unilaterally deviate (pure-strategy Nash equilibrium (Fudenberg and Tirole 1991)) and to design the agent utilities such that maximizing the local utilities corresponds to maximizing the global utility. The problem formulation consists of selecting assignments for a set of vehicles I , f1; : : : ; Na g given a set of targets J , f1; : : : ; Nt g. A vehicle’s assignment or action, ai 2 Ai , consists of choosing a target from an allowable subset of targets, Ai fJ [ ;g, where the action ; is a “null” target or an empty assignment and the total action space for all vehicles is A , A1  : : :  ANa . A joint assignment profile a 2 A consists of a vector describing the actions of all agents, a D .a1 ; : : : ; aNa /, and is sometimes written as a D .ai ; ai / to highlight agent i ’s action given the actions of the other agents, ai . Each agent wishes to maximize its own local utility such that for vehicle i , ai? D argmax Ui .ai ; ai /; (60.28) ai 2Ai

where ai is vehicle i ’s assignment and ai is the assignment profile of the other vehicles. Solving these for each agent gives a joint assignment profile a 2 A corresponding to a global mission utility Ug .a/. If every target has a reward value Rj .a/ as a function of the joint assignment profile, then the global utility is the total reward over all targets and is written as

1466

S.S. Ponda et al.

Ug .a/ D

Nt X

Rj .a/:

(60.29)

j D1

In game theory, a framework where maximizing local agent utilities corresponds to maximizing the global utility is referred to as a potential game (Monderer and Shapley 1996), and several recent approaches have considered fitting the vehicletarget assignment problem within this architecture (Arslan et al. 2007; Marden et al. 2009; Chapman et al. 2010; Marden and Wierman 2009). A potential game consists of a game where there exists a potential function .a/W A ! R such that for every vehicle i 2 I, for every ai 2 Ai , and for every ai0 ; ai00 2 Ai , the local utility functions satisfy Ui .ai0 ; ai /  Ui .ai00 ; ai / D .ai0 ; ai /  .ai00 ; ai /;

(60.30)

where the potential function is representative of the global utility. Potential games have several useful properties that can be exploited to obtain near-optimal solutions to the global problem (the reader is referred to Monderer and Shapley (1996)). The most important of these is that potential games are guaranteed to have at least one pure-strategy Nash equilibrium (PSNE), which, in the vehicletarget assignment problem, corresponds to a valid assignment profile where no vehicle has any incentive to deviate from its strategy given the other vehicles’ strategies. Several multiplayer learning algorithms that are guaranteed to converge to a PSNE have been recently developed (e.g., fictitious play, spatial adaptive play, and regret matching) (Arslan et al. 2007). However, the main issue with the potential game framework is that for most multi-agent problems, there usually exist many PSNE, and while several current algorithms guarantee convergence to a PSNE, they typically do not find the optimal PSNE (and in practice can produce arbitrarily bad assignment profiles). To mitigate this problem, it is necessary to carefully design the vehicle utilities and the negotiation strategies such that the resulting algorithm converges to a near-optimal PSNE, a challenging task that has generated much interest in the research community (Arslan et al. 2007; Monderer and Shapley 1996). Designing local utility functions to align agent utility with global utility is difficult, since consideration must be given to what information agents have access to and how much they need to communicate with each other in order to properly represent the mission goals. An obvious choice to align agent utilities with the global utility is the identical interest utility, where each vehicle’s utility is exactly equal to the global mission utility. Unfortunately, this requires that each vehicle have complete knowledge about the global state, including other vehicles’ actions, all target reward functions including targets that are not of interest to the specific agents, and all environmental parameters that may affect any agent. Given that autonomous networked teams typically have limited bandwidth and communication constraints, it is desirable to find a utility function structure that keeps the amount of global information an agent requires to a minimum while still aligning the agent’s local utility with the global utility. Better choices for localized utility functions that

60 Cooperative Mission Planning for Multi-UAV Teams

1467

have been considered are the range-restricted utility, the equally shared utility, and the wonderful life utility (see Arslan et al. (2007), Monderer and Shapley (1996), and Tumer and Wolpert (2004) for further details). The most popular of these is the wonderful life utility (WLU) due to its computational simplicity and favorable properties. The WLU is defined as Ui .ai ; ai / D Rj .ai ; ai /  Rj .;; ai /

(60.31)

for ai D j and represents the marginal utility obtained by an agent for doing task j given the assignments of the other vehicles. The WLU has the useful feature that the local vehicle utility is equal to the marginal contribution made to the global utility, leading to a potential game with potential function equal to the global utility (.a/ D Ug .a/). The next section discusses solution algorithms for potential games that ensure convergence to near-optimal equilibria.

60.2.3.2 Solution Algorithms As mentioned before, depending on the negotiation strategy, stable outcomes of most potential game algorithms might result in arbitrarily bad pure-strategy Nash equilibria (PSNE), motivating the design of negotiation strategies that increase the probability of convergence to a near-optimal PSNE. Furthermore, it is desirable for vehicles to negotiate with each other at every time step given all agents’ assignments only, without requiring internal knowledge about other vehicles’ utility functions or internal state. Several current negotiation strategies are reviewed in Arslan et al. (2007), including algorithms such as action-based fictitious play, utilitybased fictitious play, regret matching (and variants), and spatial adaptive play, and the benefits and drawbacks of each are described. One negotiation strategy that ensures convergence to a near-optimal PSNE is the spatial adaptive play (SAP) algorithm. The SAP algorithm is described as follows: at stage k, a vehicle is randomly chosen from the set of vehicles (e.g., vehicle i ). Vehicle i then proposes a target according to probability distribution pi .k/ that maximizes 31 Ui .1; ai .k  1// 7C B6 :: pi .k/ @4 5A C H Œpi .k/ ; : 02

(60.32)

Ui .Nt ; ai .k  1// where H./ is the entropy function which represents the level of randomization and the parameter  sets the reward for randomization. The closed-form solution for the maximizing probability distribution pi .k/ is then 0 2

31 Ui .1; ai .k  1// e xi 7C B1 6 :: pi .k/ D @ 4 : 5A ; .x/i D x1 :  e C : : : C e xn Ui .Nt ; ai .k  1//

(60.33)

1468

S.S. Ponda et al.

where ./ is the soft-max function. The SAP algorithm converges to a near-optimal PSNE with arbitrarily high probability which is tunable by the user through the parameter  (Arslan et al. 2007). Furthermore, the computational overhead required to run SAP is low, making real-time implementations feasible. Several extensions to the baseline game-theoretic task assignment formulation have been proposed in the literature. In Marden et al. (2009), weakly acyclic games are introduced to account for time-varying objective functions and action sets, as well as a restricted-action set variant of SAP termed binary restricted SAP. In Marden and Wierman (2009), it has been shown that noncooperative control formulations make it impossible to design budget-balanced agent utilities that guarantee that the optimal control is a PSNE. The proposed solution to address this issue consists of designing a modified agent utility that is conditioned on additional state information and recasting the problem to a stochastic statebased game framework. Finally, to address the problem of dynamic task allocation with task deadlines, Chapman et al. (2010) present a stochastic game formulation approximated by a series of overlapping potential games. The resulting overlapping potential game approximation (OPGA) is shown to be solvable in a distributed manner and is robust to communication restrictions.

60.3

Planning Architectures

There are many important aspects to consider when deciding what type of planning algorithm is most appropriate for a given application. An important observation is to note that, for most problems of interest, the planners being discussed in this chapter will only be able to approximately solve the problem. This is because for most applications, (1) optimality is computationally infeasible, even in the most efficient computation structures (Bertsimas and Weismantel 2005), and (2) the optimal allocation will only be accurate with respect to the objective function, which is often necessarily approximate. This section introduces the idea that communication between algorithmic modules is a fundamental aspect of partitioning classes of algorithms and will formalize definitions for three communication architectures: centralized, distributed, and decentralized. It is then shown how each of these architectures place constraints on whether the algorithm can run synchronously or not, which may affect overall algorithmic performance. Finally, this section introduces different methods for handling coordination and cooperation in the assignment space, discusses what types of information should be shared among agents, and explains how these choices affect the performance of different algorithms in their respective environments. The partition introduced in this section is not new, and similar insights have been presented in the literature (Shima and Rasmussen 2009; Smith and Davis 1981). In (Shima and Rasmussen 2009), the authors partition the design space into classes of algorithms according to three factors: message cost, distribution of computation, and severity of task coupling. Each of these three issues is roughly subsumed into the categories presented in this section. Earlier work described in

60 Cooperative Mission Planning for Multi-UAV Teams

1469

Smith and Davis (1981) stressed the difference between distributed problem solving (where decisions are made in a distributed fashion) and distributed processing (where data is processed in a distributed way, but decisions are still made in a relatively centralized way), and the term decentralized is used to describe systems that utilize both distributed problem solving and distributed processing. This section redefines these terms to more heavily rely on the communication aspects of the problem; however, the classes of algorithms contained in Smith and Davis (1981) remain consistent with the definitions proposed in this section. Furthermore, Smith and Davis (1981) also introduce a distinction in the level of cooperation required between what is referred to as task-sharing (where the module sub-problems are assumed to be relatively separable) and results-sharing (which requires more intense communication), and this distinction is also mirrored in the discussion on handling cooperation in the assignment space presented at the end of this section.

60.3.1 Centralized A centralized computation model refers to an algorithm running on a single machine, and where information passing between modules of the algorithm is done through shared memory. Fast centralized algorithms are often implemented using parallelized computation structures that can take advantage of the large number of cores available in modern computer systems. Centralized implementations can also be quite fast for algorithms that are significantly parallelizable, since each module in the algorithm has access to the current global algorithmic state almost instantaneously. This means that there is very little communication cost between the modules (communication cost includes both the extra resources needed to facilitate communication and the associated delays introduced). Despite this, some environments may not be ideal for centralized approaches. The following list describes four environments where centralized computation may not be advisable: 1. If large amounts of data have to be transmitted to the centralized solver’s location, a centralized architecture may not be ideal. Other architectures could allow processing of the data remotely to identify the “usefulness” of information prior to sending metadata to a centralized location (referred to as distributed processing in Smith and Davis (1981)). Additionally, more autonomous implementations could use this remote data to actually produce candidate assignments for these remote agents without ever communicating with a centralized location (referred to as distributed problem solving in Smith and Davis (1981)), thus creating localized cooperative assignments. 2. The solution speed of centralized solvers is throttled by the rate at which the important information reaches the computation structure(s). If communications are slow, unreliable, or expensive, it becomes harder to justify passing large amounts of possibly irrelevant data through the network. 3. If most of the information needed by each agent to create its cooperative assignments is local, much faster reaction times might be obtained by keeping

1470

S.S. Ponda et al.

all of the computations local as well and sharing only the results of these local computations with other agents through communication. 4. In an abstract sense, distributing the computation on multiple machines might also be preferred if there are very different types of calculations that need to be performed. If, for example, the computation has large portions that can be sped up using an optimized computational structure such as a GPU, it would be desirable to outsource these computations to more appropriate hardware. If the above criteria are not a concern for a particular application, centralized parallel algorithms are likely to perform quite well.

60.3.2 Distributed This section defines distributed algorithms as algorithms consisting of separate modules that are able to utilize reliable message passing. Distributed systems can be implemented on a single machine or on separate machines. In a distributed system, this communication between the agents can be either passed over a network or implemented as internal memory passing, but the aspect that defines the algorithm as distributed computation (as opposed to centralized computation) is that it utilizes separate memory partitions between the distributed algorithmic pieces. This introduces an extra layer of complexity over centralized solutions because shared information must be transmitted through a communication channel. Given this description, the defining characteristic of algorithms that are designed to operate in distributed environments is that strong message-passing channels are assumed to exist. A connection is defined as strong if each distributed node has knowledge of the existence of all other nodes it is able to communicate with, knows how to transmit messages to each of these nodes, has low message latency between each of these nodes, and has guarantees that messages will arrive reliably. One application of distributed algorithms is if the computational load is too immense for a single (possibly multi-core) machine. This strategy would utilize some number of machines communicating through messages in a local and reliable way. The trade-off to consider when moving from centralized to distributed computation strategies is that time lost in message communication must be offset by the additional computation available by using multiple machines. Distributed algorithms also perform well (as compared to centralized algorithms) when information is being acquired remotely and external processing can be utilized before metadata is transmitted to other modules. This is because the overall communication load on the system can be reduced dramatically for the same amount of computation. For example, in cooperative planning environments, this could involve an agent observing a local event, changing its own plan based on this newly observed information, and then reliably communicating the results to the rest of the distributed modules. Again, the most important aspect of these distributed algorithms is that they rely on stable communications and depend on sharing information among other modules reliably. If communication links are not sufficiently reliable, performance may degrade significantly, and decentralized assumptions may be more appropriate.

60 Cooperative Mission Planning for Multi-UAV Teams

1471

60.3.3 Decentralized Decentralized architectures are most useful in systems where information sharing is performed exclusively through messages (e.g., separate agents linked via a communication network but with no direct memory sharing). Decentralized algorithms are designed for environments where there are no rigid constraints placed on message delays, network connectivity, program execution rates, and message arrival reliability. This means that decentralized algorithms are the most robust to catastrophic changes in the communication environment. The price of this robustness may be conservative performance when communication conditions are actually favorable, but for truly decentralized environments, this trade-off may be preferable. Fully decentralized algorithms allow a high degree of autonomy for the sparsely connected agents. This is because decentralized architectures do not place rigid rules on how agents must interact; therefore, there is flexibility on how individual agents can make decisions. Given that communications between agents may be weak, decentralized algorithms could allow large teams to interact efficiently, without bogging the entire network down with an overly restrictive infrastructure. As discussed below in Sect. 60.3.4, decentralized architectures are also able to handle asynchronous communications much more naturally than distributed and centralized approaches.

60.3.4 Synchronous Versus Asynchronous Parallelized computation can be used to increase the performance in cooperative assignment algorithms. Given this, choices can be made about how to organize this parallel computation: there are highly structured, synchronous algorithms that enforce constraints on when computation can be done or more flexible structures that utilize asynchronous computation. Algorithms that use synchronous computation define rigid rules about when computation can happen. Often, synchronization takes the form of queuing up computation that may only be executed after certain event driven triggers. This is to ensure that the algorithm has predictable state during program execution. Synchronization is heavily utilized in most iterative algorithms. In these synchronous iterative algorithms, a set of modules often perform parallel computations, share state variables, and then wait until a certain trigger before computing the next iteration of the algorithm. In this process, guarantees can be made about the state of each agent. Furthermore, these synchronous structures are also useful because they allow algorithms to make assumptions about the information state of other modules. The cost of using synchronous algorithms is that, in some environments, it may take significant effort to enforce synchronous behavior. In distributed and decentralized algorithms, the computation triggers are often nonlocal and thus may need to be inferred through inter-module communication. In the literature, the effort involved in enforcing this synchronization has been referred to as the synchronization penalty (Bertsekas and Tsitsiklis 1989). In centralized approaches,

1472

S.S. Ponda et al.

the mechanisms required to enforce synchronization are fairly lightweight, but when information is being shared across a network and modules must spend time waiting for messages to arrive from physically separated machines, this penalty can become severe. Conversely, asynchronous computation is used when modules of the algorithm can run relatively independently of one another. Because of this, these structures work well in decentralized algorithms where they can utilize information whenever available and not on a rigid schedule. This introduces a fundamental tradeoff between the rigidity of forcing an algorithm to run on a synchronous schedule versus allowing the algorithm to run asynchronously but take a performance hit for losing information assumptions about the state of the other modules. Typically, in centralized and distributed environments, the synchronization penalty may be weak enough that it makes sense to utilize synchronous computation, whereas in decentralized approaches, the synchronization penalty can become crippling, and asynchronous computation often becomes necessary.

60.3.5 Performance Metrics Once the computational structures of the algorithm are chosen, there are several metrics that the designer will care about during the run time of the algorithm. Depending on the relative importance of each of the following metrics, different algorithms may be preferred: • Score performance specifically refers to how important it is that the final solution optimizes the global cost function. There can be cases where significant coupling exists in the score function, increasing the difficulty of finding a good plan. In these problems, it may be desirable for an algorithm to have strong performance guarantees with respect to the optimal. Conversely, in other scenarios it may be easy to find near-optimal allocations and other considerations may be more important. • Run time specifically refers to how much computational time the entire algorithm takes to complete, and, as will be contrasted below, this is typically thought of as a global measure for the entire team. For off-line solutions, acceptable measures of run time can be in the hours (or possibly days). When algorithms are run in the field, run time is typically thought of in the seconds range. • Convergence detection can be a difficult thing to quantify in general. It is usually trivial to post process when convergence occurred, but in some cases (especially decentralized implementations), it may be difficult to determine if the algorithm has actually converged. This leads to different definitions of what convergence should be. Does the algorithm require global convergence? Does it just require local convergence (only the individual module)? Does the algorithm require only partial convergence (in some cases, being sure about the next task to execute is enough)? The answers to these questions will impact what type of information is communicated as well as how much total communication is actually needed. • Convergence time is typically identical to run time in terms of global convergence metrics, but for local convergence or partial convergence metrics, it measures the

60 Cooperative Mission Planning for Multi-UAV Teams

1473

time until this local convergence is detected. With convergence time, typically smaller is better, and the smaller this number is, the faster the planner can react to dynamic changes in the environment. In very dynamic environments, convergence time may be one of the most important metrics for the mission designer. • Reaction time to situational awareness is closely related to convergence time. It looks specifically at the turn around time between acquiring information and incorporating it into the plan cycle. In some algorithms, information cannot be incorporated mid-convergence (while others allow for the inclusion of new information at any time). In designing algorithms, trade-offs will need to be made between convergence time and score performance to enable faster reaction times to situational awareness.

60.3.6 Coordination Techniques Coupling between the goals of distributed agents in decentralized environments is typically required for most missions. Because of this, the assignment variables that specify mission behavior can contain nontrivial constraints between them. As a result, cooperative allocation algorithms must be able to produce consistent values for these coupled variables. In centralized environments, the problem of checking that constraints between modules are satisfied can be handled trivially by using shared memory. In distributed and decentralized algorithms, the problem becomes more complex because the modules are communicating using messages. There are three main types of coordination approaches to consider: 1. One technique is to create an a priori partition in the assignment space, so each distributed agent can trivially generate conflict-free plans. 2. A second technique, commonly known as implicit coordination, involves performing situational awareness (SA) consensus (i.e., consensus on all variables that are not assignments) and hope that, given sufficiently close information, the agents are able to independently create conflict-free allocations. 3. The last technique is to directly incorporate the constraint feasibility of the allocations into the convergence process of the planning algorithm. In general, cooperative algorithms may use combinations of the three techniques listed above. The following discusses the main factors associated with determining which of these techniques is more appropriate given the application at hand. 1. Creating an a priori partition in the task environment effectively chops up the task space and only allows each agent to bid on a subset of the overall task set. Given a higher-level task allocation or a human operator, it may be reasonable to use this method for small teams in relatively static environments. This is because, for some environments, it may be obvious what agents should be servicing which tasks. However, in environments with large numbers of relatively homogeneous agents, or in dynamic environments, the task partitioning problem can become very complicated. By creating this partition, the algorithm is placing artificial

1474

S.S. Ponda et al.

constraints on which allocations are available, which may result in arbitrarily poor task assignments. 2. The main emphasis of implicit coordination is to arrive at situational awareness consensus (where situational awareness consensus refers to the process by which all agents agree on all variables relevant to the initial conditions of the task allocation problem) before starting the planning algorithm, then run effectively centralized planners on each agent independently. The basic premise of this method is that a consensus algorithm can generate consistent parameter values across an entire team, in which case each agent will generate identical final allocations. This method has been explored in the literature (Ren et al. 2005, 2007; Ren and Beard 2005; OlfatiSaber and Murray 2004; Alighanbari and How 2008a; Moallemi and Roy 2006; Olshevsky and Tsitsiklis 2006; Hatano and Mesbahi 2005; Wu 2006; Tahbaz-Salehi and Jadbabaie 2006) and is widely used because it is a relatively straightforward way to decentralize a task allocation algorithm. The benefit of using implicit coordination over task space partitioning (see 1 above) is that by not limiting the assignment space a priori, the algorithm is able to account for changes in information explicitly (e.g., dynamic tasks and additional state information). Using task space partitioning, any changes in information are likely to necessitate a repartitioning and redistribution of the task space, and if this is not executed, the resulting assignment is likely to perform poorly. In contrast, with implicit coordination, agents explicitly account for information changes through the situational awareness consensus protocol every time they replan, thus producing more relevant higher-performing assignments. Implicit coordination also has the added benefit that if the task environment is highly coupled (with many agent-to-agent and task-to-task constraints), then the plans produced are able to recognize and exploit this structure, thus producing assignments that include highly coordinated behaviors between the agents. All of these benefits come with the caveat that the situational awareness consensus must happen before the planner can start producing plans, and, in order to guarantee that agents produce the same assignments, this consensus process may require large amounts of communication bandwidth (Alighanbari and How 2008a). A potential difficulty with algorithms that utilize this implicit coordination strategy is that they often require the planning parameters to be known precisely, thus requiring that the consensus process be conducted until errors in situational awareness become very small. In fact, it is often the case that, if the final estimates of the planning parameters do not converge to within arbitrarily tight bounds, there is no guarantee that the resulting assignment will be conflict free. 3. The third technique ignores gaining any insight about the situational awareness of other agents and only requires that assignments between the team remain conflict free. The power of this approach is that all of the consensus effort is spent on achieving consistent values for the agent assignments. This solution is preferable if there are few inter-agent constraints and inter-task constraints (the more coupled the task environment becomes, the more difficult the task consensus problem becomes) or if the communication environment is not necessarily reliable such that it would be difficult to reach complete team-wide consistent situational awareness. For many decentralized applications, the primary goal is achieving a conflict-free distribution

60 Cooperative Mission Planning for Multi-UAV Teams

1475

of tasks that performs well, rather than requiring intense cooperation between agents to achieve the best possible assignment (such as the CBBA algorithm discussed in Sect. 60.4.1.1). Given that in these environments the communication links are often non-robust, especially across larger distances, only broadcasting information directly relevant to the assignment constraints may be preferable. There is a tradeoff between implicit coordination and algorithms using assignment consensus. The choice is really between deciding if it is easier to converge on a consistent assignment vector or converge to arbitrarily tight bounds on all other significant state variables.

60.3.7 Consensus As discussed above, a key component of cooperative decision making involves performing consensus among agents, which is defined as reaching an agreement on quantities of interest, such as plans, situational awareness, or other desired parameters. Most distributed planning approaches employ consensus algorithms, which are sets of rules, or protocols, that determine how information is exchanged between agents to ensure that the team will converge on the parameters of interest. As a simple illustrative example, the following linear consensus protocol can be used to converge on a continuous parameter z: xP i .t/ D

X

.xj .t/  xi .t// 8i;

(60.34)

j 2Ni

xi .0/ D zi 2 R; where each agent i computes errors with its set of neighbors Ni and uses these to correct its parameter estimate (Olfati-Saber et al. 2007). Collectively, the team dynamics for n agents can be written as an nth order linear system: xP .t/ D Lx.t/;

(60.35)

where L D D  A is known as the graph Laplacian, which is computed using an adjacency matrix A describing connectivity between agents (the elements aij are 1 if j is a neighbor of iP and 0 otherwise) and a degree matrix D D diag.d1 ; : : : ; dn /, with elements di D j 2Ni aij . The maximum degree, denoted as  D maxi di , is useful in bounding the eigenvalues of L, which, for an undirected connected network, can be ordered sequentially as 0 D 1  2      n  2 :

(60.36)

The eigenvalues of L can be used to predict convergence rates and stability properties of these linear consensus algorithms (in particular, 2 is related to speed of convergence, and n provides stability bounds in time-delayed networks).

1476

S.S. Ponda et al.

As shown above, the nontrivial eigenvalues are all positive, and exploiting the fact that Eq. 60.35 describes a linear system, the consensus algorithm is globally asymptotically stable and converges exponentially to an equilibrium with rate given by 2 (Olfati-Saber et al. 2007). Furthermore, for the system described P above, the algorithm is guaranteed to achieve a unique equilibrium zN D 1=n i zi , which is the average of all the agents’ initial values. Recent research has explored the effects of more realistic mission environments on these types of linear consensus algorithms for multi-agent teams. Some examples include analyzing the impact of time-delayed messages and dynamic network topologies on convergence and stability properties of the consensus algorithms. The work in Olfati-Saber et al. (2007) shows that global exponential convergence guarantees can be extended to dynamic networks as long as the network remains connected at each time t. The agents are guaranteed to reach consensus with convergence rate greater than or equal to ?2 D mint 2 .G.t//, where 2 .G.t// is the second eigenvalue of the Laplacian for graph G.t/. Similar guarantees can be made for time-delayed networks, where messages are received after a delay  instead of instantaneously. The system dynamics can be modified as follows: xP .t/ D Lx.t  /;

(60.37)

and global exponential convergence guarantees are retained for delays within the range  < =2 n . Note that convergence rates and robustness to time-delays can be improved by actively controlling the network structure (modifying G and L), which is an active area of research (Ren and Beard 2005; Ren et al. 2007; Jadbabaie et al. 2003; Blondel et al. 2005; Olfati-Saber et al. 2007). Consensus algorithms have been applied to a wide variety of distributed decisionmaking applications, ranging from flocking to rendezvous (Jadbabaie et al. 2003; Beard and Stepanyan 2003; Lin et al. 2003; Fax and Murray 2004; Ren 2006; Ren et al. 2007; Olfati-Saber et al. 2007). Most of these consensus algorithms are computationally inexpensive and guarantee convergence of the situational awareness, even over large, complex, and dynamic network topologies (Hatano and Mesbahi 2005; Wu 2006; Tahbaz-Salehi and Jadbabaie 2006). A common issue with classical consensus algorithms is that agents’ observations are often treated with equal weight, whereas in reality, some agents may have more precise information than others. Extending classical consensus algorithms to account for this uncertainty in local information, Kalman consensus approaches have been developed that approximate the inherent uncertainty in each agent’s observations using Gaussian distributions (Ren et al. 2005; Alighanbari and How 2008b). These algorithms produce consensus results that are more heavily influenced by agents with smaller covariance (therefore higher certainty) in their estimates. A limitation of Kalman consensus approaches is that Gaussian approximations may not be well suited to model systems with arbitrary noise characteristics, and applying Kalman filter-based consensus methods to the mean and covariance of other distributions can sometimes produce biased steady-state estimation results (Fraser et al. 2009).

60 Cooperative Mission Planning for Multi-UAV Teams

1477

Other Bayesian decentralized data and sensor fusion methods have been explored to determine the best combined Bayesian parameter estimates given a set of observations (Waltz and Llinas 1990; Grime and Durrant-Whyte 1994; Makarenko and Durrant-Whyte 2006). A major challenge, however, is that these decentralized data fusion approaches require channel filters to handle common or repeated information in messages between neighboring nodes. These channel filters are difficult to design for arbitrary network structures, and generic channel filter algorithms have not been developed other than for simple network structures (e.g., fully connected and tree networks), thus limiting the applicability of decentralized data fusion methods (Grime and Durrant-Whyte 1994). Recent work has addressed this issue by showing that, through a combination of traditional consensus-based communication protocols and decentralized data fusion information updates, scalable representative information fusion results can be achieved, without requiring complex channel filters or specific network topologies (Olfati-saber 2006; Xiao et al. 2005; Fraser et al. 2009, 2012). In particular, the work in Olfati-saber (2006) utilized dynamicaverage consensus filters to achieve an approximate distributed Kalman filter, while (Xiao et al. 2005) implemented a linear consensus protocol on the parameters of the information form of the Kalman filter, permitting agents to execute a Bayesian fusion of normally distributed random variables. However, as previously noted, these Kalman-based methods are derived specifically for normally distributed uncertainties (Olfati-saber 2006; Xiao et al. 2005), and thus can produce biased results if the local distributions are non-Gaussian. Recent work has extended these combined filtering and data fusion approaches to allow networked agents to agree on the Bayesian fusion of their local uncertain estimates under a range of non-Gaussian distributions (Fraser et al. 2012). In particular, the approach exploits conjugacy of probability distributions (Gelman et al. 2004) and can handle several different types of conjugate distributions including members of the exponential family (e.g., Dirichlet, gamma, and normal distributions) (Fraser et al. 2009, 2012). The approach in Fraser et al. (2012) is termed hyperparameter consensus and has demonstrated flexibility in handling several realistic scenarios, including ongoing measurements and a broad range of network topologies, without the need for complex channel filters.

60.4

Decentralized Algorithms

This section reviews current decentralized approaches proposed in the literature, describing the benefits, challenges, and limitations of each.

60.4.1 Market-Based Approaches Market-based algorithms are integer programming-based formulations that are able to solve distributed and decentralized cooperative assignment problems (Dias and Stentz 2000; Dias et al. 2006). Many successful market-based approaches use an

1478

S.S. Ponda et al.

auction mechanism (Bertsekas 1989, 2007). Auctions may be run via an auctioneer (Gerkey and Mataric 2002; Mataric et al. 2003; Gerkey and Mataric 2004), where each agent computes the reward associated with a task based on their own local understanding of the world and then uses this reward value to place bids on the most desirable tasks. Each agent communicates their bid to the auctioneer, and the auctioneer determines the winner for each task. These types of methods guarantee conflict-free solutions since the auctioneer only selects one agent as the winner. They are distributed because the bids are calculated at spatially separated locations, but they are synchronous in that they rely on one winner at every iteration. Some auction-based protocols do not need to designate a single agent as the auctioneer but utilize a different protocol where the winner is decided based on a set of self-consistent rules (Castanon and Wu 2003). Other such methods, called distributed auction algorithms (Sariel and Balch 2005; Ahmed et al. 2005; Atkinson 2004; Lemaire et al. 2004), have been shown to efficiently produce suboptimal solutions when reward functions satisfy a property called submodularity. One such method is the consensus-based bundle algorithm (CBBA) (Choi et al. 2009), which is a polynomial-time, market-based approach that uses consensus to reach agreement on the cooperative assignment. The following section describes CBBA and illustrates how distributed auction-based allocation algorithms can produce provably good approximate solutions in distributed and decentralized environments.

60.4.1.1 Consensus-Based Bundle Algorithm (CBBA) CBBA is a distributed auction algorithm that provides provably good task assignments for multi-agent, multi-task allocation problems. This algorithm has been shown to work well in practice because it can handle complicated constraints in a real-time distributed way. The algorithmic structure of CBBA consists of iterations between two phases: a bundle-building phase where each agent greedily generates an ordered bundle of tasks and a task consensus phase where conflicting assignments are identified and resolved through local communication between neighboring agents. These two phases are repeated until the algorithm has reached convergence. To further explain the relevant details of the algorithm, some notation will first be formalized: 1. A bid is represented as a triple: sij D .i; j; cij /, where i represents the bidding agent’s index, j represents the task’s index, and cij represents the bid value for this agent-task pair. 2. A bundle is an ordered data structure internal to each agent i , bi D fsij1 ; : : : ; sijn g that consists of all n of its current bids. When new bids are made, they are appended to the end of the bundle; thus, the order in the bundle reflects the relative age of each bid. 3. The bid space is an unordered set of bids, defined as A D fsi1 j1 ; : : : ; siN jN g. This set contains a globally consistent set of the current winning bids in the team. A local bid space Ai is defined as a set that contains agent i ’s current local understanding of the global bid space. In a fully connected network, Ai D A after each consensus phase, but in general, the geometry of agents in the network may lead to information propagation latencies and thus nonidentical bid spaces.

60 Cooperative Mission Planning for Multi-UAV Teams

1479

Bundle-Building Phase For each agent i , the bundle-building phase is run independently. 1. At the beginning of the bundle-building phase, the previous bundle is cleared and all bids that were won by agent i located in the local bid space, Ai , are also removed. This step is required for the optimality guarantees of the algorithm, but in most cases, the agent will re-add each of the tasks it has just dropped from Ai as it is building back up the bundle. 2. For each task j available in the environment, each agent i uses its local internal score function Fij .bi / D cij , which is a function of its current bundle, to create a score cij . These scores are then ordered from highest to lowest. 3. The ordered scores cij are compared with the winning bid information for the corresponding task j located in the agent’s local bid space Ai . The first (and thus largest) score that would outbid the current winner in the local bid space is chosen as agent i ’s next bid. A bid sij is created and placed at the end of the bundle and replaces the current bid on task j in the local bid space. Steps 2 and 3 are repeated until no tasks are able to outbid Ai or the maximum bundle length is reached, at which point the bundle-building phase terminates. It is worth noting that in this formulation, the values cij are used to rank the tasks, and the values sij are communicated as bids to other agents. Algorithm performance can be increased by explicitly separating these two values (i.e., internal scores vs. external bids) (Johnson et al. 2012), enabling a larger class of allowable (more representative) score functions while still ensuring that properties required for algorithmic convergence are met. Consensus Phase After the bundle-building phase completes, each agent i synchronously shares its current local bid space Ai with each of its adjacent neighbors. This local bid space, in combination with time-stamp information, is then passed through a decision table (see Choi et al. 2009, Table 1 for details) that provides all of the conflict resolution logic to merge local bid spaces. In general, the consensus logic prefers larger and more recent bids. If the consensus phase has occurred more than twice the network diameter times without any bids changing, the algorithm has converged and terminates; if not, each agent reenters the bundle-building phase and the algorithm continues. Score Function One requirement that is fundamental for all of the convergence and performance guarantees of CBBA is that the score functions used must satisfy a diminishing marginal gains (DMG) property. This DMG condition is a subset of the well-studied submodularity condition, was recognized in the original description of CBBA (Choi et al. 2009), and further studied in recent work (Johnson et al. 2012). It is defined as Fij .bi /  Fij .bi ˚end si k ? / 8k ? 2 f1; : : : ; Nt g;

(60.38)

where bi ˚end si k ? refers to adding a bid si k ? to an already existing bundle bi . Roughly, this condition means that no bids si k ? can be made that would increase cij , agent i ’s score for task j . Given score functions that satisfy DMG, CBBA is guaranteed to converge to a conflict-free solution despite possible inconsistencies

1480

S.S. Ponda et al.

in situational awareness and to achieve at least 50 % optimality (Choi et al. 2009) (although empirically its performance has been shown to be above 90 % optimality for some UAV environments (Bertuccelli et al. 2009)). The bidding process of CBBA runs in polynomial time, demonstrating good scalability with increasing numbers of agents and tasks, making it well suited to real-time dynamic environments. The real-time implementation of CBBA has been demonstrated for heterogeneous teams, and the algorithm has been recently extended to account for timing considerations associated with task execution, asynchronous communication environments, cooperative planning missions, and uncertain operation environments (Ponda et al. 2010, 2011, 2012; Whitten et al. 2011; Johnson et al. 2011).

60.4.2 Decentralized Markov Decision Processes As mentioned previously in Sect. 60.2.2, the centralized multi-agent MDP formulation implicitly assumed that each agent had individual full observability of the joint state. When agents are making decisions and observations locally, this is generally an impractical assumption, since it would require each agent to communicate their observations to every other agent at every time step with zero communication cost (commonly referred to as a free comm environment). In realistic missions, however, there are operational costs and constraints associated with communication (e.g., bandwidth or dynamic network topologies) which need to be accounted for when planning. To address this issue, several decentralized MDP formulations have been proposed which extend the centralized multi-agent MDP (or POMDP) framework to explicitly account for communication between agents. However, this additional layer increases the dimensionality of the problem which exacerbates computational intractability. This section reviews some of these decentralized algorithms, discusses associated challenges, and presents approximations that can be made to address computational complexity. The decentralized MDP (Dec-MDP), originally proposed in Bernstein et al. (2002), relaxes the limiting assumption of individual full observability by replacing it with a joint full observability requirement (i.e., together the agents have full observability of the state, but individual agents are not required to know the full state on their own). Agents can share parts of the state with each other through communication after a certain number of time steps, but this additional level of communication increases the computational complexity required by the algorithm (double exponential complexity or NEXP). As a reminder, the complexity hierarchy is given by P NP PSPACE EXP NEXP denoting which classes of problems are subsets of other classes with regard to increasing computational complexity. For more details and discussion on complexity classes, the reader is referred to Bernstein et al. (2002), Papadimitriou and Tsitsiklis (1987), and Papadimitriou (2003). Similar to the Dec-MDP, the Dec-POMDP (Bernstein et al. 2002) is the decentralized version of the POMDP, where there is additional uncertainty in the state given information provided by the observations (the global state cannot be uniquely determined given a joint observation). Like the Dec-MDP,

60 Cooperative Mission Planning for Multi-UAV Teams

1481

the Dec-POMDP assumes that agents have different partial information about the global state. If it is assumed that each agent communicates its observations to every other agent at every time step with zero cost, then the Dec-MDP and Dec-POMDP reduce to the MMDP or MPOMDP described in Sect. 60.2.2. This situation is referred to as free comm, and the complexity of the Dec-MDP/Dec-POMDP reduces to that of its MDP/POMDP counterpart (P or PSPACE complexity, respectively). Therefore, the lack of immediate communication (or isolation) between agents can be associated with a jump in computational complexity (Goldman and Zilberstein 2004). Another framework similar to the Dec-POMDP that has been considered in the literature is the multi-agent team decision problem (MTDP) (Pynadath and Tambe 2002) which also consists of decentralized MDP solutions for cooperating agents. The MTDP and the Dec-POMDP have been shown to be equivalent in terms of computational complexity and expressiveness of representation (Seuken and Zilberstein 2008). Since communication is a key component of decentralized planning, the above formulations have been extended to explicitly model communication between agents. The resulting algorithms, namely, the Dec-POMDP-COM (Goldman and Zilberstein 2003) and the COM-MTDP (Pynadath and Tambe 2002), treat these communications as actions which agents must make explicit decisions about at every time step. In particular, sending a message is considered an action which incurs an associated cost, and each agent needs to decide what messages to send at every given time, where not sending anything is modeled as a null message with zero cost. The reward function is accordingly modified to include the cost of communication. As such, the problem formulation can be extended to make sequential decisions about which messages to send as well as which actions to perform. The tuple now becomes < S; A; P; R; ˝; O; ˙; C˙ ; RT ;  >, where the additional terms are defined below: • ˙: Defines the joint message space ˙ D ˙1   ˙n composed of a discrete set of message options, where ˙i is the message space for agent i and joint message vector 2 ˙ is of the form D . 1 ; : : : ; n / representing the messages sent by each agent. • C˙ : Defines a function quantifying the cost of communication where C˙ . / W ˙ ! R is the cost of sending message and is typically a sum over all individual agent message costs, with a null message having zero cost. • RT : Defines a total reward function RT .s; a; s 0 ; / W S  A  S  ˙ ! R specifying the total rewards and costs associated with applying action a 2 A in state s 2 S, transitioning to state s 0 2 S, and communicating message 2 ˙. Typically, RT .s; a; s 0 ; / D R.s; a; s 0 /  C˙ . /. In this framework, a joint policy  D .1A ; 1˙ ; : : : ; nA ; n˙ / contains individual agent policies i D .iA ; i˙ / composed of local policies for actions and for communications, respectively. In Seuken and Zilberstein (2008), the computational complexity of the Dec-POMDP, MTDP, Dec-POMDP-COM, and COM-MTDP is analyzed, and all of these algorithms have been shown to be NEXP-complete (nondeterministic exponential time), suffering from bad scalability as the number of agents increases.

1482

S.S. Ponda et al.

The varying degrees of complexity between the different problem formulations are directly connected to the assumptions on observability and communication (Seuken and Zilberstein 2008), motivating the development of approximate problem formulations that make assumptions on independence and communication between agents to reduce the computational complexity. These approximation algorithms exploit the structure of the domain to make intelligent decisions about which assumptions to make given the scenario at hand (independence, feature-based representations, etc.), leading to a trade-off between optimality and computational tractability. It is worth noting that these approximations are made in the problem formulation itself rather than in the solution algorithm employed. The most notable of these approximation algorithms is the transition-independent decentralized MDP (TI-Dec-MDP) (Becker 2004), which assumes that agents are independent collaborating entities connected through a global reward that is a function of all agents’ states and actions. The TI-Dec-MDP extends the factored Dec-MDP, where the global state can be broken up into local agent states plus some world state that represents the environment variables of interest and which is independent of agents’ states and actions (S D S0  S1      Sn , where Si is each agent’s local state space and S0 is the world state space). In this formulation, the transition and observation dynamics for each agent are independent (i.e., only functions of the local agent state and the world state S0  Si ), and the only coupling between agents is given by the joint reward function. These independence assumptions reduce the computational complexity of the Dec-MDP; however, although this approach scales to higher dimensions better than many others, it requires an extensive list of state transition histories per agent, referred to as events (Becker 2004; Redding et al. 2012). Furthermore, the independence assumption does not allow the representation to explicitly model global situations that are functions of the combined agents’ states (e.g., vehicle collisions), although these can be prevented by assigning very large negative rewards (which are joint), thus making collision avoidance part of the reward function instead of an explicit constraint. Some other approximation architectures include the decentralized sparse-interaction MDP (Dec-SIMDP) (Melo and Veloso 2011), which deals with computational limitations by separating the parts of the state space where agents need to cooperate from the parts where they can act independently; the group-aggregate Dec-MMDP (Redding et al. 2012), which uses features to compress other agents’ state-action spaces to highlight the properties of interest for each agent (e.g., how many agents are in an area of interest vs. where is every agent in the environment); and the auctioned POMDP (Capit´an et al. 2011), where each agent solves its own POMDP locally and communicates with other agents via an auction algorithm to achieve consensus on plans. Although these algorithms have demonstrated reduced computational complexity and real-time feasibility for large teams of cooperating agents, the approximation strategies involved are typically ad hoc and problem dependent, and developing good approximation strategies for cooperating agents remains an active area of research.

60 Cooperative Mission Planning for Multi-UAV Teams

60.5

1483

Conclusions and Future Directions

This chapter presents autonomous cooperative task allocation and planning methods for heterogeneous networked teams, providing an overview of the numerous approaches that have been considered in the literature. In particular, three standard planning frameworks are discussed: integer programming, Markov decision processes, and game theory. The chapter also outlines various architectural decisions that must be addressed when implementing online planning systems for multiagent teams, providing insights on when centralized, distributed, and decentralized architectures might be good choices for a given application, and how to organize the communication and computation to achieve the desired mission performance. Algorithms that can be utilized within the various architectures are identified, and the main technical challenges and future research directions are discussed. As mentioned throughout the various sections in this chapter, one of the major challenges associated with cooperative planning algorithms is maintaining computational tractability, especially for distributed and decentralized systems. Designing efficient approximation algorithms that can be used in real-time dynamic environments remains an active area of research in the integer programming, MDP, and game theory communities. Furthermore, designing algorithms that can handle message delays, dynamic networks, and unreliable communication links involves making trade-offs between performance and robustness which can vary depending on the scenario at hand. Many of the open issues in this field revolve around identifying problem-specific features that can be exploited to improve performance or reduce the computational effort. For example, submodularity in the score functions has recently been used to develop greedy algorithms that provide bounds on the suboptimality of the performance (Singh et al. 2009; Golovin and Krause 2010; Krause et al. 2006). The challenge in this case is ensuring that the problem formulation satisfies the submodularity (or adaptive submodularity) conditions. Another major consideration is that planning algorithms rely on underlying system models, which are usually approximations of the real systems and thus are often inaccurate. Discrepancies between these planner models and the actual system dynamics can cause significant degradations in mission performance. Thus, the planning systems must be efficiently integrated with online parameter identification and learning algorithms to adapt these models in real time (Redding et al. 2010). This process could also include elements of active learning, wherein part of the task assignment is specifically focused on deciding how best to reduce uncertainty to improve mission performance (Bertuccelli and How 2011). Robust planning and model learning, especially in decentralized and dynamic environments, remains an active area of research. Much of the literature to date has been devoted to developing algorithms that perform well in a single type of environment. Future algorithms should be able to exploit dynamic environmental conditions and adapt their solutions accordingly. For example, if the communication environment is particularly harsh, an algorithm should be robust enough to handle the situation. If conditions improve, however,

1484

S.S. Ponda et al.

the algorithm should be able to leverage the additional communication structure and information to improve performance. These hybrid adaptive planners would be able to make efficient use of whatever information is available, yet remain robust to worst-case environments. Finally, the algorithms presented in this chapter were discussed in the context of autonomous operations, but many future applications will require human operators on the loop (as opposed to being in the loop). In those scenarios, future research must develop techniques to efficiently allocate the roles and responsibilities of the operator and the autonomy/planning algorithms to ensure synergistic operations while performing the mission. Determining the objective function structure in the optimization to develop plans that are consistent with the human operator’s overall objectives is also typically quite challenging, especially for decentralized operations (Cummings et al. 2012), and thus is a promising area for future research. Acknowledgments This work was supported in part by the AFOSR and USAF under grant (FA9550-08-1-0086) and MURI (FA9550-08-1-0356). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Air Force Office of Scientific Research or the U.S. government.

References A. Ahmed, A. Patel, T. Brown, M. Ham, M. Jang, G. Agha, Task assignment for a physical agent team via a dynamic forward/reverse auction mechanism, in International Conference on Integration of Knowledge Intensive Multi-Agent Systems (IEEE, Piscataway, 2005) B. Alidaee, H. Wang, F. Landram, A note on integer programming formulations of the real-time optimal scheduling and flight path selection of UAVs. IEEE Trans. Control Syst. Technol. 17(4), 839–843 (2009) B. Alidaee, H. Wang, F. Landram, On the flexible demand assignment problems: case of unmanned aerial vehicles. IEEE Trans. Autom. Sci. Eng. 8(4), 865–868 (2011) M. Alighanbari, J.P. How, A robust approach to the UAV task assignment problem. Int. J. Robust Nonlinear Control 18(2), 118–134 (2008a) M. Alighanbari, J.P. How, An unbiased Kalman consensus algorithm. AIAA J. Aerosp. Comput. Inf. Commun. 5(9), 298–311 (2008b) G. Arslan, J. Marden, J. Shamma, Autonomous vehicle-target assignment: a game-theoretical formulation. J. Dyn. Syst. Meas. Control 129, 584 (2007) M.L. Atkinson, Results analysis of using free market auctions to distribute control of UAVs, in AIAA 3rd Unmanned Unlimited Technical Conference, Workshop and Exhibit (AIAA, Reston, 2004) A.G. Banerjee, M. Ono, N. Roy, B.C. Williams, Regression-based LP solver for chanceconstrained finite horizon optimal control with nonconvex constraints, in Proceedings of the American Control Conference (San Francisco, 2011) R. Beard, V. Stepanyan, Information consensus in distributed multiple vehicle coordinated control. IEEE Conf. Decis. Control 2, 2029–2034 (2003) R.W. Beard, T.W. McLain, M.A. Goodrich, E.P. Anderson, Coordinated target assignment and intercept for unmanned air vehicles. IEEE Trans. Robot. Autom. 18, 911–922 (2002) R. Becker, Solving transition independent decentralized Markov decision processes, in Computer Science Department Faculty Publication Series, 2004, pp. 208

60 Cooperative Mission Planning for Multi-UAV Teams

1485

J. Bellingham, A. Richards, J.P. How, Receding horizon control of autonomous aerial vehicles. Am. Control Conf. 5, 3741–3746 (2002) R. Bellman, Dynamic Programming (Dover, Mineola, 2003) A. Ben-Tal, A. Nemirovski, Robust convex optimization. Math. Oper. Res. 23(4), 769–805 (1998) D. Bernstein, R. Givan, N. Immerman, S. Zilberstein, The complexity of decentralized control of Markov decision processes, in Mathematics of operations research (2002), pp. 769–805. http:// dl.acm.org/citation.cfm?id=2073951 D.P. Bertsekas, The auction algorithm for assignment and other network flow problems, Technical report, MIT, 1989 D.P. Bertsekas, Dynamic Programming and Optimal Control, vol. I–II, 3rd edn. (Athena Scientific, Belmont, 2007) D.P. Bertsekas, J.N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods (Prentice-Hall, Englewood Cliffs, 1989) D. Bertsimas, D. Brown, Constructing uncertainty sets for robust linear optimization. Oper. Res. 57(6), 1483–1495 (2009) D. Bertsimas, R. Weismantel, Optimization over integers (Dynamic Ideas, Belmont, 2005) D. Bertsimas, D.B. Brown, C. Caramanis, Theory and applications of robust optimization. SIAM review. 53(3), 464–501 (2011) L. Bertuccelli, J. How, Active exploration in robust unmanned vehicle task assignment. J. Aerosp. Comput. Inf. Commun. 8, 250–268 (2011) L. Bertuccelli, H. Choi, P. Cho, J. How, Real-time multi-UAV task assignment in dynamic and uncertain environments, in AIAA Guidance, Navigation, and Control Conference (AIAA, Reston, 2009). (AIAA 2009–5776) B.M. Bethke, Kernel-based approximate dynamic programming using bellman residual elimination, Ph.D. thesis, Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, Cambridge, 2010 B. Bethke, J.P. How, J. Vian, Group health management of UAV teams with applications to persistent surveillance, in American Control Conference (ACC), Seattle (IEEE, New York, 2008), pp. 3145–3150 L. Blackmore, M. Ono, Convex chance constrained predictive control without sampling, in AIAA Proceedings. (np) (2009) V.D. Blondel, J.M. Hendrickx, A. Olshevsky, J.N. Tsitsiklis, Convergence in multiagent coordination, consensus, and flocking, in Proceeding of the IEEE Conference on Decision and Control (2005) S. Bradtke, A. Barto, Linear least-squares algorithms for temporal difference learning. J. Mach. Learn. Res. 22, 33–57 (1996) L. Bus¸oniu, R. Babuˇska, B. De Schutter, D. Ernst, Reinforcement Learning and Dynamic Programming Using Function Approximators (CRC, Boca Raton, 2010) J. Capit´an, M. Spaan, L. Merino, A. Ollero, Decentralized multi-robot cooperation with auctioned POMDPs, in Sixth Annual Workshop on Multiagent Sequential Decision Making in Uncertain Domains (MSDM-2011), 2011, p. 24 D. Castanon, J. Wohletz, Model predictive control for stochastic resource allocation. IEEE Trans. Autom. Control 54(8), 1739–1750 (2009) D.A. Castanon, C. Wu, Distributed algorithms for dynamic reassignment. IEEE Conf. Decis. Control 1, 13–18 (2003) P.R. Chandler, M. Pachter, D. Swaroop, J.M. Fowler, J.K. Howlett, S. Rasmussen, C. Schumacher, K. Nygard, Complexity in UAV cooperative control, in American Control Conference (ACC), Anchorage, 2002 A. Chapman, R. Micillo, R. Kota, N. Jennings, Decentralized dynamic task allocation using overlapping potential games. Comput. J. 53, 1462–1477 (2010) W. Chen, M. Sim, J. Sun, C. Teo, From CVaR to uncertainty set: implications in joint chance constrained optimization. Oper. Res. 58(2), 470–485 (2010) T. Chockalingam, S. Arunkumar, A randomized heuristics for the mapping problem: the genetic approach. Parallel Comput. 18(10), 1157–1165 (1992)

1486

S.S. Ponda et al.

H.-L. Choi, L. Brunet, J.P. How, Consensus-based decentralized auctions for robust task allocation. IEEE Trans. Robot. 25(4), 912–926 (2009) T. Cormen, Introduction to Algorithms (MIT, Cambridge, 2001) J. Cruz Jr., G. Chen, D. Li, X. Wang, Particle swarm optimization for resource allocation in UAV cooperative control, in AIAA Guidance, Navigation, and Control Conference and Exhibit, Providence (AIAA, Reston, 2004), pp. 1–11 M.L. Cummings, J.P. How, A. Whitten, O. Toupet, The impact of human–automation collaboration in decentralized multiple unmanned vehicle control. Proc. IEEE 100(3), 660–671 (2012) P. De Boer, D. Kroese, S. Mannor, R. Rubinstein, A tutorial on the cross-entropy method. Ann. Oper. Res. 134(1), 19–67 (2005) E. Delage, S. Mannor, Percentile optimization for Markov decision processes with parameter uncertainty. Oper. Res. 58(1), 203–213 (2010) M.B. Dias, A. Stentz, A free market architecture for distributed control of a multirobot system, in 6th International Conference on Intelligent Autonomous Systems IAS-6 (IOS, Amsterdam/Washington, DC, 2000), pp. 115–122 M.B. Dias, R. Zlot, N. Kalra, A. Stentz, Market-based multirobot coordination: a survey and analysis. Proc. IEEE 94(7), 1257–1270 (2006) Y. Eun, H. Bang, Cooperative task assignment/path planning of multiple unmanned aerial vehicles using genetic algorithms. J. Aircr. 46(1), 338 (2010) A.M. Farahmand, M. Ghavamzadeh, C. Szepesv´ari, S. Mannor, Regularized policy iteration, in Advances in Neural Information Processing Systems (NIPS), ed. by D. Koller, D. Schuurmans, Y. Bengio, L. Bottou (MIT, Cambridge, 2008), pp. 441–448 J. Fax, R. Murray, Information flow and cooperative control of vehicle formations. IEEE Trans. Autom. Control 49(9), 1465–1476 (2004) C.A. Floudas, Nonlinear and Mixed-Integer Programming – Fundamentals and Applications (Oxford University Press, New York, 1995) C.S.R. Fraser, L.F. Bertuccelli, J.P. How, Reaching consensus with imprecise probabilities over a network, in AIAA Guidance, Navigation, and Control Conference (GNC), Chicago, 2009. (AIAA-2009-5655) C.S. Fraser, L.F. Bertuccelli, H.-L. Choi, J.P. How, A hyperparameter consensus method for agreement under uncertainty. Automatica 48(2), 374–380 (2012) E.W. Frew, B. Argrow, Embedded reasoning for atmospheric science using unmanned aircraft systems, in AAAI 2010 Spring Symposium on Embedded Reasoning: Intelligence in Embedded Systems, Palo Alto (AAAI, Menlo Park, 2010) D. Fudenberg, J. Tirole, Game Theory (MIT, Cambridge, 1991) A. Gelman, J. Carlin, H. Stern, D. Rubin, Bayesian Data Analysis, 2nd edn. (Chapman and Hall, Boca Raton, 2004) A. Geramifard, F. Doshi, J. Redding, N. Roy, J. How, Online discovery of feature dependencies, in International Conference on Machine Learning (ICML), ed. by L. Getoor, T. Scheffer (ACM, New York, 2011), pp. 881–888 B. Gerkey, M. Mataric, Sold!: auction methods for multirobot coordination. IEEE Trans. Robot. Autom 18(5), 758–768 (2002) B.P. Gerkey, M.J. Mataric, A formal analysis and taxonomy of task allocation in multi-robot systems. Int. J. Robot. Res. 23(9), 939–954 (2004) F. Glover, R. Marti, Tabu search, in Metaheuristic Procedures for Training Neutral Networks (Springer, Boston, 2006), pp. 53–69 C. Goldman, S. Zilberstein, Optimizing information exchange in cooperative multi-agent systems, in Proceedings of the second international joint conference on Autonomous agents and multiagent systems (ACM, New York, 2003), pp. 137–144 C. Goldman, S. Zilberstein, Decentralized control of cooperative systems: categorization and complexity analysis. J. Artif. Intell. Res. 22, 143–174 (2004) D. Golovin, A. Krause, Adaptive submodularity: a new approach to active learning and stochastic optimization, Proceedings of the International Conference on Learning Theory (COLT), 2010

60 Cooperative Mission Planning for Multi-UAV Teams

1487

S. Grime, H. Durrant-Whyte, Data fusion in decentralized sensor networks. Control Eng. Pract. 2(5), 849–863 (1994) C. Guestrin, D. Koller, R. Parr, Multiagent planning with factored MDPs, in NIPS, ed. by T.G. Dietterich, S. Becker, Z. Ghahramani (MIT, Cambridge, 2001), pp. 1523–1530 Y. Hatano, M. Mesbahi, Agreement over random networks. IEEE Trans. Autom. Control 50(11), 1867–1872 (2005) ILOG, CPLEX (2006), http://www.ilog.com/products/cplex/ A. Jadbabaie, J. Lin, A.S. Morse, Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans. Autom. Control 48(6), 988–1001 (2003) L.B. Johnson, S.S. Ponda, H.-L. Choi, J.P. How, Asynchronous decentralized task allocation for dynamic environments, in Proceedings of the AIAA Infotech@Aerospace Conference, St. Louis (AIAA, Reston, 2011) L.B. Johnson, H.-L. Choi, S.S. Ponda, J.P. How, Allowing non-submodular score functions in distributed task allocation, in IEEE Conference on Decision and Control (CDC), 2012 (submitted) Y. Kim, D. Gu, I. Postlethwaite, Real-time optimal mission scheduling and flight path selection. IEEE Trans. Autom. Control 52(6), 1119–1123 (2007) E. King, Y. Kuwata, M. Alighanbari, L. Bertuccelli, J.P. How, Coordination and control experiments on a multi-vehicle testbed, in American Control Conference (ACC), Boston, (American Automatic Control Council, Evanston; IEEE, Piscataway, 2004), pp. 5315–5320 A. Krause, C. Guestrin, A. Gupta, J. Kleinberg, Near-optimal sensor placements: maximizing information while minimizing communication cost, in Information Processing in Sensor Networks, 2006. IPSN 2006. The Fifth International Conference on (ACM, New York, 2006), pp. 2–10, 0–0 M.G. Lagoudakis, R. Parr, Least-squares policy iteration. J. Mach. Learn. Res. 4, 1107–1149 (2003) G. Laporte, F. Semet, Classical heuristics for the capacitated VRP, in The Vehicle Routing Problem, ed. by P. Toth, D. Vigo (Society for Industrial Mathematics, Philadelphia, 2002) S. Leary, M. Deittert, J. Bookless, Constrained UAV mission planning: a comparison of approaches, in Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, Barcelona (IEEE, Piscataway, 2011), pp. 2002–2009 T. Lemaire, R. Alami, S. Lacroix, A distributed task allocation scheme in multi-UAV context. IEEE Int. Conf. Robot. Autom. 4, 3622–3627 (2004) J. Lin, A. Morse, B. Anderson, The multi-agent rendezvous problem. IEEE Conf. Decis. Control 2, 1508–1513 (2003) S. Mahadevan, M. Maggioni, C. Guestrin, Proto-value functions: a Laplacian framework for learning representation and control in Markov decision processes. J. Mach. Learn. Res. 8, 2007 (2006) A. Makarenko, H. Durrant-Whyte, Decentralized Bayesian algorithms for active sensor networks. Int. Conf. Inf. Fusion 7(4), 418–433 (2006) N.D. Manh, L.T.H. An, P.D. Tao, A cross-entropy method for nonlinear UAV task assignment problem, in IEEE International Conference on Computing and Communication Technologies, Research, Innovation, and Vision for the Future (RIVF) (IEEE, Piscataway, 2010), pp. 1–5 J. Marden, A. Wierman, Overcoming limitations of game-theoretic distributed control, in Joint 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference (IEEE, Piscataway, 2009) J. Marden, G. Arslan, J. Shamma, Cooperative control and potential games. IEEE Trans. Syst. Man Cybern. Part B Cybern. 39(6), 1393–1407 (2009) M.J. Mataric, G.S. Sukhatme, E.H. Ostergaard, Multi-robot task allocation in uncertain environments. Auton. Robots 14(2–3), 255–263 (2003) I. Maza, F. Caballero, J. Capit´an, J. Mart´ınez-de Dios, A. Ollero, Experimental results in multiUAV coordination for disaster management and civil security applications. J. Intell. Robot. Syst. 61(1), 563–585 (2011)

1488

S.S. Ponda et al.

T.W. McLain, R.W. Beard, Coordination variables, coordination functions, and cooperative-timing missions. J. Guid. Control Dyn. 28(1), 150–161 (2005) F.S. Melo, M. Veloso, Decentralized MDPs with sparse interactions. Artif. Intell. 175, 1757–1789 (2011) C.C. Moallemi, B.V. Roy, Consensus propagation. IEEE Trans. Inf. Theory 52(11), 4753–4766 (2006) D. Monderer, L. Shapley, Potential games. Games Econ. Behav. 14, 124–143 (1996) R. Murphey, Target-based weapon target assignment problems. Nonlinear Assign. Probl. Algorithms Appl. 7, 39–53 (1999) A. Nemirovski, A. Shapiro, Convex approximations of chance constrained programs. SIAM J. Optim. 17(4), 969–996 (2007) I. Nikolos, E. Zografos, A. Brintaki, UAV path planning using evolutionary algorithms, in Innovations in Intelligent Machines-1 (Springer, Berlin/New York, 2007), pp. 77–111 Office of the Secretary of Defense, Unmanned aircraft systems roadmap, Technical report, OSD (2007), http://www.acq.osd.mil/usd/UnmannedSystemsRoadmap.2007-2032.pdf R. Olfati-saber, Distributed Kalman filtering and sensor fusion in sensor networks, in Network Embedded Sensing and Control, vol. 331 (Springer, Berlin, 2006), pp. 157–167 R. Olfati-Saber, R.M. Murray, Consensus problems in networks of agents with switching topology and time-delays. IEEE Trans. Autom. Control 49(9), 1520–1533 (2004) R. Olfati-Saber, A. Fax, R.M. Murray, Consensus and cooperation in networked multi-agent systems. IEEE Trans. Autom. Control 95(1), 215–233 (2007) A. Olshevsky, J.N. Tsitsiklis, Convergence speed in distributed consensus and averaging, in IEEE Conference on Decision and Control (CDC) (IEEE, Piscataway, 2006), pp. 3387–3392 C. Papadimitriou, Computational Complexity (Wiley, Chichester, 2003) C.H. Papadimitriou, J.N. Tsitsiklis, The complexity of Markov decision processes. Math. Oper. Res. 12(3), 441–450 (1987) S. Paquet, L. Tobin, B. Chaib-draa, Real-time decision making for large POMDPs. Adv. Artif. Intell. 3501, 450–455 (2005) K. Passino, M. Polycarpou, D. Jacques, M. Pachter, Y. Liu, Y. Yang, M. Flint, M. Baum, Cooperative control for autonomous air vehicles, in Cooperative control and optimization (Kluwer, Dordrecht/Boston, 2002), pp. 233–271 S.S. Ponda, J. Redding, H.-L. Choi, J.P. How, M.A. Vavrina, J. Vian, Decentralized planning for complex missions with dynamic communication constraints, in American Control Conference (ACC), Baltimore, 2010 S.S. Ponda, L.B. Johnson, H.-L. Choi, J.P. How, Ensuring network connectivity for decentralized planning in dynamic environments, in Proceedings of the AIAA Infotech@Aerospace Conference, St. Louis (AIAA, Reston, 2011) S.S. Ponda, L.B. Johnson, J.P. How, Distributed chance-constrained task allocation for autonomous multi-agent teams, in American Control Conference (ACC), 2012 A. Pongpunwattana, R. Rysdyk, J. Vagners, D. Rathbun, Market-based co-evolution planning for multiple autonomous vehicles, in Proceedings of the AIAA Unmanned Unlimited Conference, San Diego (AIAA, Reston, 2003) D. Pynadath, M. Tambe, The communicative multiagent team decision problem: analyzing teamwork theories and models. J. Artif. Intell. Res. 16(1), 389–423 (2002) S. Rathinam, R. Sengupta, S. Darbha, A resource allocation algorithm for multivehicle systems with nonholonomic constraints. IEEE Trans. Autom. Sci. Eng. 4(1), 98–104 (2007) J. Redding, A. Geramifard, A. Undurti, H. Choi, J. How, An intelligent cooperative control architecture, in American Control Conference (ACC), Baltimore, 2010, pp. 57–62 J.D. Redding, N.K. Ure, J.P. How, M. Vavrina, J. Vian, Scalable, MDP-based planning for multiple, cooperating agents, in American Control Conference (ACC) (2012, to appear) W. Ren, Consensus based formation control strategies for multi-vehicle systems, in American Control Conference (ACC) (American Automatic Control Council, Evanston; IEEE, Piscataway, 2006), pp. 6–12

60 Cooperative Mission Planning for Multi-UAV Teams

1489

W. Ren, R. Beard, Consensus seeking in multiagent systems under dynamically changing interaction topologies. IEEE Trans. Autom. Control 50(5), 655–661 (2005) W. Ren, R.W. Beard, D.B. Kingston, Multi-agent Kalman consensus with relative uncertainty. Am. Control Conf. 3, 1865–1870 (2005) W. Ren, R.W. Beard, E.M. Atkins, Information consensus in multivehicle cooperative control. IEEE Control Syst. Mag. 27(2), 71–82 (2007) A. Richards, J. Bellingham, M. Tillerson, J.P. How, Coordination and control of multiple UAVs, in AIAA Guidance, Navigation, and Control Conference (GNC), Monterey (AIAA, Reston, 2002). AIAA Paper 2002–4588 G.A. Rummery, M. Niranjan, Online Q-learning using connectionist systems (Technical Report No. CUED/F-INFENG/TR 166), Cambridge University Engineering Department (1994) R.O. Saber, W.B. Dunbar, R.M. Murray, Cooperative control of multi-vehicle systems using cost graphs and optimization, in Proceedings of the 2003 American Control Conference, 2003, vol. 3 (IEEE, Piscataway, 2003), pp. 2217–2222 A. Salman, I. Ahmad, S. Al-Madani, Particle swarm optimization for task assignment problem. Microprocess. Microsyst. 26(8), 363–371 (2002) S. Sariel, T. Balch, Real time auction based allocation of tasks for multi-robot exploration problem in dynamic environments, in AIAA Workshop on Integrating Planning Into Scheduling (AAAI, Menlo Park, 2005) K. Savla, E. Frazzoli, F. Bullo, On the point-to-point and traveling salesperson problems for Dubins’ vehicle, in American Control Conference (ACC), June 2005. pp. 786–791 B. Scherrer, Should one compute the temporal difference fix point or minimize the Bellman Residual? The unified oblique projection view, International Conference on Machine Learning (ICML) (IEEE, Los Alamitos, 2010) D.G. Schmale, B. Dingus, C. Reinholtz, Development and application of an autonomous unmanned aerial vehicle for precise aerobiological sampling above agricultural fields. J. Field Robot. 25(3), 133–147 (2008) C. Schumacher, P.R. Chandler, S. Rasmussen, Task allocation for wide area search munitions via network flow optimization, in Proceedings of the American Control Conference, Anchorage, 2002, pp. 1917–1922 S. Seuken, S. Zilberstein, Formal models and algorithms for decentralized decision making under uncertainty. Auton. Agents Multi-Agent Syst. 17(2), 190–250 (2008) T. Shima, S.J. Rasmussen, UAV Cooperative Decision and Control: Challenges and Practical Approaches, vol. 18 (Society for Industrial Mathematics, Philadelphia, 2009) T. Shima, S. Rasmussen, A. Sparks, K. Passino, Multiple task assignments for cooperating uninhabited aerial vehicles using genetic algorithms. Comput. Oper. Res. 33(11), 3252–3269 (2006) A. Singh, A. Krause, W. Kaiser, Nonmyopic adaptive informative path planning for multiple robots, in International Joint Conference on Artificial Intelligence (IJCAI) (AAAI, Menlo Park, 2009) R.G. Smith, R. Davis, Frameworks for cooperation in distributed problem solving. IEEE Trans. Syst. Man Cybern. 11(1), 61–70 (1981) M.T.J. Spaan, N. Vlassis, Perseus: randomized point-based value iteration for POMDPs. Int. J. Robot. Res. 24, 195–220 (2005) P. Stone, R.S. Sutton, G. Kuhlmann, Reinforcement learning for RoboCup-Soccer keepaway. Int. Soc. Adapt. Behav. 13(3), 165–188 (2005) P.B. Sujit, D. Kingston, R. Beard, Cooperative forest fire monitoring using multiple UAVs, in IEEE Conference on Decision and Control, New Orleans (IEEE, Piscataway, 2007), pp. 4875–4880 R.S. Sutton, Generalization in reinforcement learning: successful examples using sparse coarse coding, in Advances in Neural Information Processing Systems 8 (MIT, Cambridge/London, 1996), pp. 1038–1044 R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction (MIT, Cambridge, 1998) R.S. Sutton, H.R. Maei, D. Precup, S. Bhatnagar, D. Silver, C. Szepesv´ari, E. Wiewiora, Fast gradient-descent methods for temporal-difference learning with linear function approximation,

1490

S.S. Ponda et al.

in International Conference on Machine Learning (ICML), ICML ’09 (ACM, New York, 2009), pp. 993–1000 A. Tahbaz-Salehi, A. Jadbabaie, On consensus over random networks, in 44th Annual Allerton Conference, 2006 P. Toth, D. Vigo, The Vehicle Routing Problem (Society for Industrial and Applied Mathematics, Philadelphia, 2001) J.N. Tsitsiklis, B.V. Roy, An analysis of temporal difference learning with function approximation. IEEE Trans. Autom. Control 42(5), 674–690 (1997) K. Tumer, D. Wolpert, A survey of collectives, in Collectives and the Design of Complex Systems (Springer, New York, 2004), pp. 1–42 A. Undurti, J.P. How, A Cross-entropy based approach for UAV task allocation with nonlinear reward, in AIAA Guidance, Navigation, and Control Conference (GNC) (AIAA, Reston, 2010). AIAA-2010-7731 U.S. Air Force Chief Scientist (AF/ST), Technology horizons: a vision for air force science & technology during 2010–2030, Technical report, United States Air Force (2010) U.S. Army UAS Center of Excellence, Eyes of the Army: U.S. Army unmanned aircraft systems roadmap 2010–2035, Technical report (2010), http://www.fas.org/irp/program/collect/ uas-army.pdf M. Valenti, B. Bethke, J.P. How, D.P. de Farias, J. Vian, Embedding health management into mission tasking for UAV teams, in American Control Conference (ACC), New York (IEEE, New York, 2007), pp. 5777–5783 E. Waltz, J. Llinas, Multisensor Data Fusion (Artech House, Boston/London, 1990) R.V. Welch, G.O. Edmonds, Applying robotics to HAZMAT, in The Fourth National Technology Transfer Conference and Exposition, vol. 2 (2003), pp. 279–287 A.K. Whitten, H.-L. Choi, L. Johnson, J.P. How, Decentralized task allocation with coupled constraints in complex missions, in American Control Conference (ACC), 2011, pp. 1642–1649 C.W. Wu, Synchronization and convergence of linear dynamics in random directed networks. IEEE Trans. Autom. Control 51(7), 1207–1210 (2006) L. Xiao, S. Boyd, S. Lall, A scheme for robust distributed sensor fusion based on average consensus, in International Symposium on Information Processing in Sensor Networks (ACM, New York, 2005), pp. 63–70 R. Zhou, E.A. Hansen, An improved grid-based approximation algorithm for POMDPs, in International Joint Conference on Artificial Intelligence, vol. 17, number 1 (Morgan Kaufmann, San Francisco, 2001), pp. 707–716

Multi-team Consensus Bundle Algorithm

61

Matthew E. Argyle, Randal W. Beard, and David W. Casbeer

Contents 61.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1492 61.2 Task Assignment Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1494 61.3 Multi-team Task Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1495 61.3.1 Local Task Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1495 61.3.2 Task Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1496 61.3.3 Phase 1: Build Bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1499 61.3.4 Phase 2: Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1500 61.4 Convergence Time Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1501 61.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1503 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1506

Abstract

In this chapter, the consensus-based bundle algorithm (CBBA) is incorporated into a hierarchical concept of operation. In the team CBBA, agents are divided into teams and each team plans for its agents to service a set of tasks. This team planning is carried out in a distributed manner using the traditional CBBA. An “outer-loop” team CBBA strategy is presented that coordinates teams’ plans. The hierarchical structure of the team CBBA facilitates a manageable architecture for large numbers of unmanned agents through

M. Argyle () Department of Electrical Engineering, Brigham Young University, Provo, UT, USA e-mail: [email protected] R. Beard Electrical and Computer Engineering Department, Brigham Young University, Provo, UT, USA e-mail: [email protected] D.W. Casbeer Control Science Center of Excellence, Air Force Research Laboratory, WPAFB, OH, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 17, © Springer Science+Business Media Dordrecht 2015

1491

1492

M. Argyle et al.

human-centered operations. This is because each team is managed by a human operator with the team CBBA aiding coordination between teams.

61.1

Introduction

Currently deployed unmanned air vehicles (UAV) are individually controlled by a group of operators that actively manage different components of the UAV. One operator might control the motion of the vehicle, either directly through “stick-and-rudder” inputs or through a waypoint navigation mode. Simultaneously, different groups of operators remotely control the onboard sensors and evaluate any received data. Recent research has been focused on inverting this ratio of operators per UAV to have a single operator manage a team of UAVs (Cummings and Bruni 2009; Cummings et al. 2012). Planning algorithms have been developed to aid the operator with the management of a team of UAVs and current research is improving on these algorithms (Shima and Rasmussen 2009; Rasmussen and Shima 2008; Choi et al. 2009). To date, planners are generally implemented centrally, at a base station. Four main factors motivate this centralized paradigm: 1. Poor inter-UAV communication 2. Inadequate processing capability (applies to the computation and algorithms needed for both planning and sensor data processing) 3. Operators’ desire for a continuous stream of video and images 4. Pilot safety in airspace with both unmanned and manned aircraft First, trustworthy sense-and-avoid capability will increase safety in airspace with piloted and unmanned aircraft. However, until such capabilities are operational, a human operator is necessary to ensure safe operation in mixed environments, which leads to an architecture with planning performed centrally. Second, in military applications, commanders want as much information as possible to facilitate informed decisions, hence the desire for continual video and image footage. If a system is designed to feed high-bandwidth video data to a central human-processor for evaluation, then it also makes sense to plan centrally since communicating C2 data causes negligible change to bandwidth requirements when compared to the transmitted video and images. The continual video and image feeds are, also, necessary because of the UAV’s inadequate ability to autonomously evaluate sensor data and derive actionable intelligence, e.g., detection and classification (factor 2). Lastly, most UAVs do not possess the onboard processing capability to perform the complex calculations necessary for combinatorial optimization (factor 2), and there is not dependable communication between the UAVs, making a distributed planner impractical since the UAVs cannot reliably coordinate their actions. Decentralized control of UAVs, where the UAVs operate independently from human monitors, is a long-term goal, and in order to accomplish this, the four factors mentioned above will need to be resolved. In the short term, human operators will need to be closely involved with the UAV operation. As more and more UAVs and UAV teams are deployed simultaneously, one could naively allow separate UAV teams to function independently, where each team is monitored by an operator with

61 Multi-team Consensus Bundle Algorithm

1493

his/her own tasks to accomplish. However, a logical step would be to allow these separate teams to share resources. Ideally, the mission, as a whole, could be realized more efficiently when compared with independent operation. Thus, it is timely to consider alternative management schemes. More specifically, a system-of-systems management architecture needs to be considered that is capable of dynamic distributed task assignment across multiple autonomous vehicles, ranging from small to large. This requires a new distributed management architecture that can coordinate a large number of intelligent agents accomplishing a large number of tasks in time-critical, dynamically changing environments. Moreover, the management approach must still retain human operators, given mission objectives can vary constantly as does the ground situation. The operators’ involvement would be supervisory, with attention focused at the mission execution level. Yet, the operators must also be kept informed of the autonomous vehicles’ states and task progress, in order to benefit from the human’s unique ability to provide critical intuition and judgment for the execution of successful collaborative missions. Current stick-and-rudder UAV operation is a conservative first step toward fully autonomous aircraft. It allows a human to make decisions for the UAV, avoiding the problems associated with factors 2 and 4 listed above. The first step toward achieving functional autonomous UAVs is to reduce the workload on the UAV operator(s) by allowing the UAV system to takeover duties that current technology allows a machine to perform robustly. The first step for small UAVs was to allow an onboard autopilot to take over the low-level control of the UAV, ensuring stable flight by moving operator control to a higher level of control, e.g., waypoint navigation and heading/altitude commands (Chao et al. 2010). Autopilots generally have different modes of operation: one mode for direct remote control (RC) and a seperate autopilot mode. Such a system could loosely be classified as an adjustably autonomous system or a system that allows the operator to take direct control of an aircraft but also allows the operator to step back and manage the UAV(s) or even simply monitor the UAV(s). Adjustable autonomy, sometimes called incremental autonomy or variable autonomy (see Bonasso et al. 1997; Pell et al. 1998), refers to a system with the ability to continuously change the level of autonomy between full autonomous and manual modes during system operation (Dorais et al. 1998). Adjustable autonomy is different from increment assistance in that an autonomous system is able to function on its own while an “assistance system” solely aids the operator (Baudin et al. 1994). The idea of adjustable autonomy really goes much farther than the basic modes currently used by autopilots today. Adjustable autonomy really has its roots in the artificial intelligence community, the autonomous system can range from manual control to a fully autonomous mode (Scerri et al. 2002, 2003b). In the UAV setting, this implies the UAV is able to observe itself and its surroundings, place itself, in context, within those surroundings, and in pursuit of its mission’s objective make appropriate decisions on actions that effect itself and its surroundings. To further reduce the operator workload, a team paradigm is needed, which reduces the man-hours needed to operate multiple UAVs (Christoffersen and Woods 2004; Sycara and Sukthankar 2006). In the human factors arena, there has been research performed to investigate the effects of managing a team of robots on the

1494

M. Argyle et al.

human operator (Kilgore et al. 2007; Lewis et al. 2010; Crandall et al. 2011; Sycara and Lewis 2004). In order for the robots to function independently as part of a team, the agents must be capable of determining teammates intent, meaning the intent of both robot teammates, as well as human teammates (Grosz and Kraus 2003; Scerri et al. 2003a; Sycara et al. 2006, 2003). However, due to the large scope of this problem, there is a lot of research still needed to make mixed human-robot teams practical. In this chapter, an overview is given of the team consensus-based bundle algorithm (TCBBA) that was presented in Argyle and Casbeer (2011). TCBBA is a hierarchical and distributed task assignment algorithm based on the paradigm that human operators are responsible for their assigned UAVs. Rather than having humans input tasks to a swarm/cloud of agents and then the autonomy deciding which agent will do which task, here the agents are fixed to an operator and the tasks are shifted between teams. This algorithm follows current rules of engagement and is a novel distributed task assignment algorithm, in that it places end responsibility for the welfare of the agents and the success or failure of the mission directly on the human operator. In the TCBBA, each team is composed of heterogeneous UAVs, and a human operator is responsible for the UAVs assigned to his/her team. Each team of UAVs is given tasks that must be accomplished. These tasks are allotted to specific agents in the team using a distributed algorithm called the CBBA (Choi et al. 2009). The teams try to assign all the tasks while maximizing the total score of some cost function. If there are tasks that are not assigned, a second stage starts. In the second stage, the teams offer and request help to and from neighboring teams. This second stage of bidding consists of a higher-level auction algorithm running “outside” a lower level CBBA auction. The format of this chapter is as follows: Sect. 61.2 lays out the specific problem that is addressed by the TCBBA. Section 61.3 discusses the multi-team extension to the CBBA, which is followed by analysis in Sect. 61.4. Lastly, simulations results and conclusions are presented in Sects. 61.5 and 61.6.

61.2

Task Assignment Problem

Consider the multi-agent multitask assignment problem where a group of Na agents attempt to service Nt tasks while trying to maximize a reward. This can be stated as:

maximize

Nt Na X X

cj k .xj ; pj /xj k

j D1 kD1

subject to

Na X

xj k  1 8k 2 f1; : : : ; Nt g

j D1 Nt X kD1

xj k  Lm

8j 2 f1; : : : ; Na g;

(61.1)

61 Multi-team Consensus Bundle Algorithm

1495

where xj k 2 f0; 1g is one if agent j services task k and zero otherwise, xj 2 f0; 1gNt is a vector whose kth element is xj k , pj is the path for agent j and indicates the order agent j will service its assigned tasks, cj k ./ is the non-negative reward, and Lm is the maximum number of tasks per agent. Due to the curse of dimensionality, an exact solution to the multi-agent multitask problem becomes impractical for large numbers of agents (Bellman 1957). To overcome this problem, current operations have multiple operators controlling either one UAV or possibly a group of UAVs. Each group of operators would service tasks independently. In this setup, a human remains responsible for his or her team of agents. The teams interact, with other teams, through higher-level coordination. Here, the Nt tasks and Na agents are divided between No operators. Each team i 2 f1; : : : ; No g tries to service all of its assigned tasks Nt i  f1; : : : ; Nt g with its assigned agents Nui  f1; : : : ; Na g. The team assignment problem can be stated as:

maximize

X X

cj k .xj ; pj /xj k

j 2Nui k2Nt i

subject to

X

xj k  1 8k 2 Nt i

j 2Nui

X

xj k  Lm

(61.2)

8j 2 Nui :

j 2Nt i

A problem occurs when a team is unable to perform all of its assigned tasks due to time constraints, resource constraints, or agent capabilities. The TCBBA is a hierarchical task assignment method, that allows neighboring teams to request and provide help when this situation occurs.

61.3

Multi-team Task Assignment

The team consensus-based bundle algorithm (TCBBA) consists of two main stages: local task assignment and task sharing. In the first stage, each team implements the traditional CBBA to create an initial task assignment. If there are tasks that were unable to be assigned, the teams attempt to correct this by implementing the CBBA between the teams.

61.3.1 Local Task Assignment To create the initial task assignment, each team runs the CBBA introduced in Choi et al. (2009). This description of the CBBA follows the explanation given in Choi et al. (2010). The CBBA is a decentralized task assignment algorithm which

1496

M. Argyle et al.

consists of iterations between two phases: a bundle-building phase where each agent generates an ordered bundle of tasks and a consensus phase where conflicting assignments are resolved through communication with nearby agents.

61.3.1.1 Phase 1 (Build Bundle) The first phase of the CBBA is bundle construction, during which each agent adds tasks to its bundle until it is unable to add any more tasks. Each agent j in team i maintains four data vectors: a bundle bij 2 .Nt i [ f;g/Lm , the corresponding path ti pij 2 .Nt i [ f;g/Lm , a winning bid list yij 2 yij k /  hij k where ./ is one if the argument is true and zero otherwise, and hij k is one if agent j can perform task k and zero otherwise. The path is updated by pij D pij ˚nijKij fKij g   p ˚ fK g where nijKij D argmaxn Sijij n ij . 61.3.1.2 Phase 2 (Conflict Resolution) In the conflict resolution phase, three data vectors are transmitted to neighboring agents. Two were used in the bundle construction phase: the winning bid list yij and the winning agent list zij . The third data vector sij 2 Si or yOj > yOi

if Sj > Si Sj > Si; and yOj > yOi if Sj > Si else if Sj > Si if Sj > Si and Sj > Si if Sj > Si and yOj > yOi if Sj > Si and Si > Sj

if Sj > Si

UPDATE UPDATE UPDATE LEAVE RESET RESET UPDATE UPDATE RESET UPDATE UPDATE UPDATE RESET LEAVE UPDATE UPDATE

61 Multi-team Consensus Bundle Algorithm

1501

There are three possible outcomes from the decision table (A4 : 6): update, reset, or leave. An update for team i means team i ’s information is changed to what team h knows (A4 : 8). A reset re-initializes the bid and winner ID to zero (A4 : 10), and the task is removed from the bundle (A4 : 15–16). A leave makes no change to any of the data vectors. If an update or a reset occurred on task k, then the reward values for the tasks that appear in the bundle after task k are no longer valid and all of the following tasks in the bundle need to be reset (A4 : 15–16). The time-stamp vector is then updated using the current time and Sj (A4 : 21–27). After the team information Bi ; Yi , and Zi has been updated, the individual agent’s information needs to be updated (A4 : 26–28). Each agent j 2 Nui goes through its entire bundle bij looking for any tasks that were removed from the team’s bundle (A5 : 1–2). If it finds one then, for that task and each task that follows it in the bundle, the bid and the winner ID are reset to zero (A5 : 3) and the tasks are removed from the bundle and the path (A5 : 4–5).

61.4

Convergence Time Analysis

It is informative to determine the maximum number of communication steps required for all teams to arrive at a conflict-free solution in a static communication network and worst case scenario. Such analysis cannot be made on Algorithm 2 since it requires knowledge of the number of tasks assigned by each team during each iteration. In the worst case situation, the diameter of the communication network equals the number of agents. Let Di be the diameter of the communication network of the agents in team i , Ni be the number of tasks assigned to team i , DT be the diameter of the team communication network, Nu be the total number of unassigned tasks after the local task assignment phase, N be the total number of tasks, and D be the diameter of the communication network of all the agents. Choi et al. (2009) showed that the maximum number of communication steps needed for the CBBA to converge to a conflict-free solution is ND. Applying this result to the TCBBA, the number of communications steps needed in the worst case for the initial task assignment is maxi .Ni Di /. The second stage will have at most Nu DT iterations, and each iteration will have maxi Nu Di communication steps between agents as well as one communication step between teams. The total number of communication steps is   Nc D max.Ni Di / C DT Nu max.Nu Di / C 1 : i

i

(61.3)

In the worst case, maxi .Nu Di /  1 so Nc  max.Ni Di / C DT Nu2 max.Di /: i

i

(61.4)

1502

M. Argyle et al.

Assuming that the agents and tasks are divided up equally among the No teams then Ni D

N No

and (61.4) becomes Nc D

and Di D

D ; No

ND DT Nu2 D C : No2 No

(61.5)

Assuming worst case communication between the teams then No D DT and (61.5) becomes   ND N 2 2 (61.6) C N D D C N u u D: No2 No2 Comparing (61.6) to the worst case convergence bound for the full CBBA assignment ND gives   N 2 (61.7) C N u D  ND; No2 and solving (61.7) for Nu gives the following: s Nu 

1

1 p N: No2

(61.8)

Notice as the number of teams increases, the number of allowed unassigned tasks, p Nu approaches N . Given the assumption of full communication between the teams in lieu of worst case communication then DT D 1 and (61.5) becomes ND Nu2 D C D No2 No



Nu2 N C No2 No

 D:

(61.9)

Comparing (61.9) to the worst case convergence bound for the full CBBA assignment ND yields   N Nu2 D  ND; (61.10) C No2 No where, as with the analysis of above, (61.10) is solved for Nu to obtain s Nu 

No 

1 p N: No

(61.11)

Figure 61.1 shows how the maximum number of unassigned tasks change as the number of teams increase while the number of tasks and agents stay the same. Notice in the worst case team communication case, the maximum number of

61 Multi-team Consensus Bundle Algorithm 45 Worst Case Team Communication Full Team Communication

40

N1/2

35 Unallocated Tasks

Fig. 61.1 Maximum number of unassigned tasks for TCBBA to converge quicker than CBBA in worst case as the number of teams changes (Nt D 150; D D 12)

1503

30 25 20 15 10 5 0 0

2

4

6

8

10

12

Number of Teams

Worst Case Team Communication Full Team Communication N1/2

Unallocated Tasks

20

Fig. 61.2 Maximum number of unassigned tasks for TCBBA to converge quicker than CBBA in worst case as the number of tasks changes (No D 4; D D 12)

15

10

5

0

0

50 100 Number of Tasks

150

p unassigned tasks quickly approaches the maximum value of Nt . In the p full team communication case, the number of unassigned tasks quickly exceeds Nt . Figure 61.2 shows how the maximum number of unassigned tasks changes as the number of tasks increase while the number of teams and agents stay the same. Notice that both the worst p case team communication and full team communication increase proportionally to Nt .

61.5

Simulation Results

Four scenarios are used to compare the TCBBA to the CBBA. In each scenario there are three teams of four agents within 1;200  400 m world. Each team is assigned Nu tasks which take 25 s to complete. There are four types of tasks

1504

M. Argyle et al.

fA; B; C; Dg and four types of agents fa; b; c; d g. Each agent type is only able to service its corresponding task type and is modeled as a point-mass with a constant speed of jvj D 40 m/s and no turning constraints, i.e., the agent’s position is given by rP D v; where the direction of v points toward the next task to be visited. The timediscounted reward for the tasks is p

Si i D p

X

j

 .pi /

ji

cj ;

(61.12)

where Si i is the total score of agent i with path pi , j D 0:001 is the discounting j factor for task j , i .pi / is the estimated time that agent i will service task j when following path pi , and cj D 100 is the value of the task. The objective is the maximize the total score over 500 s. The four scenarios are created by changing two parameters. The first parameter is if the teams have their own region of the map or if they are in the same overlapping region. If the teams are not overlapping, then they each have their own 400  400 m region with team 2 located in the center. If the teams are overlapping then the teams’ agents and tasks are distributed randomly throughout the entire world. The second parameter is if there are 50 or 100 tasks per team. Teams 1 and 3 do not include all types of agents, so they will be forced to cooperate with other teams in order to accomplish their tasks. Team 1 has two type a agents and two type b agents, team 2 has one of each type of agent, and team 3 has two type c agents and two type d agents. Each team is assigned an equal number of each task type which are randomly distributed within the team’s assigned region. Providing an incomplete set of agents to teams 1 and 3 guarantees there will be unassigned tasks after the initial CBBA, it also guarantees that the proportion of unassigned tasks to total tasks will be higher than the threshold developed in the previous section. This almost guarantees that the TCBBA will take more communication steps than the CBBA. Five hundred Monte-Carlo simulations are run for each scenario. In each simulation, the tasks are assigned using five different methods: running Algorithm 1 once, running Algorithm 2 once, running Algorithm 2 until the change in score is less than 2:5 %, running Algorithm 2 until either all the tasks have been assigned or the unassigned tasks list stops changing, and merging all the tasks and agents into one team and running the CBBA. Figure 61.3 shows the mean percent difference in total score between the four TCBBA versions and the CBBA assignment. Figure 61.4 shows the mean percent difference in communication steps needed to arrive at a conflict free solution between the TCBBA and the CBBA. Figure 61.5 shows the mean percent difference in computation time between the TCBBA and CBBA. The results show several interesting things about the TCBBA. First, as seen in Fig. 61.3, all the variations of the TCBBA consistently have a lower overall score than the CBBA. This difference varies based on the which TCBBA variation is used

61 Multi-team Consensus Bundle Algorithm

1505

0% −1% −2% −3% −4% −5% −6%

Insert

−7%

2.50%

Once Complete

−8% −9% No Overlap 50 Tasks/Team

No Overlap 100 Overlap 50 Tasks/ Tasks/Team Team

Overlap 100 Tasks/Team

Fig. 61.3 Mean percent difference in overall score between TCBBA and CBBA 900% Insert 800% 700%

Once 2.50% Complete

600% 500% 400% 300% 200% 100% 0% No overlap 50 Tasks/Team

No overlap 100 Tasks/Team

Overlap 50 Tasks/Team

Overlap 100 Tasks/Team

Fig. 61.4 Mean percent difference in communication steps between TCBBA and CBBA

but is typically within 3 %. What is not shown is that occasionally the TCBBA can do as well as the CBBA but in all 500 runs, it never beat it. Second, the TCBBA always required more communication steps than the CBBA as shown in Fig. 61.4. This result was expected because of the large number of unassigned tasks. Third, there is a point of diminishing returns, such that repeating the second variation of TCBBA until the unassigned task list is empty or unchanging is, generally, not worth

1506

M. Argyle et al. 300%

250%

Insert Once 2.50%

200%

Complete

150%

100%

50%

0% −50% No Overlap 50 Tasks/Team

No Overlap 100 Tasks/Team

Overlap 50 Tasks/Team

Overlap 100 Tasks/Team

Fig. 61.5 Mean percent difference in computation time between TCBBA and CBBA

the computation. Finally, the amount of computation time required is proportional to the number of communication steps as seen in Fig. 61.5. Because the larger problem is partitioned into a set of smaller problems in the TCBBA, it is interesting to note that the TCBBA can take less time to arrive at a solution than the CBBA when implemented in a centralized manner. Conclusion

This chapter presents a modification of the consensus-based bundle algorithm that handles the multi-team, multitask problem using a combination of auctions and consensus to achieve feasible, conflict free solutions. While these solutions are less optimal than running the CBBA, typically within 3 %, the TCBBA allows the agents to be divided into separate teams, where, following current rules of engagement, responsibility for each agent/team can be kept under the supervision of a human supervisor.

References M. Argyle, D.W. Casbeer, R. Beard, Proceedings of the American Control Conference, San Francisco, CA (IEEE, New York, 2011), pp. 5376–5381 C. Baudin, S. Kedar, B. Pell, Knowl. Aquis. 6(2), 179–196 (1994) R.E. Bellman, Dynamic Programming (Princeton University Press, Princeton, 1957) R. Bonasso, D. Kortenkamp, T. Whitney, Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative Applications of Artificial Intelligence, Providence, Rhode Island (AAAI MIT, Palo Alto/Cambridge, 1997), pp. 949–956 H. Chao, Y. Cao, Y. Chen, Int. J. Control Autom. Syst. 8(1), 36–44 (2010)

61 Multi-team Consensus Bundle Algorithm

1507

H.L. Choi, L. Brunet, J. How, IEEE Trans. Robot. 25(4), 912–926 (2009) H.L. Choi, A.K. Whitten, J. How, Proceedings of the American Control Conference, Baltimore, MD (IEEE, New York, 2010), pp. 3057–3062 K. Christoffersen, D.D. Woods, Advances in Human Performance and Cognitive Engineering Research, vol. 2 (Emerald Group Publishing Limited, Bingley, 2004), pp. 1–12 J.W. Crandall, M.L. Cummings, M.D. Penna, P.M.A. de Jong, IEEE Trans. Syst. Man Cybern. 41(2), 385–397 (2011) M.L. Cummings, S. Bruni, in Handbook of Automation, ed. by S.Y. Nof (Springer, New York/Dordrecht, 2009), pp. 437–447 M.L. Cummings, J. How, A.K Whitten, O. Toupet, Proc. IEEE 100(3), 660–671 (2012) G. Dorais, R. Bonasso, D. Kortenkamp, B. Pell, D. Schreckenghost, Proceedings of the First International Conference of the Mars Society, vol. 1 (The Mars Society, Lakewood, 1998), pp. 369–378 B. Grosz, S. Kraus, in Foundations and Theories of Rational Agency, ed. by A. Rao, M. Wooldridge (Kluwer Academic, Dordrecht, 2003) R.M. Kilgore, K.A. Harper, C.E. Nehme, M.L. Cummings, Proceedings AIAA Infotech@ Aerospace, Sonoma, CA (AIAA, Reston, 2007) M. Lewis, H. Wang, S.Y. Chien, P. Velagapudi, P. Scerri, K.P. Sycara, Hum. Factors 52(2), 225–233 (2010) B. Pell, S. Sawyer, D.E. Bernard, N. Muscettola, B. Smith, Proceedings of the IEEE Aerospace Conference, Snowmass, CO, vol. 2 (IEEE, New York, 1998), pp. 289–313 S.J. Rasmussen, T. Shima, Int. J. Robust Nonlinear Control 18(2), 135–153 (2008) P. Scerri, D.V. Pynadath, M. Tambe, J. Artif. Intell. Res. 17, 171–228 Norwell, Massachusetts (2002) P. Scerri, D. Pynadath, L. Johnson, P. Rosenbloom, M. Si, N. Schurr, M. Tambe, Proceedings of the Second International Joint Conference on Autonomous Agents and Multiagent Systems, Melbourne, VIC (ACM, New York, 2003a), pp. 433–440 P. Scerri, D.V. Pynadath, M. Tambe, in Agent Autonomy, ed. by H. Hexmoor, C. Castelfranchi, R. Falcone, G. Weiss, vol. 7 (Kluwer Academic Publishers, Norwell, 2003b), pp. 211–241 T. Shima, S.J. Rasmussen (eds.), UAV Cooperative Decisions and Control: Challenges and Practical Approaches (SIAM, Philadelphia, 2009) K.P. Sycara, G. Sukthankar, Tech. Rep., Robotics Institute, Carnegie Mellon University, 2006 K.P. Sycara, M. Lewis, in Team Cognition: Understanding the Factor that Drive Process and Performance, ed. by E. Salas, S. Fiore (American Psychological Association, Washington DC, 2004) K.P. Sycara, M. Paolucii, J. Giampapa, M. van Velsen, Auton. Agents Multi-agent Syst. 7(1), 29– 48 (2003) K.P. Sycara, K. Decker, A. Pannu, M. Williamson, D. Zeng, IEEE Expert Intell. Syst. Appl. 2(6), 36–46 (2006)

Cooperative Mission and Path Planning for a Team of UAVs

62

Hyondong Oh, Hyo-Sang Shin, Seungkeun Kim, Antonios Tsourdos, and Brian A. White

Contents 62.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1510 62.2 Path Planning Using Dubins Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1512 62.2.1 Generating Dubins Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1513 62.2.2 Conditions for the Existence of Dubins Paths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1515 62.2.3 Length of Dubins Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1516 62.3 Cooperative Mission Planning I: Road-Network Search Route Planning . . . . . . . . . . . . . . . . 1517 62.3.1 Road-Network Search Route by Multiple Unmanned Aerial Vehicles . . . . . . . . . 1518 62.3.2 Optimization via MILP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1519 62.3.3 Approximation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1521 62.3.4 Numerical Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1525 62.4 Cooperative Mission Planning II: Communication Relay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1531 62.4.1 Waypoint Sequence Decision Making for Communication Relay . . . . . . . . . . . . . . 1531 62.4.2 Nonflying Zone Constraint Against UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1534 62.4.3 Speed Constraint of UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1538 62.4.4 Numerical Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1539 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1543

Abstract

This chapter addresses the cooperative mission and path-planning problem of multiple UAVs in the context of the vehicle-routing problem. Since the conventional vehicle-routing algorithms approximate their path to straight lines

H. Oh () • H-S. Shin • A. Tsourdos • B.A. White Department of Engineering Physics, School of Engineering, Cranfield University, Cranfield, Bedfordshire, UK e-mail: [email protected]; [email protected]; [email protected]; [email protected] S. Kim Department of Aerospace Engineering, Chungnam National University, Daejeon, South Korea e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 14, © Springer Science+Business Media Dordrecht 2015

1509

1510

H. Oh et al.

to reduce computational load, the physical constraints imposed on the vehicle are not to be taken into account. In order to mitigate this issue, this chapter describes a framework allowing integrated mission and path planning for coordinating UAVs using the Dubins theory based on the differential geometry concepts which can consider non-straight path segments. The main advantage of this approach is that the number of design parameters can be significantly reduced while providing the shortest, safe, and feasible path, which leads to a fast design process and more lightweight algorithms. In order to validate the integrated framework, cooperative mission and path-planning algorithms for two missions are developed: (1) road-network search route-planning patrolling every road segment of interest efficiently based on the optimization and approximation algorithm using nearest insertion and auction negotiation and (2) communication-relay route planning between a ground control station and the friendly fleet satisfying the constraints on the UAV speed and the avoidance of nonflying zones. Lastly, the performance of the proposed algorithms is examined via numerical simulations.

62.1

Introduction

The large scale of UAV applications has proliferated vastly within the last few years with the fielding of Global Hawk, Pioneer, Pathfinder Raven, and Dragoneye UAVs, among others (Samad et al. 2007). The operational experience of UAVs has proven that their technology can bring a dramatic impact to the battlefield. This includes, but is not limited to, obtaining real time, relevant situational awareness before making contact; helping commanders to lead appropriate decision making; and reducing risk to the mission and operation. Groups of UAVs are of special interest due to their ability to coordinate simultaneous coverage of large areas or cooperate to achieve common goals. The intelligent and autonomous cooperation of multiple UAVs operating in a team/swarm offers revolutionary capabilities: improved situation awareness; significant reductions in manpower and risk to humans; the ability to perform in hostile, hazardous, and geometrically complex environments; and cost efficiency. Specific applications under consideration for groups of cooperating UAVs include, but are not limited to, border patrol, search and rescue, surveillance, mapping, and environmental monitoring. In these applications, a group of UAVs becomes a mobile resource/sensor and, consequently, tasks and routes of each UAV need to be efficiently and optimally planned in order to cooperatively achieve their mission. Therefore, this chapter addresses the cooperative mission and pathplanning problem, which here is considered as a vehicle-routing problem of multiple UAVs for given missions. Since general vehicle-routing algorithms approximate their paths to straight lines in order to reduce the computational load, the physical constraints imposed

62 Cooperative Mission and Path Planning for a Team of UAVs

1511

on the vehicle are not to be taken into account. In order to mitigate this issue, this chapter describes a framework which allows integrated mission and path planning for coordinating UAVs. One of key enablers of this approach is the pathplanning scheme which was developed in a previous study (Tsourdos et al. 2010) based on the differential geometry concepts, especially Dubins paths. Path-planning algorithms using differential geometry examine the evolution of guidance geometry over time to derive curvature satisfying UAV constraints. Guidance commands defining a maneuver profile can be then computed using the derived curvature of the guidance geometry. One of the main advantages of this approach is that the number of design parameters can be significantly reduced while maintaining the guidance performance. Therefore, this approach will enable not only a fast design process and more lightweight algorithms but will also generate safe and feasible paths for multiple UAVs. This is required for the integration of operational and physical constraints of the UAVs into the cooperative mission and path-planning solution. Road-network search and communication-relay problems are also addressed to validate the integrated framework of mission and path planning. The road-network search routing problem enables multiple airborne platforms to efficiently patrol every road segment identified in the map of interest. This problem has been mainly handled in the operational research area (Gibbons 1999; Ahr 2004; Gross and Yellen 2003; Bektas 2006) and can be generally classified into two categories: one is the traveling salesman problem (TSP) which finds a shortest circular trip through a given number of cities, and the other is the Chinese postman problem (CPP) which finds the shortest path to travel along each road in the road network. The TSP using multiple UAVs can be considered as a task assignment problem to minimize the mission time or energy by assigning each target to an UAV, for which binary linear programming (Bellingham et al. 2003), iterative network flow (Chandler et al. 2002), tabu search algorithm (Ryan et al. 1998), and receding horizon control (Ahmadzadeh et al. 2006) have been proposed. Recently, Royset and Sato (2010) proposed a route optimization algorithm for multiple searchers to detect one or more probabilistically moving targets incorporating other factors such as environmental and platform-specific effects. Meanwhile, the CPP is normally used for ground vehicle applications such as road maintenance, snow disposal (Perrier et al. 2007), boundary coverage (Easton and Burdick 2005), and graph searching and sweeping (Alspach 2006; Parsons 1976). The communication-relay problem described in this chapter is used to extend the mission area of a main friendly fleet (FF) such as aircraft and ground convoys using relatively inexpensive subsystems, swarms of UAVs. Therefore, the focus is on cooperative route planning of multiple UAVs to maintain communication between the FF and a ground control station (GCS). UAVs in this problem are used as an effective communication-relay platform in environments characterized by poor radio frequency connectivity, which includes urban, forested, or mountainous regions where no line-of-sight exists between ground transmitters and receivers (Cerasoli and Eatontown 2007). In the late 1990s, a feasibility study which uses

1512

H. Oh et al.

the UAV as a communication relay develops a Battlefield Information Transmission System (BITS) system (Pinkney et al. 1996). The main objective of this study was to provide beyond line of sight communications within an area of operations without using scarce satellite resources. The study predicted that, with the advances in miniaturization technology and improved transmitter efficiencies, the UAV could carry multifunction and multiband transponders to handle the communication relay within size, weight, and power dissipation budgets. Cerasoli (Cerasoli and Eatontown 2007) assessed the practical effectiveness of a UAV communication relay in an urban area using a ray tracing method. The paper concluded that a UAV at 2,000 m provided coverage for over 90 % of the ground receivers with 10 dB of LOS path loss by analysis of UAVs placed at various positions and heights over an approximately 500-m2 urban area. The remainder of this chapter is organized as follows. Section 62.2 introduces an overview of path planning using Dubins paths and explains how to allow for physical constraints of the fixed-wing UAVs in Dubins path planning. Then, Sect. 62.3 proposes road-network search route-planning algorithms for multiple UAVs based on mixed integer linear programming (MILP) optimization and the approximation algorithm using nearest insertion and auction negotiation. Section 62.4 describes route-planning and decision-making algorithms for a group of UAVs in order to guarantee communication between the GCS and the FF. Simulation results and analysis of each problem are also included in each section. Conclusions are given in Sect. 62.4.4.2.

62.2

Path Planning Using Dubins Paths

Several path-planning methodologies based on differential geometry have been proposed: Dubins curves, clothoid arcs, and Pythagorean hodograph curves (Shanmugavel et al. 2007; Tsourdos et al. 2010). This section will briefly introduce the concept of Dubins path planning based on the reference (Shanmugavel 2007; Kim et al. 2011), and the detailed algorithm will be developed to design flyable and safe Dubins path transiting between waypoints for cooperative missions of multiple UAVs. The Dubins path is considered because it is the shortest path, has simple geometry, and is computationally efficient. Path planning is defined as the geometric evolution of curves between two desired poses in free space, Cfree . The pose in 2D comprises the position coordinates .x; y/ and the orientation . A simple case of producing path between two poses is first considered. This can be extended into any number of waypoints/poses: r.q/

Ps .xs ; ys ; s / ! Pf .xf ; yf ; f /;

r.q/ 2 Cfree ;

j.t/j  max

(62.1)

where Ps and Pf denote a starting and final pose, respectively, .t/ represents the curvature of path r.q/ with parameter q, and max is the maximum curvature

62 Cooperative Mission and Path Planning for a Team of UAVs

1513

imposed by the UAV dynamic constraints such as maximum lateral acceleration. Motion in the plane is composed of either rectilinear or turning or angular motions. A straight line provides the shortest rectilinear motion, and the circular arc provides the shortest turning or angular motion. Also, the arc provides a constant turning radius which also satisfies the maximum curvature constraint, that is, the minimum turning radius, which is a function of speed and maximum lateral acceleration. The Dubins path (Dubins 1957) is the shortest path between two vectors in a plane, and the path meets the minimum bound on turning radius. The Dubins path is a composite path formed either by two circular arcs connected by a common tangent or three consecutive tangential circular arcs or a subset of either of these two. The first path is a CLC path, and the second one is a CCC path, where “C” stands for circular segment and “L” stands for line segment. Either of these two curves will form the shortest path between two poses and so is a good approach for UAV path planning. This section focuses on Dubins paths of CLC type, and the details of CCC type paths can be found in Oh et al. (2011a).

62.2.1 Generating Dubins Path The Dubins path can be produced by solving (Eq. 62.1). However, it is computationally efficient if it is produced by geometric principles owing to its simple geometry that it is formed by two circular arcs connected by their common tangents. Therefore, the principles of analytic geometry are used to produce the Dubins paths. There are two possible tangents between the arcs: (1) an external tangent, where the start and finish maneuvers have same turning directions, and (2) an internal tangent, where the turning maneuvers have opposite turning directions (e.g., if the start maneuver is clockwise, the finish maneuver will be anticlockwise and vice versa). Here, only the Dubins path with an external tangent is derived (the case of an internal tangent is analogous). Referring to Fig. 62.1, the input parameters for producing the Dubins path are: i) Initial pose: Ps .xs ; ys ; s / ii) Finish pose: Pf .xf ; yf ; f / iii) Initial turning radius: s .D 1s / iv) Finish turning radius: f .D 1f / 1. Find the centers of the turning circles Os .xcs ; ycs / and Of .xcf ; ycf /: .xcs ; ycs / D .xs ˙ s cos.s ˙ =2/; ys ˙ s sin.s ˙ =2//

(62.2a)

.xcf ; ycf / D .xf ˙ f cos.f ˙ =2/; yf ˙ f sin.f ˙ =2// (62.2b) where Os and Of are called primary circles represented by Cs and Cf , respectively. 2. Draw a secondary circle of radius jf  s j at Of for s  f .

1514

H. Oh et al.

3. Connect the centers Os and Of to form a line c, called center line, where jcj D p .xcs  xcf /2 C .ycs  ycf /2 . 4. Draw a line Os T 0 tangent to the secondary circle Csec . 5. Draw a perpendicular straight line from Of to Os T 0 , which intersects the primary circle Cf at TN , which is called as a tangent entry point on Cf . 6. Draw a line TX TN parallel to Os T 0 , where TX is called as a tangent exit point on Cs . 7. Connect the points Ps and TX by an arc of radius s and TN and Pf by an arc of radius f . 8. The composite path is then formed by the starting arc Ps TX , followed by the external tangent line TX TN and the finishing arc TN Pf . The triangle 4Os Of T 0 is a right-angled triangle with hypotenuse Os Of , and the other two sides are Of T 0 and Os T 0 , where jjOf T 0 jj D jf  s j. The included angle between Os Of and Os T 0 is e , where f  s e D arcsin jcj The slope of the line c is

! (62.3)

, where ycf  ycs D arctan xcf  xcs

! (62.4)

The angles ex D †.X1 Os TX and en D †.X2 Of TN are calculated from ! ! Table 62.1 where Os X1 and Of X2 are parallel to the positive TN pf

T’ f

C s

Fig. 62.1 Dubins – design of CLC path

ps

x2

of

TX

os

Csec

Cf

x1 X

Cs

Table 62.1 Calculation of tangent exit and entry points Start-turn Clockwise

Finish-turn Clockwise

Anticlockwise

Anticlockwise

e     arcsin f c s     arcsin f c s

ex e C

 2

C

en e C

 2

C

e 

 2

C

e 

 2

C

62 Cooperative Mission and Path Planning for a Team of UAVs

1515

Fig. 62.2 Tangent circles L

R P C2 C1

direction of x-coordinate. The values of ex and en , the tangent exit and entry points, are calculated as   TX D xcs C s cos.ex /; ycs C s sin.ex /   TN D xcf C f cos.en /; ycf C f sin.en /

(62.5a) (62.5b)

From the figure, it can be seen that the turning maneuvers have clockwise rotations. Similarly, other type of Dubins path using the internal tangent can be produced by drawing the secondary circle of radius equal to js C f j. It is worth pointing out that the calculation of the tangent exit and entry points TX and TN is central in generating the Dubins path. For a given pose, there are two circles tangent to it. Referring to Fig. 62.2, the pose P has a right turn R on the arc C1 and a left turn L on the arc C2 . If s and f are fixed, four Dubins paths are possible between Ps and Pf , which are fRSR, RSL, LSR, LSLg, where represents the tangent. However, if the final orientation is taken either ˙f , the number of Dubins paths between Ps and Pf will become eight. Figure 62.3 shows the eight possible Dubins paths: four paths each from the primary circle C1 to the secondary circles C3 and C4 and four from C2 to C3 and C4 .

62.2.2 Conditions for the Existence of Dubins Paths From Sect. 62.2.1, it can be seen that the existence of the Dubins path is determined by the common tangents between the turning arcs. These tangents vanish under two conditions. The external tangent vanishes when one of the primary circles Cs and Cf contains the other, while the internal one vanishes when the primary circles intersect. Both of these conditions together determine the existence of the Dubins

1516

H. Oh et al.

C4

C4

C4

C4 C3

C3

C3

C3 C2

C2

C2

C1

C1 Right to Right C1 to C3

Right to Right C1 to C2

Right to Left C1 to C3

C3

Left to Left C2 to C4

C2

C1 Left to Left C2 to C3

C3

C3

C2

C1

Right to Left C1 to C4

C4

C4

C3 C2

C1

C1

C4

C4

C2

C2

C1 Left to Right C2 to C4

C1 Left to Right C2 to C3

Fig. 62.3 Dubins paths with f as a free variable

path between given two poses, which in turn are a function of the turning radii and hance curvature: External tangent W .c C s / > f ; f > s

(62.6a)

Internal tangent W c > .s C f /; f > s

(62.6b)

62.2.3 Length of Dubins Paths The Dubins path is a composite path of two circular arcs and a straight line. Hence, the path length sDubins is the sum of the lengths of individual path segments. Since the length of the common tangent connecting the arcs is determined by the radii of the arcs, the length is also the function of the turning radii. Hence, the length of the path can be varied by changing the radii (curvatures). Also, any two paths can be made equal in length by simply varying the curvature of the arcs: sDubins D ss C st C sf

(62.7a)

62 Cooperative Mission and Path Planning for a Team of UAVs

D js ˛s j C kTX TN k C jf ˛f j sDubins D f .s ; f /

1517

(62.7b) (62.7c)

path, ˛s and ˛f are the included angles, where sDubins is the length of the Dubins p ˛s D ex , ˛f D en , and kTX TN k D .yN  yX /2 C .xN  xX /2 . Using this length of the Dubins path, a reference velocity for each UAV can be computed in order to control the time taken for the UAV to traverse each path for cooperative mission planning.

62.3

Cooperative Mission Planning I: Road-Network Search Route Planning

For a road search route-planning mission, a road network is established as a set of straight line connecting a set of waypoints. These waypoints are located either on road junctions or along the roads with sufficient separations between them to allow for accurate representation of a curved road by using a set of straight lines. A sample road network is shown in Fig. 62.4a, which is based on the Google map of a village in the UK. This road network can be translated to a graph composed of straight line segments connecting a set of vertices, as shown in Fig. 62.4b (Oh et al. 2011b). In order to search all the roads within the map of interest, there are generally two typical routing problems (Ahr 2004): Traveling Salesman Problem (TSP): A salesman has to visit several cities (or road junctions). Starting at a certain city, the TSP finds a route with minimum travel distance on which the salesman traverses each of the destination cities exactly once (and for the closed TSP, leads him back to his starting point). Chinese Postman Problem (CPP): A postman has to deliver mail for a network of streets. Starting at a given point, for example, the post office, the CPP finds a route with minimum travel distance which allows the postman to stop by each street at least once (and for the closed CPP, leading him back to the post office). Consider the CPP and its variants, which involve constructing a tour of all the roads with the shortest distance of the road network. Typically, the road network is mapped to an undirected graph G D .V; E/; having edge weights w W E ! R0C , where the roads are represented by the edge set E D fe1 ; : : : ; en g and the road junctions are represented by the vertex set V D fv1 ; : : : ; vm g as numbered in Fig. 62.4b. Each edge ei D fvei;1 ; vei;2 g is weighted with its length or the amount of time required to traverse it. The basic CPP algorithm involves first constructing an even graph from the road network which has a set of vertices with an even number of edges attached to them producing a pair of entry and exit points. When the roadnetwork graph has junctions with an odd number of edges, some roads therefore must be selected as exceptions for multiple visits by the postman to make the even graph. The search pattern (tour) of the even graph is calculated by determining the Euler tour of the graph (Gross and Yellen 2003), which visits every edge of the even graph exactly once.

1518

a 500 Y position (m)

Fig. 62.4 The graphic representation of a road network. (a) The satellite map of a village in the UK. (b) Its graphic representation by straight line segments

H. Oh et al.

1000 1500 2000 2500 500 1000 1500 2000 2500 3000 3500 X position (m) The satellite map of a village in the UK

b

2500

Y position (m)

2000

1500

1000

500

500

1000 1500 2000 2500 3000 3500 X position (m)

Its graphic representation by straight line segments

62.3.1 Road-Network Search Route by Multiple Unmanned Aerial Vehicles The conventional CPP algorithm has been applied to a fully connected road network for use by ground vehicles. However, since UAVs do not have any restrictions such that they must only move along the roads, the CPP algorithm needs to be modified for the case that UAVs search a general road map having unconnected road segments. The modified CPP (mCPP) generates a tour of the road network traveling all the roads once no matter what the type area of interest map is: an even or odd graph. Even searching the area having no road somewhere in it can be tackled by the mCPP algorithm by generating a virtual road pattern with a lawnmower (Maza and Ollero 2007) or spiral-like (Nigam and Kroo 2008) algorithm.

62 Cooperative Mission and Path Planning for a Team of UAVs Fig. 62.5 An illustration of the modified Chinese Postman Problem (mCPP)

1519

500

4

UAV 1

450

2

1 400

3

2

Unconnected/ unstructured area

350

Structured area

14

300 9

250

3

8

7

12

6

6

10

200

5 7

Dynamic target

9

8

100 100

5

4

Static target

13

150

Road Segment

Sensor footprint

11

200

300

400

500

600

700

Figure 62.5 exemplifies a sample road-network search problem to be solved by the mCPP. As the CPP algorithms generally approximate their path to a straight line for simplicity, the mCPP algorithm using a straight line path is called the modified Euclidean CPP (mECPP) for the rest of this chapter. However, in order to produce the shortest path flyable by the UAV that connects the road segment sequence selected by the search route algorithm, the flight constraints of the UAV have to be taken into account. To accommodate this, the Dubins path is incorporated into the mCPP algorithm instead of using just a straight line to connect the roads, and this is called the modified Dubins CPP (mDCPP). A road-network search routeplanning algorithm using the optimization via MILP for the mCPP problem is first developed. A new approximation algorithm is then proposed as a more practical approach to reduce the complexity of the algorithm.

62.3.2 Optimization via MILP The mCPP is solved by use of the MILP optimization using the multidimension multi-choice knapsack problem (MMKP) formulation to find an suboptimal solution minimizing the total flight time of UAVs. The classical MMKP picks up items for the knapsacks to have maximum total value so that the total resource required does not exceed the constraints of knapsacks (Hifi et al. 2006). In order to apply the MMKP to the road-network search, the UAVs are assumed to be the knapsacks, the roads to be searched are the resources, and the limited flight time or energy for each UAV as the capacity of knapsack. The MMKP formulation allows the search problem to consider the characteristics of each UAV such as flight time capacity and minimum turning radius. The details of the proposed road-network search algorithm for multiple UAVs are explained as follows:

1520

H. Oh et al.

Fig. 62.6 An example of CCC and CLC paths following the road

Sensor footprint

CLC Path only 5

4

CLC & CCC Path

3

8 6

6

5 7

62.3.2.1 Dubins Path Planning Once the shortest edge permutations are determined, the next step is to compute and to store the cost (path length or flight time) and to then compute the Dubins paths to connect them. Although the CLC path is being used in general case, this study also explores the CCC path for a densely distributed road environment, because the CLC path cannot be applied in all cases (cf. Sect. 62.2.2). Moreover, to follow the road precisely taking into account the sensor footprint coverage, the path should consist of both CCC and CLC forms of Dubins path. Figure 62.6 shows an example of a road search path using CLC and CCC path for a small sensor footprint, which results in a detour at the road intersection.

62.3.2.2 Generation of the Shortest Edge Permutation First of all, unordered feasible edge (i.e., road) permutations to be visited by the UAV are generated for all possible cases with a given petal size. The petal size means the maximum number of edges that can be visited by one UAV and is determined by the amount of resources available to it. If the end vertex of one edge and any vertex of the next edge are not connected, they are connected with an additional edge which has a shorter distance. Then, the shortest order-of-visit edge permutations considering the initial position of each UAV are computed under the assumption that a path is represented as a straight line.

62.3.2.3 MMKP Formulation and MILP Optimization The final step of the proposed algorithm is to allocate the edge permutations to each UAV so as to cover all the roads with a minimum flying time. This can be expressed by a MMKP formulation as

62 Cooperative Mission and Path Planning for a Team of UAVs

min J D

Npi N UAV X X

Tij xij

1521

(62.8)

i D1 j D1

s.t

Npi N UAV X X

Ekj xij  1; for k D 1; : : : ; Nedge

(62.9)

i D1 j D1 Npi X

xij D 1; xij 2 f0; 1g; for i D 1; : : : ; NUAV ; j D 1; : : : ; Npi (62.10)

j D1

where NUAV , Nedge , and Npi denote the number of UAVs, the edges to be visited, and the permutations generated by the i -th UAV, respectively. Tij represents the mission cost (flight time) of the j -th permutation of the i -th UAV, and Ekj represents the matrix whose k-th element of the j -th permutation is 1 if edge k visited; otherwise, 0 and xij is either 0, implying permutation j of the i -th vehicle is not picked or 1 implying permutation j of the i -th UAV is picked. The first constraint states that the UAVs should visit every edge once or more, and the second constraint represents the allocation of exactly one edge permutation to the each UAV. This MMKP problem is solved using a SYMPHONY MILP solver (Ralphs et al. 2010). It should be noted that, depending on the petal size, the computational burden of the mission cost Tij of all edge permutations can be significant.

62.3.3 Approximation Algorithm 62.3.3.1 Nearest Insertion-Based mDCPP Due to the complexity of the problem, instead of using the optimization method explained above, an approximation algorithm is developed as a more practical solution to the mCPP (Oh et al. 2011a, 2012). To develop the approximation algorithm, the TSP algorithm is first studied. Although there are a lot of algorithms for the TSP (Rosenkrantz et al. 2009), one heuristic approach, a nearest insertion method, is adopted here since it is fast and easy to implement. The basic idea of the insertion method is to construct the approximate tour by a sequence of steps in which tours are constructed for progressively larger subsets of the nodes of the graph. It produces a tour no longer than twice the optimal regardless of the number of nodes in the problem and runs in a time proportional to the square of the nodes (Rosenkrantz et al. 2009). Having this in mind, the nearest insertion-based mDCPP (NI-mDCPP) algorithm is developed as illustrated in Fig. 62.7. The NI-mDCPP algorithm for the single UAV is described as follows. Algorithm Description 1. Start from a certain point or road junction and select the nearest road to it using the Dubins path length. 2. Make and grow a tour by finding the nearest road to any of the selected tour roads.

1522

H. Oh et al.

3. Compute the cost of insertion to decide whether to insert before or stack after the closest road to the tour. 4. Insert the selected road in the decided position. 5. 2  4 are repeated until all roads are included in the tour.

62.3.3.2 Euclidean Distance Order Approximation To reduce the computation burden further for the dynamic environment, an additional approximation algorithm which uses the Euclidean distance order is incorporated into the NI-mDCPP algorithm. This algorithm is described as follows. First of all, make an ascending order of road list for both the Euclidean distance and the Dubins path length between all pairs of end points of the road network, and find the maximum number, norder;max which guarantees that road list of Euclidean distance within that number contains the shortest Dubins path. Note that although norder;max , is determined before running the algorithm given information of the road network, a size of norder;max would remain almost the same against minor changes of road information for an uncertain dynamic environment. Then, when finding the nearest roads, make the ascending order list of distances between the edge of interest and all the other roads using the Euclidean distance first, and find the nearest road whose Dubins path length is the shortest among the roads in the norder;max Euclidean distance order. In other words, this method computes only norder;max Dubins path distances instead of computing all the Dubins lengths between one road the other roads. 62.3.3.3 Negotiation for Multiple UAVs Having developed the NI-mDCPP algorithm for the single UAV, it can be extended to the case of multiple UAVs using auction-based negotiation. The algorithm is described as follows. Algorithm Description 1. Start with N initial positions or roads of N UAVs. 2. Make and grow a tour by using single NI-mDCPP algorithm while storing the cost (path length or flight time) between selected tour roads and remaining edges (which was needed for finding the nearest vertex). 3. When conflict occurs, that is, more than one UAV wants the same road for the next tour, the auction algorithm using stored cost is used to match UAVs with the task (road) to minimize the cost. Figure 62.8 illustrates the procedure of the algorithm. Since road 3 is not searched yet in Fig. 62.8d, each UAV sends its cost for the given task (in this case, visiting road 3), and then an auction or bipartite (linear programming) approach is used to match UAVs with the task to minimize the cost. Although overlapping road segments are avoided using the auction algorithm, collision between UAVs might occur during the transition from one road to the next. In that case, if necessary, the path can be replanned either by visiting a new road or by modifying the curvature of the arc of the Dubins path (Tsourdos et al. 2010). For simplicity, it is assumed that the collision avoidance is done by a local flight controller or by operating the UAVs

62 Cooperative Mission and Path Planning for a Team of UAVs

1523

Fig. 62.7 An example simulation of the NI-mDCPP road search algorithm. (a) Find the nearest edge. (b) First insertion. (c) Find the nearest edge. (d) Second insertion. (e) Find the nearest edge. (f) Third insertion. (g) Fourth insertion. (h) Eighth insertion

1524

H. Oh et al.

Fig. 62.8 Negotiation procedure for multiple UAVs. (a) Step 0. (b) Step 1. (c) Step 2. (d) Step 3. (e) Step 4

at different altitudes. The proposed algorithm is rather simple but straightforward and can be run in real time. Moreover, by including additional factors such as different minimum turning radii and total path length (or the number of roads) assigned to each UAV into the cost, the auction-based negotiation can be made very flexible and can deal with heterogeneous UAVs.

62 Cooperative Mission and Path Planning for a Team of UAVs

1525

62.3.4 Numerical Simulations 62.3.4.1 Performance Comparison To evaluate the performance of the proposed road-network search algorithms for multiple UAVs, numerical simulations are performed for a specific scenario with four UAVs, and the road network shown in Fig. 62.4a. Each UAV has different dynamic constraints given by: • Minimum turning radius min : [100 90 80 70] m • Maximum cruise speed Vc;max : [60 50 40 30] m/s The maximum curvature max of the UAVs are given by max D 1=min . The UAVs are assumed to have maximum cruising speed during the entire mission, and the maximum petal size of the edge permutation is set to five. Figure 62.9a shows the result of the road-network search using MILP optimization. The flight path is smooth and flyable due to the Dubins path planning, and since the UAV does not need to only fly along the road, the results include additional paths connecting some of the unconnected roads. The total flight length of all UAVs is 2,798.1 m, and its flight duration is 583.1 s. In this scenario, the total computation time exceeds a reasonable limit (>5 min) using a normal PC system (Core 2 CPU, 2.0 Ghz, and 512 MB RAM). Figure 62.9b shows the search results of the NI-mDCPP algorithm. The NI-mDCPP gives a solution within less than a second having about 12 % longer flight time (653.4 s) than that of the MILP optimization. Considering both the computation time and performance, the NI-mDCPP can be regarded as a preferable approach for the given sample map or a more complex scenario. 62.3.4.2 Monte Carlo Simulation of Approximation Method To evaluate the properties and performance of the proposed approximation algorithm, Monte Carlo simulations are performed using a random map with different parameters defining the map size and the number of UAVs. The map environment is composed of 10 by 10 vertices, numbered as shown in Fig. 62.10, and the road edges are generated by connecting two randomly selected vertices. To check the impact of the map size on the Dubins path planning, the distance dmap between the adjacent vertex is set to be proportional to the minimum turning radius min of the UAV as dmap D Ks  min

(62.11)

where Ks is the scale factor. Moreover, some of the selected edges whose lengths are longer than three times dmap are discarded to get a well-distributed road network and to distribute the roads to each UAV with a similar length. Figure 62.10 shows the sample road network with 20 randomly chosen edges. By Monte Carlo simulations, the effect of three major factors for the road-network search route planning can be investigated. These are: The distribution density of road network: densely or sparsely distributed road relative to the minimum turning radius of UAVs The type of path planning: straight line or Dubins path

1526

a

0 500

Y position (m)

Fig. 62.9 Road-network search route-planning results using multiple UAVs. (a) MILP optimization. (b) NI-mDCPP

H. Oh et al.

1000 1500 2000 2500

b

0

500

1000 1500 2000 2500 3000 3500 X position (m) MILP Optimisation

0

500

1000 1500 2000 2500 3000 3500 X position (m)

0

Y position (m)

500 1000 1500 2000 2500

NI-mDCPP

The number of UAVs: computation time, path length, and the longest path length of a single UAV This will provide information on priorities for the efficient use of the UAV group in the planning phase of the autonomous search mission. In the simulations, the UAVs are assumed to have a constant velocity and minimum turning radius of min D 50 m. The simulation results are the average of 50 runs. Single UAV Case The first set of simulations are performed using a single UAV with different road map scales. For the rest of this section, the terms of mDCPP and the mECPP are used for the NI-mDCPP and the NI-mECPP, respectively. One of the search route-planning results using the mDCPP for a random map is shown in Fig. 62.11, which covers all the roads satisfying turning constraints of UAVs. Figure 62.12a displays the computation time ratio between the Dubins path (mDCPP) and the

62 Cooperative Mission and Path Planning for a Team of UAVs

1527

Fig. 62.10 A sample road network with 20 randomly chosen edges (Ks D 1; min D 50 m)

Fig. 62.11 NI-mDCPP road search path with different map scale factors. (a) Ks D 1. (b) Ks D 2. (c) Ks D 3. (d) Ks D 4

1528

a

H. Oh et al.

b

40

18 16

30

norder,max

tmDCPP/tmECPP

35

25 20 15

14 12 10 8

10

6

NI-mDCPP NI-mDPP with Euclidean approximation

5 0 0.5

20

1

1.5

2

2.5

3

3.5

4

4.5

4 5

2 0.5

1

1.5

2

2.5

3

3.5

4

4.5

Map scale (KS)

Map scale (KS)

Computation time ratio

norder,max for Euclidean approximation

c

5

1.05

LmDCPP/L*mECPP

1 0.95 0.9 0.85 0.8

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Map scale (KS) Path length ratio

Fig. 62.12 NI-mDCPP results with different map scale factors, average for 50 simulations. (a) Computation time ratio. (b) norder;max for Euclidean approximation. (c) Path length ratio

straight line (mECPP). Regardless of the map scale, the mDcPP algorithm is around consistently 35 times slower than the mECPP. Meanwhile, the computation time of the mDCPP along with the Euclidean distance order approximation, as explained in Sect. 62.3.4, decreases as the map scale increases as a result of a decrease in the maximum order norder;max as shown in Fig. 62.12b. Figure 62.12c compares the total path length to cover the entire road map using the mDCPP and the mECPP. For fair comparison, the length of the mECPP (denoted by LmECPP ) is computed by road search route using the mECPP algorithm but connecting the roads using a Dubins path. This is because although road search route planning is performed using the mECPP, a real trajectory of the UAV connecting the roads should be of Dubins restricted by its maximum curvature. When the minimum turning radius is relatively small compared to the distance between roads, that is, when the map scale is small, the path length of the mDCPP is shorter than that of the mECPP. However, as the map scale gets bigger, the path length ratio gets closer to one (or even greater than one) since the road search order using Dubins paths is almost the same as using a straight line, as one can expect from Fig. 62.11.

62 Cooperative Mission and Path Planning for a Team of UAVs

1529

Fig. 62.13 NI-mDCPP road search path with different number of UAVs (Ks D 2). (a) 1 UAV. (b) 2 UAVs. (c) 3 UAVs. (d) 4 UAVs. (e) 5 UAVs. (g) 6 UAVs

Multiple UAVs Case The second simulation is performed using multiple UAVs with different road map scales, and one of the search route-planning results using the mDCPP is shown in Fig. 62.13. The initial position of each UAV is equally distributed around the road area. Figure 62.14 shows the normalized simulation results for

1530

b

1 0.9

Normalised Dubins path length

Normalised computation time

a

H. Oh et al.

Ks=0.5 Ks=1 Ks=2 Ks=3 Ks=4

0.8 0.7 0.6 0.5 0.4 0.3

1.05

1

0.95 Ks=0.5 Ks=1 Ks=2 Ks=3 Ks=4

0.9

0.85

0.2 1

2

4 5 3 Number of UAVs

6

7

1

2

Computation time

c

d

0.6

7

3.4 3.2

Ks=0.5 Ks=1 Ks=2 Ks=3 Ks=4

Performance index

Normalised longest Dubins path length of one UAV

0.9

0.7

6

Total Dubins path length

1

0.8

3 4 5 Number of UAVs

0.5 0.4

Ks=0.5 Ks=1 Ks=2 Ks=3 Ks=4

3 2.8 2.6 2.4

0.3 2.2

0.2 0.1

2 1

2

3

4

5

6

7

1

2

3

4

5

Number of UAVs

Number of UAVs

The longest Dubins path of one UAV

Performance index cost

6

7

Fig. 62.14 NI-mDCPP results with different number of UAVs (Ks D 2), average for 50 simulations. (a) Computation time. (b) Total Dubins path length. (c) The longest Dubins path of one UAV. (d) Performance index cost

a single UAV. In particular, the longest path length (Fig. 62.14c) of the UAV is of interest since it is equivalent to the mission completion time of the entire UAV team. The normalized computation time and the longest path length of the UAV decrease as the size of the UAV team is increased, regardless of the map scale as each UAV takes partial charge of the road search mission cooperatively using the auction-based task assignment. The total path length (Fig. 62.14b) is significantly affected by the map scale. When the map scale is small, the total path length decreases in proportion to the number of UAVs. Whereas in a relatively big map environment, the normalized path length remains close to one since each UAV will fly a long distance from the initial position or from one road to another road. Apparently, the simulation results show that the bigger the UAV team size is, the better performance it shows in terms of the computation time and path length. However, using a large number of UAVs on the team require more operational cost, such as fuel and communication resources. Therefore, the performance index to determine the optimal size of the UAV team for the search mission is proposed as shown in (Eq. 62.12), which includes an additional operational cost for each UAV,

62 Cooperative Mission and Path Planning for a Team of UAVs

1531

normalized by the number of UAVs (by maximum seven UAVs in this case), nN v , and its weighting factor, wnv : J D wt tN C wl lN C wL LN C wnv nNv

(62.12)

where wt , wl , and wL represent the weighting factors of computation time, the total path length and the longest path length, of one UAV, respectively. Under the assumption that all of the weighting factors are set to one, the number of UAVs to minimize the performance index J can be selected as four for all map scale factors as shown in Fig. 62.14d.

62.4

Cooperative Mission Planning II: Communication Relay

This section presents the development of trajectory planning for multiple UAVs making use of Dubins path theory in order to maintain and optimize the communication relay between a FF (friendly fleet) performing a main mission and a GCS (ground control station) centrally administrating the whole mission (Kim et al. 2011). To apply Dubins path theory to the optimization of communication between UAVs and with the GCS, various strategic or dynamic constraints have to be considered: mission planning of the FF, the positions of the GCS, any existing nonflying zones, and the limits of UAV dynamics. These constraints affect the decision making of the waypoint sequences of the UAV members. (The final, definitive version of the Section 62.4 of this chapter has been published in Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 225/1, Jan/2011 (http://pig.sagepub.com/content/225/1/12. c 2011, Institution of abstract) by SAGE Publications Ltd, All rights reserved.  Mechanical Engineers.)

62.4.1 Waypoint Sequence Decision Making for Communication Relay The decision making of the UAVs will be based on the waypoint sequence WF of the FF which is a priory known since the friendly fleet’s future movement can be broadcasted to the UAV swarm and the GCS: WF D ŒWF 0 ; WF 1 ; WF 2 ; : : : ; WF n 

(62.13)

where WF i D .xF i ; yF i /T and n is the number of waypoints to be passed by the FF. Awareness of the asset speed VF makes it possible to get the arrival-time sequence TF at the waypoint sequence defined in (Eq. 62.13): TF D ŒtF 0 ; tF 1 ; tF 2 ; : : : ; tF n 

(62.14)

1532

H. Oh et al.

Fig. 62.15 Decision making on spreading out the UAV members: the distance from the GCS to the FF is shorter than twice of the communication range ko Rcom

WFi

Friendly Fleet

UAVs

ds

Rcom

yi

Wuc

yF

ko Rcom

P GCS

Rcom

where tF i is the arrival time at the i -th waypoint. This can be computed using the speed information of the FF which is assumed constant over the mission. Given these waypoint and arrival-time sequences of the FF, the waypoint sequence of Nu UAVs can be determined by considering communication relay between the GCS and the FF. Let us consider a ratio of the distance from the GCS position P D .xp ; yp /T to the FF over the communication range Rcom as shown in Fig. 62.15 given as kF D max i

jWF i  P j ; i 2 f0; 1;    ; ng Rcom

(62.15)

It is assumed that each UAV has the same communication range as the GCS. Firstly, consider the case where the distance from the GCS to the FF is shorter than .1 C ko /-times the communication range Rcom , that is, kF  .1 C ko / as shown in Fig. 62.15 (ko is given by (Eq. 62.16)). If the FF’s next waypoint and arrival time are given in (Eqs. 62.13) and (62.14), the target center position of UAVs, Wuc , is assigned as the intersection between a circle of the communication range adjusted by an overlap coefficient ko and the line-of-sight from the GCS position to the asset’s next waypoint WF i , where Wuc D P C ko Rcom Œcos and where F

D tan1

F ; sin

yF i  yp xF i  xp

T F

(62.16) (62.17)

The distribution of UAVs on the boundary of the communication range aims at maximizing the coverage of communication when the FF goes out of the

62 Cooperative Mission and Path Planning for a Team of UAVs

1533

communication range. Note that 0 < ko  1 is adjustable to make the boundaries between the communication circles overlap in order to prevent loss of communication. The choice of the design parameter ko is made by a trade-off analysis. A target waypoint of an individual UAV is next determined to distribute all of the swarm members relatively to the target center position Wuc . All the UAV members should spread out in a single line, as shown in Fig. 62.15, because the other UAVs enable redundant communication channels in an emergency, although in this case, a single UAV is enough to maintain the communication coverage. The distribution ! line needs to be perpendicular to the line-of-sight P WF i in order to maximize the lateral communication coverage. Thus, the azimuth of the distribution line relative to the GCS site, l , is obtained using the azimuth of the FF relative to the GCS site, F , computed with (Eq. 62.17) to give l

D

F

 =2

(62.18)

As a result, the position vector of the j -th UAV is decided by a rotational transformation using the number of UAVs Nu , the angle of the distribution line l , and the target center position Wuc : Wuj D Wuc C



cos. l / sin. l / sin. l / cos. l /



..Nu  1/=2 C .j  1//ds 0

 (62.19)

where j 2 f1; 2;    ; Nu g and ds is a separation distance between the UAVs at the target waypoint. In the general case, consider the case where the distance from the GCS to the FF exceeds .1 C ko /-times the communication range Rcom , that is, kF > .1 C ko / in (Eq. 62.15). To maintain the communication relay, the distribution of the UAVs has to be restructured since a single UAV cannot link the communication between the FF and the GCS. This study adopts a restructuring policy such that some of the UAV members can move to communication-relay positions having the shape of a chain defined by kF . First, choose the number of UAVs Nm that should move to the new communication-relay positions in order not to lose the communication as kF Nm D< 1> (62.20) ko where < x > rounds the value of x to the nearest integer toward minus infinity. Note that the closer that kF to ko is, the fewer UAVs are needed, but the smaller the overlapping areas are. For example, when the FF flies to the waypoint defined by 2ko  kF < 3ko with the help of a four-UAV swarm, Nm becomes 1; hence, only the fourth UAV member of the swarm should move to the new relay waypoint point Wur4 , as shown in Fig. 62.16. In a similar manner, generalize the relay position of Wurk by Wurk D WF i C



cos. sin.

F/

 sin. F / F / cos. F /



.Nu  k C 1/ko Rcom 0

 (62.21)

1534

H. Oh et al.

Friendly Fleet

W Fi

k o R com

R com W ur4

UAVs

yi

W uc

(kF - 2ko)Rcom

ds

yF

k o R com P

GCS

R com

R com

Fig. 62.16 Decision making on spreading out the UAV members: the distance from the GCS to the FF is longer than twice of the communication range ko Rcom

where k 2 fNu  Nm C 1; Nu  Nm C 2;    ; Nu g. In short, Nm number of UAVs move to the relay positions Wurk as defined in (Eq. 62.21), while the other Nu  Nm number of UAVs are distributed on the line as defined in (Eq. 62.19). In this way, the communication between the GCS and the FF can be maintained in virtue of the reconstruction of the UAV formation no matter that the FF maneuvers beyond the communication boundary of the GCS. Finally, the following waypoint and arrival time sequences of the j -th UAV can be obtained using (Eqs. 62.19), (62.21), and (62.14): j

j

j

j

j WU D ŒWu0 ; Wu1 ; Wu2 ; : : : ; Wun  j TU

D

j j j j Œtu0 ; tu1 ; tu2 ; : : : ; tun 

(62.22) (62.23)

j

where tui D tai since the UAV has to arrive at the same time as the assets to protect them effectively.

62.4.2 Nonflying Zone Constraint Against UAV In a battlefield or an urban area, a NFZ (nonflying zone) could exist as an obstacle dangerous to a swarm of UAVs generated by an enemy, a high building/structure, or an airport space. When a threat is detected, the path planner generates an intermediate waypoint so that the threat is avoided. Consider a flyable trajectory r.q/ generated for a given set of poses/waypoints. The schematic of the concept

62 Cooperative Mission and Path Planning for a Team of UAVs

1535

Fig. 62.17 Threat handling by intermediate pose

Fig. 62.18 Case 1: The line between the target waypoints crosses the NFZ

is shown in Fig. 62.17. In the figure, the central hatched circle is the threat. The intermediate waypoints are generated by first drawing a line between the current waypoint p1 and the next waypoint p2 . If a line orthogonal to this is constructed, it will intersect the safety circle at two points N and M . These are then designated as the potential intermediate waypoints. If the center C is left to the line p1 p2 , the intermediate waypoint M is selected on the right-hand side of the threat region and vice versa. Consider two cases relating to the nonflying zone avoidance.

62.4.2.1 Case 1: The Line Between the Target Waypoints Crosses the NFZ The first case happens when the line between the current waypoint Wi 1 and the next waypoint Wi cuts a nonflying zone as shown in Fig. 62.18.

1536

H. Oh et al.

Define a vector of positions of target waypoints, the intermediate waypoint to be designed, and the nonflying zone center by Wi 1 D Œxi 1 ; yi 1 T

(62.24a)

Wi D Œxi ; yi T

(62.24b)

I

ŒxiI ; yiI T

(62.24c)

Pn D Œxn ; yn T

(62.24d)

Wi D

The following equation of a line li is obtained with a simple geometrical relation between two successive waypoints, Wi 1 , Wi : y  yi D

yi  yi 1 .x  xi / xi  xi 1

(62.25)

To check the violation of the NFZ, the distance from the NFZ center to the line li is calculated as d D

j.yi  yi 1 /.xn  xi /  .xi  xi 1 /.yn  yi /j p .yi  yi 1 /2 C .xi  xi 1 /2

(62.26)

If d < s , a intermediate target waypoint WiI needs to be defined for the UAV to make a detour around the nonflying zone. For this, the tangential line lt is required, and its gradient can be obtained using the following second-order polynomial equation which represents the relationship between the tangential line lt and the safety circle radius s : am2 C bm C c D 0 (62.27) where the coefficients are defined as a D .xn  xi 1 /2  s2

(62.28a)

b D 2.xn  xi 1 /.yi 1  yn /

(62.28b)

c D .yi 1  yn /2  s2

(62.28c)

This equation has two roots, and the root mt closest to the gradient of the line li is chosen. The resulting equation of the line lt is given by y  yi 1 D mt .x  xi 1 /

(62.29)

The equation of the line ln1 is defined using the property that it is perpendicular to the line li and is located at the NFZ center Pn :

62 Cooperative Mission and Path Planning for a Team of UAVs

y  yn D mn1 .x  xn /

1537

(62.30)

I i 1 where mn1 D  yxii x yi 1 . Finally, the intermediate target waypoint Wi is defined by the intersecting point of the lines lt and ln1 . Solving (Eqs. 62.29) and (62.30) together gives

xiI D

xiI D

mt xi 1  yi 1  mn1 xn C yn mt  mn1

(62.31a)

yiI D mt .xiI  xi 1 / C yi 1

(62.31b)

mt xi 1  yi 1  mn1 xn C yn mt  mn1

(62.32a)

yiI D mt .xiI  xi 1 / C yi 1

(62.32b)

62.4.2.2 Case 2: The Next Target Waypoint Exists Inside the NFZ For this case, a replacement target waypoint WiS needs to be defined as the next target waypoint exists inside the NFZ as shown in Fig. 62.19. In a similar manner to case 1 in Sect. 62.4.2.1, the line lt tangential to the NFZ safety circle can be obtained using (Eqs. 62.27)–(62.29). Then, define a position vector of the substitute target waypoint as WiS D ŒxiS ; yiS T (62.33) The equation of the line ln2 is defined using the property that it is perpendicular to the line lt and goes through the next waypoint Wi : y  yi D mn2 .x  xi /

(62.34)

where mn2 D 1=mt . Finally, the replacement target waypoint WiS is defined as the intersection point of lines lt and ln2 . Solving (Eqs. 62.29) and (62.34) together gives

Fig. 62.19 Case 2: The next target waypoint exists inside the NFZ

1538

H. Oh et al.

Fig. 62.20 Decision making concept on speed constraints: An intersection position exists (Left) and does not exist (Right) between the maximum arrival circle of the UAV and the protection line

Friendly Fleet

WFi

Rcom

Wi

WSi Wi -1

VmaxΔt

Next waypoint New Next waypoint

Current waypoint

P GCS

xiS D

mt xi 1  yi 1  mn2 xi C yi mt  mn2

(62.35a)

yiS D mt .xiS  xi 1 / C yi 1

(62.35b)

62.4.3 Speed Constraint of UAV The UAVs have a maximum speed constraint Vmax ; thus, at times they cannot arrive at the next target waypoint Wi in time from the previous waypoint Wi 1 , as shown in Fig. 62.20. In this case, the next target waypoint has to be replaced by a new waypoint WiS . The decision making considers only the case in which there exists an intersection between the maximum arrival circle of the UAV and the communication range of the FF. To check existence of an intersection between the area of the maximum arrival circle and the next waypoint, the distance from the current waypoint Wi to the next waypoint Wi C1 is calculated as d D jWi  Wi C1 j:

(62.36)

62 Cooperative Mission and Path Planning for a Team of UAVs

Decision making on NFZ and speed Waypoint sequence of UAVs Arrival time to the next pose Position of Non-Flying Zones Maximum speed of UAVs

1539

Decision making on orientation and turning radius

Initial pose (current waypoint) Final pose (next waypoint) Initial turning radius Final turning radius

Decision making on basic waypoint sequence

Dubins path planner

Waypoint sequence of FF Speed of FF

Dubins trajectory For communication relay

Fig. 62.21 The overall block diagram of the proposed decision-making and path-planning methodology

The radius of the maximum arrival circle is computed as Vmax t where t is defined j j as tui  tu.i 1/ using (Eq. 62.23). If d > Vmax t, an intersection position does not exist between the area of the maximum arrival circle of the UAV and the next waypoint, as shown in Fig. 62.20. In this case, since the UAV has to go inside the FF’s communication range in order to maintain the communication relay, the new waypoint Wi should be determined by the intersection between the maximum arrival circle of the UAV and the communication range of the FF. The starting and final positions have been defined to generate the Dubins path between the target waypoints of the UAVs. To complete the design of the Dubins path, the orientation also need be designed. The orientation of the UAVs between the i -th and .i C 1/-th waypoints is synchronized with that of the friendly fleet in order to keep the line-of-sight stable from the GCS to the FF as i D tan1 i C1 D tan1

yF i  yF .i 1/ xF i  xF .i 1/ yF .i C1/  yF i xF .i C1/  xF i

(62.37a) (62.37b)

Figure 62.21 shows the overall block diagram of the proposed decision-making and path-planning methodology.

62.4.4 Numerical Simulations 62.4.4.1 Simulation Conditions In order to demonstrate the performance and behavior of the path-planning system, numerical simulations are carried out on a baseline scenario as depicted in Fig. 62.22. The flight plan of the friendly fleet is shown as a bold red (outside

1540

H. Oh et al.

Fig. 62.22 An illustration of basic scenario

x 104 9 8 7 North, m

NFZ 1 6

NFZ 3

5 NFZ 2 4 GCS

3 2 1

UAVs

FF −2

0

2

4

6

East, m

Table 62.2 Simulation parameters Variable Description Rcom P

Communication range Position of GCS

Pn

Positions of NFZ

s

Radius of NFZ

WF

Waypoint sequence of FF

VF ds Vmax nmax ko

Speed of FF Separation distance Maximum speed of UAV Maximum lateral acceleration of UAV Overlap coefficient

8 x 104

Value

Unit

20 Œ30; 30T   25 35 42 60 40 54 3   5 8 25 42 55 42 32 14 7 20 55 87 95 59 39 37 33 20 50 2.5 50 1 0.9

km km km km km m/s km m/s g N/A

the GCS communication range) and blue (inside the GCS communication range) line passing through a series of predefined waypoints. The circumstances are such that the flight of the FF must pass outside of the GCS communication range for a significant period. Figure 62.22 shows that the FF flies up to three times the communication ranges. Three nonflying zones (red dotted circles) are also defined around and outside the GCS communication range. The speed and the flight plan of the FF for the whole mission is known to the UAV members in advance ahead the next waypoint. The UAV group consists of three members (defined by (Eq. 62.16)) whose initial positions are illustrated as triangles. The detailed default parameters can be found in Table 62.2.

62 Cooperative Mission and Path Planning for a Team of UAVs Fig. 62.23 Simulation result on basic scenario with two UAVs: position histories of UAVs (x) and friendly fleet (o) (star: GCS, red dotted circle: NFZ, black dotted circle: GCS coverage)

1541

x 104 9 8

North, m

7 6 5 4 3 2 −2

0

2

4 East, m

9

x 104 d (GCS-FF) dmin (FF-UAVs) dmax (UAVs) dmin (GCS-UAVs)

8 7

Rcom 2Rcom GCS coverage

m

6 5 4

8 x 104

b 1.5 1 indicator

a

6

0.5

3 0

2 1 0

0

500 1000 1500 2000 2500 3000 3500 time, s

−0.5

0

500 1000 1500 2000 2500 3000 3500 time, s

Fig. 62.24 Simulation result on basic scenario with two UAVs: (a) relative distance histories among GCS, UAVs, and friendly fleet and (b) indication on communication relay (1: success, 0: failure)

62.4.4.2 Simulation Results To effectively compare performance, the proposed algorithm is applied to a twoUAV swarm. Figures 62.23 and 62.24 show that the movements of the UAV members fail to maintain the communication relay for about 1,200 s in the middle of the mission because the distance from the GCS to the UAV members exceeds the communication range Rcom , which means the GCS loses the communication with the UAV. This implicates that more than two UAVs should be required to maintain the communication relay between the FF and the GCS.

1542

H. Oh et al.

Fig. 62.25 Simulation result on basic scenario with speed and NFZ constraints: position histories of UAVs (x) and friendly fleet (o) (star: GCS, red dotted circle: NFZ, black dotted circle: GCS coverage)

x 104 9 8

North, m

7 6 5 4 3 2 −2

0

2

4 East, m

9

x 104

8 7 6

m

b d (GCS-FF) dmin (FF-UAVs) dmax (UAVs) dmin (GCS-UAVs) Rcom 2Rcom GCS coverage

5 4 3 2

8 x 104

60 50

speed m/s

a

6

40 30 20 10

1 0

0

500 1000 1500 2000 2500 3000 3500 time, s

0 0

500 1000 1500 2000 2500 3000 3500 4000 time, s

Fig. 62.26 Simulation result on basic scenario with speed and NFZ constraints: (a) relative distance histories among GCS, UAVs, and friendly fleet and (b) speed histories of UAVs (red line: Vmax )

The trajectories of the UAVs avoid entering the NFZ effectively, and at the same time, their speeds do not go over the maximum, as shown in Figs. 62.25 and 62.26. Also, this figure shows that the communication is successfully maintained. It can be shown that the minimum distance from the GCS to the UAVs, the maximum distance between the UAVs, and the minimum distance from the FF to the UAVs are always kept below the communication range Rcom . Figure 62.27 shows snapshots of the communication boundaries of the UAV members, depicted as the red circles, where the communication between the GCS and the friendly fleet is successfully relayed by restructuring the UAV positions in a chain shape.

62 Cooperative Mission and Path Planning for a Team of UAVs x 104

x 104

b

9

9

8

8

7

7 North, m

North, m

a

6 5

6 5

4

4

3

3

2

2 −2 −1

0

1

2 3 4 East, m t=776s

c

5

6

7 x 104

1543

−2

0

2

4 East, m t=1826s

6

8 x 104

x 104 9 8

North, m

7 6 5 4 3 2 −1

0

1

2

3 4 5 East, m t=2331s

6

7

8 x 104

Fig. 62.27 Simulation result on basic scenario with speed and NFZ constraints: snapshots of communication ranges of UAVs. (a) t = 776 s. (b) t = 1,826 s. (c) t = 2,331 s

Conclusions

This chapter has developed a framework enabling cooperative mission and path planning of multiple UAVs in the context of the vehicle-routing problem. The vehicle-routing algorithms often approximate paths of UAVs to straight lines to reduce computational load. In order to resolve this issue, this study focused on integration of the differential geometry concepts, especially Dubins paths, into the vehicle route-planning design procedure to consider non-straight path components. Note that physical and operational constraints of the UAV can be taken into account by these non-straight components of paths. In order to validate this approach, cooperative mission and path-planning algorithms for two missions are developed: road-network search and communication relay. In the road-network search algorithm, the conventional Chinese Postman Problem (CPP) algorithm was firstly explained and modified and applied to the general type of road map which includes unconnected roads. Then, MILP optimization

1544

H. Oh et al.

and the nearest insertion algorithm along with auction negotiation were examined for multiple unmanned aerial vehicles. To realistically accommodate the maneuvering constraints of UAVs, the Dubins path planning was used to solve the modified CPP. The performance of the MILP optimization and the approximation algorithm was then compared for a specific road-network scenario. Particularly, the performance of the approximation algorithm was investigated via a Monte Carlo simulation framework by analyzing the effects of different map sizes, pathplanning methods, and the number of UAVs. Based on these results, the efficient UAV team size and path-planning method were suggested for the road search route planning and hence showed that this framework can be applied to a variety of autonomous search missions in the initial phase of mission planning. In the communication-relay problem, a design methodology on vehicle route planning of multiple UAVs to maintain communication between the ground control station and the friendly fleet was proposed. Dubins theory was applied to get flyable paths, and a decision-making algorithm was developed to satisfy the constraints on the UAV speed and the avoidance of nonflying zones, as well as maintaining communication. The performance of the proposed algorithms is examined via numerical simulations.

References A. Ahmadzadeh, G. Buchman, P. Cheng, A. Jadbabaie, J. Keller, V. Kumar, G. Pappas, Cooperative control of UAVs for search and coverage, in Conference on Unmanned Systems (SPIE, Bellingham, 2006) D. Ahr, Contributions to multiple postmen problems. Ph.D. thesis, Heidelberg University, 2004 B. Alspach, Searching and sweeping graphs: a brief survey. Matematiche (Catania) 59, 5–37 (2006) T. Bektas, The multiple traveling salesman problem: an overview of formulations and solution procedures. Int. J. Manag. Sci. 34(3), 209–219 (2006) J. Bellingham, M. Tillerson, A. Richards, J. How, Multi-Task Allocation and Path Planning for Cooperating UAVs, Cooperative Control: Models, Applications and Algorithms (Kluwer, Dordrecht, 2003) C. Cerasoli, N.J. Eatontown, An analysis of unmanned airborne vehicle relay coverage in urban environments, in Proceedings of MILCOM (IEEE, Piscataway, 2007) P. Chandler, M. Pachter, D. Swaroop, J. Hewlett, S. Rasmussen, C. Schumacher, K. Nygard, Complexity in UAV cooperation control, in American Control Conference, Anchorage (American Control Conference, Evanston, 2002) L.E. Dubins, On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and tangents. Am. J. Math. 79(3), 497–516 (1957) K. Easton, J. Burdick, A coverage algorithm for multi-robot boundary inspection, in IEEE International Conference on Robotics and Automation (IEEE, Piscataway, 2005) A. Gibbons, Algorithmic Graph Theory (Cambridge University Press, Cambridge/New York, 1999) J. Gross, J. Yellen, Handbook of Graph Theory (CRC, Boca Raton, 2003) M. Hifi, M. Michrafy, A. Sbihi, A reactive local search-based algorithm for the multiple-choice multi-dimensional knapsack problem. Comput. Optim. Appl. 33, 271–285 (2006) S. Kim, P. Silson, A. Tsourdos, M. Shanmugavel, Dubins path planning of multiple unmanned airborne vehicles for communication relay. Proc. IMechE G 225, 12–25 (2011)

62 Cooperative Mission and Path Planning for a Team of UAVs

1545

I. Maza, A. Ollero, Multiple UAV cooperative searching operation using polygon area decomposition and efficient coverage algorithms, in Distributed Autonomous Robotic Systems 6, vol. 5 (Springer, Tokyo/New York, 2007), pp. 221–230 N. Nigam, I. Kroo, Persistent surveillance using multiple unmanned air vehicles, in Aerospace Conference, 2008 IEEE, Big Sky (IEEE, Piscataway, 2008) H. Oh, S. Kim, A. Tsourdos, B. White, Cooperative road-network search planning of multiple UAVs using dubins paths, in AIAA Guidance, Navigation and Control Conference, Portland (AIAA, Reston, 2011a) H. Oh, H. Shin, A. Tsourdos, B. White, P. Silson, Coordinated road network search for multiple UAVs using dubins path, in 1st CEAS Specialist Conference on Guidance, Navigation and Control, Munich (Springer, Berlin/Heidelberg, 2011b) H. Oh, S. Kim, A. Tsourdos, B. White, Coordinated road-network search route planning by a team of UAVs. Int. J. Syst. Sci., (2012) In press T. Parsons, Pursuit-evasion in a graph, in Theory and Applications of Graphs (Springer, Berlin, 1976) N. Perrier, A. Langevin, J. Campbell, A survey of models and algorithms for winter road maintenance. Part IV: Vehicle Routing and Fleet Sizing for Plowing and Snow Disposal. Comput. Oper. Res. 34, 258–294 (2007) M.F.J. Pinkney, D. Hampel, S. DiPierro, Unmanned aerial vehicle (uav) communications relay, in IEEE Military Communications Conference, 1996. MILCOM’96, Conference Proceedings (IEEE, Piscataway, 1996) D. Rosenkrantz, R. Stearns, P. Lewis, An analysis of several heuristics for the traveling salesman problem. Fundam. Probl. Comput. 1, 45–69 (2009) J. Royset, H. Sato, Route optimization for multiple searchers. Nav. Res. Logist. 57, 701–717 (2010) J. Ryan, T. Bailey, J. Moore, W. Carlton, Reactive tabu search in unmanned aerial reconnaissance simulations. in 30th Conference on Winter Simulation, Washington, DC, 1998 T. Ralphs, L. Ladanyi, M. Guzelsoy, A. Mahajan, SYMPHONY 5.2.4 (2010), http://projects.coinor.org/SYMPHONY T. Samad, J. Bay, D. Godbole, Network-centric systems for military operations in urban terrain: the role of UAVs. Proc. IEEE 95(1), 92–107 (2007) M. Shanmugavel, Path planning of multiple autonomous vehicles. Ph.D. thesis, Cranfield University, 2007 M. Shanmugavel, A. Tsourdos, B.A. White, R. Zbikowski, Differential geometric path planning of multiple UAVs. J. Dyn. Syst. Meas. Control 129, 620–632 (2007) A. Tsourdos, B. White, M. Shanmugavel, Cooperative Path Planning of Unmanned Aerial Vehicles (Wiley, Chichester, 2010)

Cooperative Task Assignment and Path Planning for Multiple UAVs

63

Sangwoo Moon, David Hyunchul Shim, and Eunmi Oh

Contents 63.1 63.2 63.3 63.4 63.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1548 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1550 Integrated Hierarchical Structure for Multi-UAV Coordination . . . . . . . . . . . . . . . . . . . . . . . . . . 1551 Negotiation-Based Task Assignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1552 Intersection-Based Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1554 63.5.1 Additional Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1556 63.5.2 Overall Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1560 63.6 Potential Field-Based Collision Avoidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1561 63.7 Simulation and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1564 63.7.1 Intersection-Based Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1564 63.7.2 Negotiation-Based Task Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1566 63.8 Flight Experiments and Validations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1568 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1576

Abstract

In this chapter, a hierarchical framework for path planning and task assignment for multiple unmanned aerial vehicles in a dynamic environment is presented. For multi-agent scenarios in dynamic environments, a candidate algorithm should be

S. Moon () Department of Aerospace and Mechanical Engineering, Korea Air Force Academy, Cheongju, South Korea e-mail: [email protected] D.H. Shim Department of Aerospace Engineering, Korean Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea e-mail: [email protected] E. Oh CNS/ATM & Satellite Navigation Research Center, Korea Aerospace Research Institute, Daejeon, South Korea e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 82, © Springer Science+Business Media Dordrecht 2015

1547

1548

S. Moon et al.

able to replan for a new path to perform the given tasks without any collision with obstacles or other agents. The path-planning algorithm proposed here is based on the visibility and shortest-path principles in Euclidean space. Instead of typical visibility graph-based methods that scan through all nodes, A* algorithm is adopted to find an admissible path in a “best-first” approach during the search process. Since the direct outcome from such algorithms may not produce admissible paths in complex environments due to the problems including cul-desac, additional procedures are conceived to find a solution with a lower cost by avoiding local minima and eliminating any redundant nodes. The path planner is augmented with a potential field-based trajectory planner, which solves for a detouring trajectory around other agents or pop-up obstacles. Task assignment is achieved by a negotiation-based algorithm, which assigns a task with the lowest cost to each agent after comparing all task costs of all participating agents. These algorithms are implemented on MATLAB/Simulink, which can run with simulated vehicle models or actual UAVs through a communication network. In the simulations, the algorithms were validated to perform task assignment and path planning flawlessly. In actual flight tests, the proposed algorithms were tested with a number of fixed-wing UAVs in a fully realistic situation under various reality factors such as communication loss or tracking errors. The flight test shows, even in the presence of such uncertainties and logistic factors, the algorithms were able to perform all of the given tasks without any collision with other agents or obstacles.

63.1

Introduction

Unmanned aerial vehicles (UAVs) have been found highly effective in many applications where human presence in the aircraft is deemed unnecessary or dangerous. Furthermore, as the mission becomes more complicated, it has been suggested to perform the mission more effectively by “sharing the burden” among multiple UAVs. In such scenarios, UAVs are required to fly through the designated waypoints associated with tasks in the given mission. When tasks are performed by multiple agents, the question about “who is doing which task in what order” naturally arises. Also, when multiple UAVs fly together in a given area visiting waypoints, planning a safe and efficient path for each agent is very important. Finally, all agents should be able to avoid obstacles such as terrain or buildings by replanning the path dynamically. When planning for a mission of a UAV, the flight paths for all UAVs in the scenario are determined by assigning all tasks to participating agents. When the order of the tasks to be performed by each agent is decided, the path for each agent can be easily determined by simply connecting the waypoints associated with the tasks sequentially. However, when the mission area are filled with obstacles, the original path should be modified to pass through the designated waypoints while staying clear from the forbidden zones, which may be known a priori or suddenly “pop-up.” Also, when the UAVs are required to fly close to the ground, they need

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

1549

to avoid collisions with ground objects or buildings, which are already mapped or detected by onboard sensors such as laser scanners or radars. Assigning tasks to UAVs is not a straightforward problem especially when the UAVs fly in a dynamic environment where obstacles are not fully mapped a priori or when the sensors for obstacle detection have errors. Mixed Integer Linear Programming (MILP) has been found as a powerful method because it can handle dynamic systems with discrete decision variables and there exist many efficient solvers. Bellingham used MILP for task assignment to handle waypoint visiting problems (Bellingham et al. 2001). However, since the complexity of the problem rapidly grows as the number of variables increases, it is often not suitable for realtime applications. Furthermore, since the MILP is formulated with object functions for the entire system, it needs to be implemented as a centralized system, which may not function well when the central control station or communication links are broken (Bertsekas 1991; Ren et al. 2007; Choi et al. 2009; Sujit et al. 2007; Viguria et al. 2010). Path planning is another important problem for efficient and collision-free operations for multiple UAVs. A path planner should satisfy completeness, optimality, computational tractability, and scalability. In addition, collision and threat avoidance is required for multi-agent scenarios and there are many algorithms for this problem. Many researches have been conducted on the path-planning problem since 1960s (Latombe 1991; Lavalle 2006; Hart et al. 1968; Redding et al. 2007; Yang et al. 2010). However, the path-planning problem becomes more complicated when solving for dynamic environments with incompletely known obstacles or pop-up threats. For such cases, the planner should be able to find a conflict-free path in real time. Schouwenaars et al. used mixed integer programming for multi-vehicle path planning (Schouwenaars et al. 2001). Richards showed an aircraft trajectory planning with collision avoidance using MILP (Richards et al. 2003). Also, some variations of MILP have been applied to this problem. Earl and D’Andrea developed an iterative MILP approach (Earl and D’Andrea 2005), and Jain and Grossmann proposed hybrid MILP/CP (constrained programming) models for a class of the optimization problems (Jain and Grossmann 2001). Chang and Shadden introduced the gyroscopic forces for collision avoidance, which is similar to the potential field approach (Chang et al. 2003). More recently, Shim et al. pioneered the dynamic path-planning problem using model predictive control (MPC) algorithm for a collision-avoidance problem of 16 homogeneous rotary UAVs with obstacles (Shim 2008). In this chapter, a hierarchical framework for efficient path planning and task assignment for multiple UAVs in dynamic environments is presented. For task assignment, a negotiation-based algorithm is proposed. In this approach, each agent initially chooses the task with the lowest cost, which is assumed directly proportional to the distance to the waypoint associated with a task. The choices of all other agents are notified to each agent, and when there are conflicts, i.e., when two or more agents choose the same task, their costs are compared, and the agent with the lowest cost are assigned with the task, while the remaining agents repeat the same negotiation process until all agents are assigned with a task without any

1550

S. Moon et al.

conflicts. The proposed algorithms are computationally light enough to handle realtime path planning in partially known environments. The task-assignment problem is essentially tightly coupled with the path planning. A path-planning algorithm based on the shortest-path principle in Euclidean space is chosen in this framework, where obstacles are initially modeled as polygons and a path between the two waypoints around obstacles is found using the visibility and the shortest-path principles in Euclidean space. In order to search for an admissible path without scanning over the entire solution set, A* algorithm is deployed for an effective search. When the solution is found by running the A* algorithm (and not by an exhaustive search), the solution of the planner may converge to a less optimal or even inadmissible path. Therefore, additional filters such as a cul-desac filter, a recursive filter, and a Fibonacci filter are introduced. These filters help to alleviate issues due to local minima while eliminating unnecessary segments in a candidate path. In addition, since the path solution found prior to the mission may have conflicts with pop-up obstacles or other agents, it is locally adjusted with a potential field-based collision-avoidance algorithm, which is responsible for solving a detouring trajectory around the obstacles intersecting with the original path solution. The proposed algorithms are first validated in simulations, where multiple UAVs perform the given tasks while avoiding obstacles and other agents. The proposed algorithm implemented on MATLAB/Simulink is shown to be capable of generating admissible paths in completely/partially known or completely unknown environments. Then the algorithm is validated in a flight test using three fixedwing UAVs to validate the task-assignment and path-planning algorithms together in an environment with pop-up threats. The proposed algorithms were shown to be capable of real-time path planning and task assignment in outdoor environments even with simulated dynamic obstacles.

63.2

Problem Formulation

In the following, the problem for task assignment for multiple UAVs in dynamic environments is defined. Hereinafter, tasks are defined as a set of actions to accomplish a job, a problem, or an assignment. It can be represented as a number of lower-level goals that are necessary for achieving the overall goal of the system (Pettersson and Doherty 2006; Maza et al. 2011). In this chapter, tasks is defined as visiting a waypoint, through which at least one agent should pass once to complete the associated task. First of all, the rules for the task assignment and the path planning are established as below (Moon and Shim 2010): (a) Obstacles are defined as the objects that obstruct the motion. In this research, they are represented as polygons. (b) The environment is modeled as a two-dimensional Euclidean space. (c) UAVs must avoid all obstacles. In other words, all UAVs must remain in admissible areas, clear from obstacles with a margin. Therefore, paths of all UAVs solved by the proposed procedure should be admissible. Path can intersect, but should not lead to collision among UAVs.

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

1551

Task Assignment -Task scheduling and allocation -Cooperation between UAVs -Find a suboptimal solution for task scheduling -Decentralized method

Require

Cost

Path Planning -Obstacle avoidance -Using knowledge on environment -No interactions between UAVs -Searching for an admissible path to reach to a task -Solutions are used as costs in task assignment layer

Collision Avoidance -Low-level trajectory planning -Focus on the interaction between UAVs -Can consider pop-up obstacles -Searching for an admissible trajectory to reach waypoints from the path planning

Fig. 63.1 Integration of task assignment, path planning, and collision avoidance

(e) The task assignment and the path planning should run in real time. In the following, the task-assignment and the path-planning algorithms based on rules listed above are presented.

63.3

Integrated Hierarchical Structure for Multi-UAV Coordination

Figure 63.1 explains how the task-assignment layer, the path-planner layer, and the collision-avoidance layer are integrated altogether. The integration of the task assignment and the path planner is a crucial step for architecting a mission-planning system for multi-agent scenarios (Moon et al. 2013). In this research, the task-

1552

S. Moon et al.

assignment layer needs to communicate with the path planner to compute the cost for task assignment, which is a function of the distance to reach the waypoint. During the task assignment, the path-planning layer is repeatedly queried until a feasible solution for task assignment is found. The next step is, if any unmapped obstacle is detected while flying along the original path, to find a collision-free path based on the original path. It is noted that the path-planner and the collision-avoidance layer may appear to do a similar task of finding a conflict-free path. The crucial difference is that the collisionavoidance layer finds a path locally detouring from the original path computed by the path planner, which only uses the obstacle information known before starting the mission. Therefore, these two algorithms work in a complementary manner to generate a collision-free path efficiently. As can be seen in Fig. 63.1, these two procedures are in a hierarchical structure, and the path-planning algorithm is invoked when the situation is in one of the following three cases: (i) before carrying out the mission, (ii) a new task is allocated on an agent participating in the mission, or (iii) new obstacles or threats suddenly appear, or “pop-up.” The collision-avoidance algorithm is activated when the agent flies to the target point while following the path. For multi-UAV scenarios, the collision avoidance between two or more agents should be considered. The avoidance algorithm presented later in this chapter is used to generate a collisionfree path.

63.4

Negotiation-Based Task Assignment

During the last decade, many algorithms have been developed for task-assignment problems (Bertsekas 1991; Ren et al. 2007; Choi et al. 2009; Viguria et al. 2010). The task-assignment algorithm presented in this chapter is based on the negotiation algorithm (Moon et al. 2013). The proposed method should generate a feasible solution within a reasonable time for real-time applications. In addition, this algorithm is an event trigger-based process, which means that it runs only when an agent sends a message to invoke this algorithm. This is a desirable attribute for decentralized systems running in real time with limited communication bandwidth. Events can be classified into three cases: (i) a mission is given to the system and the agents start to achieve the goal of the mission, (ii) a task in the given mission is accomplished by an agent, and (iii) the given mission is completely finished. If any of these events occurs during the mission, the presented task-assignment process is activated. Upon activation, all agents can be assigned with the different tasks, and this result is dependent on the conditions of the tasks. On the other hand, the costs for negotiation are defined by using (63.1). This formula consists of two parts: the distance from the current location of an agent to the task location and the characteristics of an agent and the assigned task, i.e., the capability set of an agent and the type of the task. The total cost is a linear combination of these two elements, and there are weights assigned for each term:

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

UAV #1 Task Assignment Requirement Before Negotiation

1553

UAV #2

UAV #3

Calculate the cost individually

Calculate the cost individually

Send the message for task assignment Calculate the cost individually

Cost: 40.68 @ Task A Cost: 40.96 @ Task A Cost: 42.09 @ Task C 1st Negotiation Cost: 42.09 @ Task C

Cost: 40.68 @ Task A

Cost: 41.60 @ Task C Cost: 42.09 @ Task C 2nd Negotiation Cost: 40.68 @ Task A Cost: 41.60 @ Task C Cost: 45.76 @ Task B

Fig. 63.2 Negotiation-based task allocation for three UAVs with three tasks

Jtask D wd Jdist C wc Jcharacter

(63.1)

In (63.1), Jdist is the cost related with the distance between the current position of an agent and the location of the given task. If there are multiple tasks with different characteristics, Jcharacter can be given with different costs to affect the behaviors of all agents in this scenario. The negotiation-based task assignment is performed in the following manner. Initially, each agent chooses its first task, which has the lowest cost from the cost list. Here, without loss of generality, the cost associated with the task is the distance to the task. The choices of all agents are broadcast to other agent, and when there are conflicts, i.e., more than one agents choose the same task, the costs of conflicting agents are compared, and the agent with the lowest cost has the right to claim the task, while other agents without assigned tasks repeat the same compare-and-claim process until all conflicts are resolved. In Fig. 63.2, an example with three UAVs is given. At first, each UAV chooses their task and the cost is compared. At the first negotiation, UAVs #1 and #2 choose the same task (task A), and they negotiate to find UAV #1 has a lower cost (40.68 vs. 40.96). So UAV #2 chooses a task other than A, which is task C with cost 41.60. This choice now conflicts with that of UAV #3 with cost 42.09, which is higher than the cost of UAV #2. Therefore, UAV #3 is forced to choose task B with cost 45.76. This approach has no guarantee to converge to the global minimum with an exception that it does converge to the global minimum when the number of the agent

1554

S. Moon et al.

is equal to the number of tasks. However, this algorithm has a low computing loads and suitable for decentralized scenarios with limited communication bandwidth.

63.5

Intersection-Based Path Planning

In Euclidean space, the shortest path between two points is the straight line between them. Denoting these points as xs and xg , the minimum cost function, i.e., the distance of the shortest path, can be represented as (63.2): min J D dist.xs ; xg / D jjxs  xg jj

(63.2)

where dist(•) function is the distance defined in the Euclidean space. However, if there are any obstacles or forbidden region on the line, the cost found by (63.2) is not valid since the path is not admissible. In order to find an admissible path in this situation, the intersection point by the path from (63.2) and the boundary of the obstacle are used (Moon et al. 2013). The first step of the proposed algorithm is to find the intersection point. If an obstacle in the two dimensional space is modeled as a polygon, then the intersection point xm in Fig. 63.3 can be found easily. In such a case, a new path that passes the boundary points of a detected obstacle, xOi1 or xOi 2 should be found. Therefore, these two points are the candidates of waypoints for path planning. The proposed algorithm uses A* approach expressed as (63.3): f .x/ D g.x/ C h.x/ D jjxs  xw jj C jjxg  xw jj

(63.3)

In (63.3), xw is the boundary point of detected edge so that the first term of (63.3) is the distance from the starting point to a waypoint candidate and the other is the distance from the waypoint candidate to the goal point. If the detected line is not the first since the operation started, this equation should be modified to (63.4) because the vehicle is already assigned to a waypoint.

Xg XOi 1

Xm Oi

Fig. 63.3 Intersection point between shortest path and obstacle

Xs

XOi 2

63 Cooperative Task Assignment and Path Planning for Multiple UAVs Table 63.1 Pseudo code for path planning

1555

Algorithm 1 Intersection-based path planning (fundamental) 1: P InitializePath(xs ; xg ) 2: xi xs 3: while Dist(xg ; xi /  dadmissible do 4: xm DrawLine(xg ; xi ) 5: if xm exist SelectOptimalNode(xg ; xs ; x1 ; x2 ; : : : ; xi ) 6: xi 7: P AddPath(P,xi ) 8: end if 9: i i+1 10: end while

f .x/ D g.x/ C h.x/ D

N X

jjxi  x.i 1/ jj C jjxg  xw jj

(63.4)

i D1

In the two-dimensional space, the suboptimal waypoints and the candidate waypoints appear as a pair, and the suboptimal waypoints receive a higher priority for calculation. If multiple obstacles are located in the area, the intersection point which is the closest among the intersection points and the obstacle of that point is considered as the next optimal candidate. As A* algorithm is used in this procedure, all the nodes found by the intersection method are prioritized, and the best solution is chosen using the priority value. Therefore, those nodes related with the intersection points are applied by this method, and the procedures from (63.2) to (63.4) will be performed until the solved node is the goal point (see Table 63.1). When A* algorithm is applied from one root to another, the resulted path is not necessarily optimal. Therefore, a number of additional steps are introduced to refine the path although this process only guarantees to converge to a local minimum. Nonetheless, this process greatly helps to improve the quality of the resulted path, which has much less redundant segments and also avoids getting stuck in a cul-de-sac. Figure 63.4 shows a result obtained using the proposed algorithm. The given environment consists of ten arbitrary-shaped obstacles, which do not have any concave parts that can cause the solver to fall into an infinite loop. Therefore, additional procedures should be added, which will be introduced in the next section. In addition, the configuration space is also considered because UAVs are assumed to occupy certain area (volume), not a point. The dashed line is the solved path, and the lines are the candidate paths to obtain the shortest path. Initially, the algorithm finds a path, which travels to the left side of an obstacle. However, this solution is discarded due to the longer traveling distance around the next obstacle. After several iterations, a new path is found (dashed line), which connects the starting point and the goal point without any conflict with all obstacles.

1556

S. Moon et al. 50 Starting Point 40 30

Y direction (m)

20 10 0 –10 –20 –30 –40 –50

Goal Point –60

–40

–20

0

20

40

60

X direction (m)

Fig. 63.4 Result from the algorithm proposed in Table 63.1

63.5.1 Additional Procedures Although the algorithm proposed in Table 63.1 can generate an admissible path in most cases, it is also possible that the planner may get stuck to local minima. Even if the path is admissible, it may contain extra nodes that make the overall path less optimal. Therefore, in this section, three additional procedures are introduced: cul-de-sac filtering, Fibonacci filtering, and recursive filtering. These additional procedures further improve the proposed algorithm to find the optimal path more reliably although these procedures render the overall algorithm computationally more demanding.

63.5.1.1 Cul-de-sac Filtering Cul-de-sac refers to a dead end where there is no outlet other than the inlet. If the proposed algorithm is applied to the situation with concave obstacles that creates cul-de-sac, the algorithm may fall into an infinite loop, thus unable to find any admissible path. Therefore, an additional filter is introduced to resolve this problem (see Table 63.2). In Fig. 63.5, a case with three concave obstacles is considered. The direct output from the intersection-based planning finds a path that falls into the cul-de-sac point located on the middle of the diagram. Two paths were generated

63 Cooperative Task Assignment and Path Planning for Multiple UAVs Table 63.2 Pseudo code for cul-de-sac filtering

50

1557

Algorithm 2 Cul-de-sac filtering 1: if xi D xi1 2: xm xi 3: xn xi1 4: while xm D xn do 5: xn SelectNeighborVertex(xi / SelectOptimalNode(xg ; xs ; x1 , . . . ,xi1; xn ; xm / 6: xm 7: end while 8: end if 9: xi xm

Starting Point

40 30

Discarded Path

Y direction (m)

20 10 0 –10 Final path

–20 –30 –40 –50

Goal Point –60

–40

–20

0

20

40

60

X direction (m)

Fig. 63.5 Cul-de-sac filtering

after applying the cul-de-sec filtering. When the path on the left is chosen, the other blue path indicated as “Discarded Path” is selected first. However, after running the original algorithm with the cul-de-sec filtering, the dashed red line is generated. Figure 63.5 shows the result after applying the cul-de-sac filtering algorithm.

63.5.1.2 Recursive Filtering As the intersection-based path planning considers only one obstacle at a time, if another obstacle is near the one being considered for path planning, the generated

1558 Table 63.3 Pseudo code for recursive filtering

S. Moon et al.

Algorithm 3 Recursive filtering 1: if CheckOverlapObstacle(O1 ; O2 ,. . . ,On; xi1 ; xi / 2: Ps InitializePath(xi1 ; xi / xi1 3: xj 4: xm DrawLine(xi ; xj / 5: if xm exist SelectOptimalNode(xi ; xi1 ; xj / 6: xi 7: Ps AddPath(Ps , xi / 8: end if 9: j j +1 10: P MergePath(P; Ps / 11: end if

path may conflict with other obstacles. So the recursive filtering is proposed, which checks if the path around one obstacle passes through any other obstacles. This filtering process checks the path between i -th and the previous node, which can be expressed as N

xi 1 xi \ [ Oj ¤ ; j D1

(63.5)

Equation (63.5) implies that the line segment between the previous node xi 1 and the current node xi intersects with any obstacles along the line. In such cases, the intersection-based path-planning algorithm is called with xi 1 and xi as the input argument for temporary starting and the goal points, respectively, find to a conflict-free path (see Table 63.3). Recursive filtering is performed after running the main planning algorithm and the cul-de-sac filtering introduced above. This filtering is frequently required when the given environment is filled with closely located obstacles and consumes a more computation time. Figure 63.6 shows an example of the recursive filtering. Since the line between the starting point and the goal point intersects with the obstacle on the right, the blue path is obtained by the intersection-based algorithm. Since it overlaps with the obstacle on the left, the recursive filtering is applied for the segment between the starting point and the first node, resulting in the red-dashed line, which is conflict-free.

63.5.1.3 Fibonacci Filtering Fibonacci filtering is named after the well-known Fibonacci sequence from the fact that the current waypoint is computed using the two preceding points. It is developed to eliminate any extra nodes that increase overall cost (i.e., a longer path) when there is an admissible path with a lower cost. In Fibonacci sequence, each number is the sum of the previous two numbers, starting with 0 and 1. Similarly, Fibonacci filtering takes only the previous two nodes as the input argument. If two nodes xi 2 and xi satisfy (63.6), this filtering moves on to the next node. Therefore, Fibonacci filtering is the opposite concept of the recursive filtering described in the previous subsection:

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

1559

50 40

Goal point

30

Y direction (m)

20 10 0 –10 –20 –30 Starting point

–40 –50 –60

–40

–20

0 X direction (m)

20

40

60

Fig. 63.6 Recursive filtering. Solid lines (blue): suboptimal paths before the filtering. Dashed line (red): optimal path obtained by the recursive filtering

Table 63.4 Pseudo code for Fibonacci filtering

Algorithm 4 Fibonacci filtering 1: if !CheckOverlapObstacle(O1 ; O2 ,. . . ,On ; xi2 ; xi / 2: xj 1 xi 3: i i 1 4: end if

N

xi 2 xi \ [ Oj D ; j D1

(63.6)

In (63.6), the line that connects the point before the previous node xi 2 and present node xi does not overlap with any obstacles, then the previous point xi 1 should be cancelled out to straighten the generated path. This procedure is performed for all nodes that satisfy (63.6), and repeated until (63.6) is not satisfied for all nodes (see Table 63.4). As is the case with the additional filtering algorithms introduced above, this filtering is more frequently needed if the given environment is complex and more computation time is needed for this algorithm along with other additional filtering algorithms mentioned above. In Fig. 63.7, the first two segments

1560

S. Moon et al. 50 40

Goal point

30

Y direction (m)

20 10 0 –10 –20 –30 –40 –50

Starting point –60

–40

–20

0 X direction (m)

20

40

60

Fig. 63.7 Fibonacci filtering. Solid line (blue): suboptimal paths before the filtering. Dashed line (red): optimal path obtained by the Fibonacci filtering

from the starting point on the blue line are less optimal than the first line segment of the red-dashed line due to the extra node, which is eliminated by running the Fibonacci filtering.

63.5.2 Overall Procedure Table 63.5 is the pseudo code for the all algorithms proposed so far. When the path planning is started, the intersection-based method is first applied and an initial candidate path is found. At this step, the cost function of the candidate paths are computed to run the A* algorithm to find the path with the least cost. Here, the path may get stuck in cul-de-sac, which is avoided by the proposed filtering (recursive filtering). In this process, the path may have extra node, which is then eliminated by the Fibonacci filtering. The filtering procedures are performed after the intersection-based algorithm. When the path reaches to the goal point, this algorithm is terminated. Figure 63.8 shows the path planning result with the recursive filtering and Fibonacci filtering in addition to the intersection-based method. Here, the recursive filtering process was responsible for finding the optimal path passing by obstacles 1

63 Cooperative Task Assignment and Path Planning for Multiple UAVs Table 63.5 Pseudo code for iterated algorithm

1561

Algorithm 5 Intersection-based path planning (overall) 1: P InitializePath(xs ; xg / 2: xi xs 3: while Dist(xg ; xi / > dadmissible do 4: xm DrawLine(xg ; xi / 5: if xm exist SelectOptimalNode (xg ; xs ; x1 ; x2 ,. . . ,xi / 6: xi 7: if xi D xi1 8: xi CulDeSacFiltering (xg ; xs ; x1 ; x2 ,. . . ,xi / 9: end if 10: if CheckOverlapObstacle (O1 ,. . . ,On; xi1 ; xi / 11: P,xi RecursiveFiltering (xg ; xs ; x1 ,. . . ,xi / 12: end if 13: if !CheckOverlapObstacle (O1 ,. . . ,On; xi2 ; xi / 14: xj 1 xi 15: i i 1 16: end if 17: P AddPath (P , xi / 18: end if 19: i i +1 20: end while

and 2. The Fibonacci filtering is responsible for finding the better (i.e., shorter) path among the three paths ending up near obstacle #3.

63.6

Potential Field-Based Collision Avoidance

Collision avoidance is a dynamic path generation technique, which finds an admissible path to avoid a collision among agents or with obstacles in the environment. The algorithm proposed here is based on the potential field approach with several improvements presented below (Moon et al. 2013). Most algorithms based on the potential field are based on the distance between the vehicle and target points at one time frame. However, in case of moving obstacles, it is desirable to consider the relative direction of motion as well. The proposed algorithm utilizes the cost function for the potential field as the function of the distance and direction of the obstacle using the normal vector as J D f n.x; v; n/

(63.7)

where x and v represent the position and velocity of the vehicle, respectively. n is the set of normal vectors of the planes on the polyhedra representing the obstacles in the environment. Figure 63.9 illustrates the relationship among a UAV, an obstacle, and a waypoint along with the current and future position and velocity vectors of a UAV and a normal vector ni .

1562

S. Moon et al. 50 40

Starting Point

30 20 Y direction (m)

1 2

10 0 –10 –20

3

–30

Goal Point

–40 –50

–60

–40

–20

0

20

40

60

X direction (m)

Fig. 63.8 Result after running recursive filtering and Fibonacci filtering. Solid lines: suboptimal paths before the additional filterings. Dashed line: optimal path selected by the filtering

Xwaypoint

qobstacle + p ni

Xobstacle

Vt+Δt

Xt

Fig. 63.9 Variables for cost function

Xt+Δt Vt

In this algorithm, a set of path candidates over a finite horizon into the future are constructed. For each path, the cost is computed with respect to the attraction forces from the waypoints or goal points and the repulsion forces from the agents. Among those candidates, the best candidate that has the lowest cost can be selected, and the

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

1563

Fig. 63.10 Cost function for waypoint tracking 0.8

cost

0.6 0.4 0.2 0 6 4 2 θ (radian)

0

0

10

5

15

20

distance (meter)

UAV moves in one step along the selected path. This procedure is iterated until the UAV reaches to the target point. The cost function for the waypoints generated from the path planning can be expressed as (63.8), which can be visualized as shown in Fig. 63.10. Note that (63.8) is not dependent on the vehicle’s heading angle. In order to prevent (63.8) from dominating other terms, a denominator that divides by the distance from the waypoint to the current position of an UAV is included: Jwaypoint D 1  exp

jjxwaypoint  xt Ct jj

!

jjxwaypoint  xt jj

(63.8)

The cost function for the obstacles and threats is expressed as (63.9), which is visualized as in Fig. 63.11, and again it is independent from the vehicle’s heading. If the angle between the normal vector of the plane on an obstacle and the vehicle’s velocity vector is very small, this indicates that the vehicle approaches to the object and therefore the cost function grows exponentially. However, if this angle is close to 180ı, the cost function decreases rapidly because the UAV moves away from the object. The cost function in Eq. (63.9) also includes a denominator to scale the magnitude of the normal vector is included to divide the term by a magnitude vector from the object to the current position of an UAV:   jjni jj Jobstacle;i D .0:5 cos obstacle C 0:5 C Cobstacle / exp  jjxt x obstable jj where obstacle D cos1



ni vt Ct jjni jjjjvt Ct jj



(63.9) 

1564

S. Moon et al.

1.5

cost

1

0.5

0 6 20

4

θ (rad)

15 10

2 5 0

0

distance (meter)

Fig. 63.11 Cost function for obstacles and threats

63.7

Simulation and Analysis

In order to evaluate the performance of the proposed algorithm, a series of simulations are conducted. In the proposed hierarchy, since the task-assignment layer cannot function without path planning, the intersection-based path planner is first implemented and validated.

63.7.1 Intersection-Based Path Planning It is assumed that the UAV has an onboard sensor for detecting obstacles such as the laser scanner, which has finite detection range. In the simulation, 20 randomly shaped obstacles are placed where only ten of them are known a priori. The simulation was constructed in MATLAB running on a PC with Intel CPU Centrino 2 Duo 2.4 GHz. Figure 63.12 shows the path-planning result when all of the obstacles are known a priori. UAV 1 flies to the goal point without any conflict with obstacles. Whenever obstacles unknown at the initial path-planning phase are discovered during the flight, the path-planning algorithm runs repeatedly in real time to find a conflict-free path (left side of Fig. 63.13). If none of the obstacles were known a priori, the pathplanning algorithm runs in real time from the start until the end, building the obstacle

63 Cooperative Task Assignment and Path Planning for Multiple UAVs Fig. 63.12 Path planning in totally known environment

1565

50 UAV 1 Start

40

Y direction (m)

30 20 10 0 –10 –20 –30 UAV 1 Goal 60

–40 –50

–60

–40

50

40 30

20

20

Y direction (m)

30

10 0 –10 –20 –30 UAV 1 Goal

–40 –50

40

50 UAV 1 Start

40

Y direction (m)

–20 0 20 X direction (m)

– 60

– 40

– 20

0

20

X direction (m)

40

60

UAV 1 Start

10 0

– 10 – 20 – 30 – 40 – 50

UAV 1 Goal

– 60

– 40

– 20

0

20

40

60

X direction (m)

Fig. 63.13 Path planning in partially known environment (left) and totally unknown environment (right)

map on the fly using onboard sensors (right side of Fig.63.13). Note that the obtained paths in the three cases with identical obstacles but with different sensing methods are different from one another. It is generally agreed that finding a feasible path in a given environment is a hard problem and finding the optimal solution, whichever the definition of optimality is, is a difficult and there is not a general solution for such work. The proposed algorithm is viable in the sense that it can find an admissible path with a smaller computation load with all obstacles modeled as polygons. Such simplified obstacle is acceptable with path planning for UAVs, which needs to stay clear from obstacles for safety and to avoid turbulence caused by the airflow bouncing back from the obstacles. The proposed algorithm runs very fast in environments with fewer obstacles. However, if the environment is filled with many obstacles, the computation load can be quite heavy because of the additional filtering processes.

1566

S. Moon et al.

0.25 Totally Known Environment Partially Known Environment Computation Time (sec)

0.2

0.15

0.1

0.05

0 0

10

20

30

40 50 60 Operation Time (sec)

70

80

90

100

Fig. 63.14 Computation time versus elapsed time

Figure 63.14 shows the computation time during the simulation. The case with totally known environment is the blue lines with square marks, and the case with partially known environment is the red lines with plus marks. The totally known environment case does not require much computation except for the first path planning. However, the partially or totally unknown case requires more computation load during the simulation whenever new obstacles are detected (t = 53 s).

63.7.2 Negotiation-Based Task Assignment In the simulation, three UAVs carry out a mission in an area with ten randomly shaped obstacles. Ten task points are arbitrarily placed in the field. As mentioned above, a task is defined as a visit to its associated waypoint. The margin for obstacle avoidance is assumed 2 m here. In this simulation, a simple kinematic model for UAVs were used such that Vx D V cos  (63.10) Vy D V sin  where

P 2 Œı; ı

(63.11)

To evaluate the performance of the proposed algorithm, a set of simulation is conducted. A scenario with three UAVs carrying out a mission in an area

63 Cooperative Task Assignment and Path Planning for Multiple UAVs 50

50 40

40

TASK 10

UAV 3 Start

10

TASK 1

UAV 1 Goal

TASK 2 TASK 9

0

TASK 3 TASK 8

–10 UAV 2 Start

–20

UAV 2 Goal

TASK 4

UAV 3 Start

10

TASK 1

TASK 9 TASK 3 TASK 8

UAV 2 Start

UAV 3 Goal

TASK 7

–40

–50

UAV 1 Start

TASK 7

–40

–20

0

20

40

60

–60

–40

–20

X direction (m)

UAV 3 Goal

0

20

40

60

X direction (m)

50

50

40

40

TASK 10

TASK 10

30

30 UAV 3 Start

10

TASK 1

UAV 1 Goal

TASK 2 TASK 9 TASK 3

0 TASK 8

–10 UAV 2 Start

–20

UAV 2 Goal

TASK 4

TASK 6

–30

UAV 3 Start

10

UAV 3 Goal

TASK 7

–50

TASK 1

TASK 2 TASK 9

TASK 8 UAV 2 Start UAV 1 Start

UAV 2 Goal

TASK 4

TASK 6

–30 –40

UAV 1 Goal

TASK 3

0 –10 –20

TASK 5

UAV 1 Start

20 Y direction (m)

20 Y direction (m)

TASK 5

–50 –60

–40

UAV 2 Goal

TASK 4

TASK 6

–30

TASK 5

UAV 1 Goal

TASK 2

0 –10 –20

TASK 6

–30 UAV 1 Start

20 Y direction (m)

20 Y direction (m)

TASK 10

30

30

–40

1567

TASK 7

TASK 5

UAV 3 Goal

–50 –60

–40

–20

0 X direction (m)

20

40

60

–60

–40

–20

0

20

40

60

X direction (m)

Fig. 63.15 Task assignment in a totally known environment (t = 10,60,120, and 185 s from left side of top to right side of bottom)

with ten randomly shaped obstacles is considered. Ten task points are arbitrarily placed in the field. As mentioned, tasks are defined as a visit to a waypoint. The margin for obstacle avoidance is 2 m. Simple kinematic equations for UAVs were used in this simulation. Figure 63.15 shows the simulation result. As the mission progresses, ten negotiations occur. UAV 1 is assigned with three tasks (task 4, task 7, and task 9) and participates in the whole negotiation process. UAV #2 is also assigned with three tasks (task 5, task 6, and task 8), and UAV #3 is assigned with four tasks (task 1, task 2, task 3, and task 10) during the mission. At fifth negotiation, UAV #1 sends the information of the determined task to UAV #2 and chooses to undertake task 4 because it is better that UAV #2 carries out task 5. Such a task swapping also occurs at the ninth negotiation process between UAV #1 and UAV #2 (see Table 63.6). There are nine peaks in the plot of the elapsed time for negotiation process (Fig. 63.16). Although the maximum computation time (=0.9 s) of the first process is greater than the iterated time to run the proposed algorithm, it does not affect the real-time performance of the while mission because the task-assignment algorithm is executed offline prior to the mission. Since the computation time to perform each negotiation in the iterative algorithm is very short, the proposed task-assignment algorithm is a viable solution for real-time application. It is also noted that the

1568

S. Moon et al.

Table 63.6 Task-assignment procedure for simulation with three UAVs

UAV #1 Task 7 Task 7 Task 7 Task 5 Task 4 Task 4 Task 4 Task 4 Task 9 Task 9

Step 1 Step 2 Step 3 Step 4 Step 5 Step 6 Step 7 Step 8 Step 9 Step 10

UAV #2 Task 8 Task 8 Task 6 Task 6 Task 5 Task 5 Task 5 Task 9 Finish Finish

UAV #3 Task 1 Task 2 Task 2 Task 2 Task 2 Task 3 Task 9 Task 10 Task 10 Finish

0.9 0.8

Computation Time (sec)

0.7 0.6 0.5 0.4 0.3 0.2 0.1

Fig. 63.16 Elapsed time versus computation time

0 0

20

40

60

80

100

120

140

160

180

200

Elapsed Time under Mission (sec)

proposed negotiation-based algorithm exhibits a greedy behavior as each agent thrives to minimize its cost at each time frame in a fully decentralized manner so that the overall task-assignment algorithm generally produces suboptimal results than centralized, non-greedy algorithms.

63.8

Flight Experiments and Validations

As the final step, in order to validate the proposed algorithm in a realistic environment, actual flight tests were conducted using a fixed-wing UAV based on a blended wing body (BWB) airframe as shown in Fig. 63.17. The airplane is powered by an electrical motor and a lightweight in-house flight control computer is installed for automatic waypoint navigation. The test UAV’s specification is given in Table 63.7. The experiment is conducted in the following steps. When the flight computer is initialized, each vehicle is launched using a bungee cord. After the vehicle climbs to

63 Cooperative Task Assignment and Path Planning for Multiple UAVs

APC 12×6 Prop. Axi 2217/16 Motor

Eleven-left

1569

Telemetry Antenna

Servo ESC Elevon-right

Video Antenna

Servo

GPS Antenna Nose Camera

Fig. 63.17 Testbed UAV based on BWB airplane Table 63.7 Testbed UAV specification

Base platform Dimensions

Weight Powerplant

Operation time Avionics

Operation Autonomy

Blended wing body Wing span: 1.52 m(W) Wing area: 0.52 m2 Aspect Ratio: 4.43 2.6 kg (fully instrumented) Axi 2217/16 DC brushless motor Lithium-ion-polymer (11.1 V 5,300 mAH) 60° q1=60° q1 0. A modified version of the Lin-Kernighan heuristic (Lin and Kernighan 1973) is implemented in linkern; this powerful solver yields approximations in the order of 5 % of the optimal tour cost very quickly for many instances. For example, in numerical experiments on a 2.4 GHz Pentium machine, approximations of random TSPs with 1,000 points typically required about 2 s of CPU time. Both concorde and linkern are written in ANSI C and, at the time of writing, are freely available for academic research use at http://www.tsp.gatech.edu/concorde/index.html. In this chapter, several routing policies were presented requiring on-line solutions of large TSPs. Practical implementations of the algorithms will rely on heuristics, such as Lin-Kernighan’s or Christofides’. If a constant-factor approximation algorithm is used, the effect on the asymptotic performance guarantees of our algorithms can be simply modeled as a scaling of the constant ˇd .

References P.K. Agarwal, M. Sharir, Efficient algorithms for geometric optimization. ACM Comput. Surv. 30(4), 412–458 (1998) M. Alighanbari, J.P. How, A robust approach to the UAV task assignment problem. Int. J. Robust Nonlinear Control 18(2), 118–134 (2008) D. Applegate, R. Bixby, V. Chv´atal, W. Cook, On the solution of traveling salesman problems, in Proceedings of the International Congress of Mathematicians, Extra Volume ICM III, Berlin, 1998, pp. 645–656. Documenta Mathematica, Journal der Deutschen MathematikerVereinigung S. Arora, Nearly linear time approximation scheme for Euclidean TSP and other geometric problems, in Proceedings of the 38th IEEE Annual Symposium on Foundations of Computer Science, Miami Beach, 1997, pp. 554–563 G. Arslan, J.R. Marden, J.S. Shamma, Autonomous vehicle-target assignment: a game theoretic formulation. ASME J. Dyn. Syst. Meas. Control 129(5), 584–596 (2007) R.W. Beard, T.W. McLain, M.A. Goodrich, E.P. Anderson, Coordinated target assignment and intercept for unmanned air vehicles. IEEE Trans. Robot. Autom. 18(6), 911–922, 2002 J. Beardwood, J. Halton, J. Hammersley, The shortest path through many points. Proc. Camb. Philos. Soc. Math. Phys. Sci. 55(4), 299–327 (1959) D.J. Bertsimas, G.J. van Ryzin, A stochastic and dynamic vehicle routing problem in the Euclidean plane. Oper. Res. 39, 601–615 (1991) D.J. Bertsimas, G.J. van Ryzin, Stochastic and dynamic vehicle routing in the Euclidean plane with multiple capacitated vehicles. Oper. Res. 41(1), 60–76 (1993a)

2108

J.J. Enright et al.

D.J. Bertsimas, G.J. van Ryzin, Stochastic and dynamic vehicle routing with general interarrival and service time distributions. Adv. Appl. Probab. 25, 947–978 (1993b) F. Bullo, E. Frazzoli, M. Pavone, K. Savla, S.L. Smith, Dynamic vehicle routing for robotic systems. Proc. IEEE 99(9), 1482–1504 (2011) G. Cannata, A. Sgorbissa, A minimalist algorithm for multirobot continuous coverage. IEEE Trans. Robot. 27(2), 297–312 (2011) P. Chandler, S. Rasmussen, M. Pachter, UAV cooperative path planning, in AIAA Conference on Guidance, Navigation, and Control, Denver (2000) N. Christofides, Bounds for the travelling-salesman problem. Oper. Res. 20, 1044–1056 (1972) Z. Drezner (ed.), Facility Location: A Survey of Applications and Methods. Series in Operations Research (Springer, New York, 1995). ISBN:0-387-94545-8 L.E. Dubins, On curves of minimal length with a constraint on average curvature and with prescribed initial and terminal positions and tangents. Am. J. Math. 79, 497–516 (1957) J.J. Enright, K. Savla, E. Frazzoli, F. Bullo, Stochastic and dynamic routing problems for multiple UAVs. AIAA J. Guid. Control Dyn. 32(4), 1152–1166 (2009) B. Golden, S. Raghavan, E. Wasil, The Vehicle Routing Problem: Latest Advances and New Challenges. Operations Research/Computer Science Interfaces, vol. 43 (Springer, New York/London, 2008). ISBN:0387777776 S. Irani, X. Lu, A. Regan, On-line algorithms for the dynamic traveling repair problem. J. Sched. 7(3), 243–258 (2004) P. Jaillet, M.R. Wagner, Online routing problems: value of advanced information and improved competitive ratios. Transp. Sci. 40(2), 200–210 (2006) D.S. Johnson, L.A. McGeoch, E.E. Rothberg, Asymptotic experimental analysis for the heldkarp traveling salesman bound, in Proceedings of the 7th Annual ACM-SIAM Symposium on Discrete Algorithms, Atlanta, 1996, pp. 341–350 S.O. Krumke, W.E. de Paepe, D. Poensgen, L. Stougie, News from the online traveling repairman. Theor. Comput. Sci. 295(1–3), 279–294 (2003) R. Larson, A. Odoni, Urban Operations Research (Prentice Hall, Englewood Cliffs, 1981) S. Lin, B.W. Kernighan, An effective heuristic algorithm for the traveling-salesman problem. Oper. Res. 21, 498–516 (1973) J.D.C. Little, A proof for the queuing formula: L = W. Oper. Res. 9(3), 383–387 (1961). ISSN:0030364X. http://www.jstor.org/stable/167570 G. Mathew, I. Mezic, Spectral multiscale coverage: a uniform coverage algorithm for mobile sensor networks, in Proceedings of the 48th IEEE Control and Decision Conference, Shanghai, 2009, pp. 7872–7877 N. Megiddo, K.J. Supowit, On the complexity of some common geometric location problems. SIAM J. Comput. 13(1), 182–196 (1984). ISSN:0097–5397 A.R. Mesquita, Exploiting Stochasticity in Multi-agent Systems. PhD thesis, University of California at Santa Barbara, Santa Barbara, 2010 B.J. Moore, K.M. Passino, Distributed task assignment for mobile agents. IEEE Trans. Autom. Control 52(4), 749–753 (2007) C.H. Papadimitriou, Worst-case and probabilistic analysis of a geometric location problem. SIAM J. Comput. 10(3), 542–557 (1981) M. Pavone, Dynamic vehicle routing for Robotic networks. PhD thesis, Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, 2010 M. Pavone, K. Savla, E. Frazzoli, Sharing the load. IEEE Robot. Autom. Mag. 16(2), 52–61 (2009) M. Pavone, A. Arsie, E. Frazzoli, F. Bullo, Distributed algorithms for environment partitioning in mobile robotic networks. IEEE Trans. Autom. Control 56(8), 1834–1848 (2011) G. Percus, O.C. Martin, Finite size and dimensional dependence of the Euclidean traveling salesman problem. Phys. Rev. Lett. 76(8), 1188–1191 (1996) H.N. Psaraftis, Dynamic programming solution to the single vehicle many-to-many immediate request dial-a-ride problem. Transp. Sci. 14(2), 130–154 (1980) D.D. Sleator, R.E. Tarjan, Amortized efficiency of list update and paging rules. Commun. ACM 28(2), 202–208 (1985)

85 UAV Routing and Coordination in Stochastic, Dynamic Environments

2109

S.L. Smith, F. Bullo, Monotonic target assignment for robotic networks. IEEE Trans. Autom. Control 54(9), 2042–2057 (2009) D. Song, C.Y. Kim, J. Yi, Stochastic modeling of the expected time to search for an intermittent signal source under a limited sensing range, in Proceedings of Robotics: Science and Systems, Zaragoza, 2010 J.M. Steele, Probabilistic and worst case analyses of classical problems of combinatorial optimization in Euclidean space. Math. Oper. Res. 15(4), 749 (1990) L.D. Stone, Theory of Optimal Search (Academic, New York, 1975) P. Toth, D. Vigo (eds.), The Vehicle Routing Problem. Monographs on Discrete Mathematics and Applications (SIAM, Philadelphia, 2001). ISBN:0898715792 P. Van Hentenryck, R. Bent, E. Upfal, Online stochastic optimization under time constraints. Ann. Oper. Res. 177(1), 151–183 (2009) E. Zemel, Probabilistic analysis of geometric location problems. Ann. Oper. Res. 1(3), 215–238 (1984)

Section XVI UAV Integration into the National Airspace Konstantinos Dalamagkidis and Richard S. Stansbury

86

UAV Integration into the National Airspace: Introduction Kimon P. Valavanis and George J. Vachtsevanos

UAV Integration into the National Airspace reviews and presents policies, procedures, and requirements for manned aviation and relates them to existing and under development equivalent policies and requirements for unmanned aviation. Integration of unmanned aviation into the (any) national airspace requires, again, that “UAVs function as if there were a human pilot onboard.” This requirement and restriction dictates that equivalent levels of safety (ELOS) must be derived for unmanned aviation (depending on the UAV classification), in addition to reliability and fault tolerance requirements. The major deviation from manned aviation requirements is that a controlled crash may be allowed for UAVs provided that human lives are not endangered and that collateral damage is minimized. Aviation Regulation by Dalamagkidis presents a brief introduction to aviation regulation. Key terms are defined and an overview of current manned aviation regulations is provided using the U.S. Federal Aviation Regulation as an example. This includes airworthiness certification, operation rules, and airspace classes. Human Factors of Unmanned Aircraft System Integration in the National Airspace System by Kaliardos and Lyall focuses on identifying human factor challenges to integrating UAS in the National Airspace System (NAS) for both pilots and air traffic controllers. The method for identifying these challenges is primarily based on the differences or “gaps” between manned aircraft in the NAS today and the unique aspects of UAS. The goal is not to generate a comprehensive

K.P. Valavanis () John Evans Professor and Chair, Department of Electrical and Computer Engineering, Daniel Felix Ritchie School of Engineering and Computer Science, University of Denver, Denver, CO, USA e-mail: [email protected]; [email protected] G.J. Vachtsevanos Professor Emeritus, School of Electrical and Computer Engineering, The Georgia Institute of Technology, Atlanta, GA, USA e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 147, © Springer Science+Business Media Dordrecht 2015

2113

2114

K.P. Valavanis and G.J. Vachtsevanos

list of human factor issues, but to focus on those that are traceable to fundamental characteristics of UAS and that are also considered challenging with respect to the current NAS and its regulatory framework. It is assumed that for integration into the NAS, future UAS will be much more standardized in many respects than in the past – closer to that of today’s civil manned aircraft, which are subject to rigorous design, operational, manufacturing, and training approvals by the FAA. Under such assumptions, human factors are not only central to the very definition of UAS but are also central to some of the most important challenges facing the integration of UAS in the NAS. Methodologies for Regulatory Compliance and Harmonization by Marshall proposes a methodology for analysis of current regulations for applicability to UAS activities. Where the rules are in development, or merely being contemplated, a structure for harmonization with international standards is proposed. The focus is on the USA as the nation that has the most comprehensive aviation regulations, while also supporting major efforts to promulgate regulations and standards to supplement those regulations with UAS specific requirements and criteria. To provide global context, the International Civil Aviation Organization’s (ICAO’s) structure and procedures are reviewed in detail and contrasted with state-sponsored regulatory processes, with an analysis of the applicability of ICAO’s standards to UAS operations. A Certification Strategy for Small Unmanned Aircraft Performing Nomadic Missions in the U.S. National Airspace System by Stachura, Elston, Argrow, Frew, and Dixon discusses the specifics of the Certificates of Authorization (COA) obtained for the second Verification of the Origin of Rotation in Tornadoes Experiment (VORTEX2) project and how the operations are conducted to satisfy the COA requirements. A strategy is outlined for operating these nomadic missions with small UAS within the confines of FAA regulations. This includes information on getting FAA COAs for a large area, specifically focusing on area selection, airworthiness, and emergency procedures, which are the keys to these applications. Hazard and Safety Risk Modeling by Dalamagkidis presents aspects of risk modeling with a focus on UAS. It provides an overview of the current level of safety of manned aviation in terms of accident statistics. These are then mapped as target levels for UAS under the “Equivalent Level of Safety” principle to provide a glimpse at what that may entail for UAS regulations. Different methodologies are presented for estimating the risk of ground impact and midair collision accidents and how these estimates can be translated to system requirements. Guidelines are provided on the use of different risk models followed by applying a selection of them to five different UAS in two distinct scenarios, to compare the results of different choices. Safety Risk Management of Unmanned Aircraft Systems by Clothier and Walker provides existing risk practitioners with a high-level introduction to some of the unique issues and challenges in the application of the safety risk management process to UAS. The scope is limited to safety risks associated with the operation of unmanned aircraft in the civil airspace system and over inhabited areas. This chapter notes the unique aspects associated with the application of the safety risk management process to UAS compared to that of conventionally piloted aircraft.

86 UAV Integration into the National Airspace: Introduction

2115

Key challenges discussed include the specification of high-level safety criteria; the identification, analysis, and evaluation of the risks; and the effectiveness of available technical and operational mitigation strategies. Some solutions to these challenges are examined including those currently in practice and those still under research and development. Certification of Small UAS by Leijgraaf describes certification procedures for (small) UAS. It focuses on the certification process, the requirements for the safe design of a UAS, and the organizational requirements for the company designing the UAS. Technology Surveys and Regulatory Gap Analyses of UAS Subsystems Toward Access to the NAS by Stansbury and Wilson provides the necessary details on how to conduct a technology survey and regulatory gap analysis of UAS technology subsystems. Four past studies performed by Embry-Riddle Aeronautical University for the FAA’s William J. Hughes Technology Center are discussed. These studies address UAS propulsion systems, sense and avoid technologies and procedures, command control and communication, and emergency recovery and flight termination systems. A recommended process for future studies is provided. Concept of Operations of Small Unmanned Aerial Systems: Basis for Airworthiness Towards Personal Remote Sensing by Stark, Coopmans, and Chen discusses the challenges of UAS integration into the NAS for a class of small UAS for personal remote sensing (PRS). This approach is centered on the three-pillared foundations for NAS integration presented by the U.S. DoD Unmanned Aircraft System Airspace Integration Plan and focuses on the specific challenges that are unique to the PRS UAS platform. This chapter presents a concept of operations for these unmanned aircrafts along with an application scenario example and concludes with a discussion of the future use of UAS in the NAS and how it is achievable through adequate regulations and standards. Standards and Certification of a UAS Sense and Avoid Capability by Zeitlin centers on “certifiable” and “standardized” UAV sense and avoid (SAA) capabilities that are needed to mitigate for the lack of a remote pilot’s ability to “see and avoid” other aircraft, which is an operational requirement. Regulators certify the airworthiness of aircraft, assuring their compliance with applicable regulations. Approvals of SAA implementations constitute part of that process. Each applicant needs to develop a Project Specific Certification Plan in close coordination with the regulator. Various tools for safety analysis and configuration control support this effort. Specification of the certification basis for SAA needs close attention, since the system addresses functions traditionally allocated to an onboard pilot. Algorithms and software will receive scrutiny due to their safety roles. Standards are developed to capture common requirements in support of systems that may use some degree of unique design innovation. A standard should suggest means of demonstrating compliance to its requirements, and so doing will greatly simplify the burden for both applicant and regulator. The closest existing standard to SAA is that for the Traffic Alert and Collision Avoidance System II (TCAS II), but SAA includes the additional function of self-separation. SAA can be implemented with various technologies in a variety of architectures. SAA should provide two

2116

K.P. Valavanis and G.J. Vachtsevanos

basic functions, each performing prescribed subfunctions. Considerable modeling and simulation will be required to validate system performance and support the safety case. The regulator may offer guidance to applicants through publication of an Advisory Circular. Collectively, this section provides nothing but a “glimpse” of challenges and issues that need to be addressed and solved before unmanned aircraft fly in unison with manned aircraft. Regardless, the information provided in this section is, to some degree, complete and comprehensive allowing for the reader to understand airspace regulation, certification procedures, airworthiness, UAV safety, and reliability.

87

Aviation Regulation Konstantinos Dalamagkidis

Contents 87.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2118 87.2 Important Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2121 87.3 Airworthiness Certification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2122 87.3.1 Type Certificate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2123 87.3.2 Standard Certificates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2123 87.3.3 Special Certificates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2124 87.4 Special Aircraft Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2125 87.4.1 Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2126 87.4.2 R/C Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2126 87.5 Pilot Certification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2126 87.6 Operation Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2127 87.6.1 Flight Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2128 87.6.2 Emergency Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2129 87.6.3 Maintenance Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2130 87.7 Airspace Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2130 87.8 Regulation Development Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2132 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2133

Abstract

This chapter presents a brief introduction to aviation regulation. Key terms are defined and an overview of current manned aviation regulations is provided using the U.S. Federal Aviation Regulation as an example. This includes airworthiness certification, operation rules, and airspace classes.

K. Dalamagkidis Institut f¨ur Informatik I6, Technische Universit¨at M¨unchen, Garching bei M¨unchen, Germany e-mail: [email protected] K.P. Valavanis, G.J. Vachtsevanos (eds.), Handbook of Unmanned Aerial Vehicles, DOI 10.1007/978-90-481-9707-1 109, © Springer Science+Business Media Dordrecht 2015

2117

2118

87.1

K. Dalamagkidis

Introduction

The need to regulate civil aviation ensuring safety and healthy competition dates back to the 1920s, with several relevant conventions addressing such issues and concerns. The most significant such convention took place in Chicago in 1944, right after the end of the Second World War with more than 50 States attending. The accomplishments of that conference set the groundwork for aviation safety and international cooperation on regulations, standards, and procedures development, all relevant even to this day. Attending States also founded the International Civil Aviation Organization (ICAO) as a means to secure progress accomplished during the conference, as well as to facilitate future cooperation. Although UAV operations were very limited before the 1944 Chicago Convention, Article 8 refers specifically to pilotless aircraft, and provisions within still apply to current systems. Some of those provisions are that a UAV cannot fly over another State without special authorization by that State (Article 8), UAVs are required to bear registration marks (Article 20), and they must have a certificate of airworthiness (Article 31). It should be noted that the Chicago Convention applies to civil aircraft, and as a result, aircraft used in military or law enforcement services may have different restrictions. In the United States, aviation regulations are collected and codified in the Code of Federal Regulations (CFR), Title 14, Chap. I, also known as Federal Aviation Regulation (FAR). Similarly, in Europe the Joint Aviation Authorities (JAA) has issued the Joint Aviation Requirements (JAR), while other countries/regions may have other similar regulatory documents. Due to an ongoing effort for harmonization between the aviation regulations, part and section numbers between the JAR and the Title 14 CFR are largely the same. This chapter presents an overview of key parts of current manned aviation regulations as defined in the Title 14 CFR, with the understanding that the provisions of other aviation regulations of other countries will be similar, if not the same. It should be noted that these regulations are currently applicable to manned aircraft but are more than likely to influence UAV regulations. Where appropriate, the relative section in the Title 14 CFR will be given. The aim of this chapter is to familiarize the reader with the general regulatory framework and with terms and concepts that are used throughout the handbook. For the latest and most accurate information, one is advised to consult with the civil aviation authority of his/her country and the current version of the respective regulations. For example, the Title 14 CFR is publicly available both online and in print from the Government Printing Office. The main focus of aviation regulation is to ensure that the aviation industry operates in a safe manner. For example, the U.S. Federal Law gives the Secretary of Transportation and the Administrator of the Federal Aviation Administration (FAA) the authority to conduct investigations, prescribe regulations, standards, and procedures, and issue orders [49 USC 40113(a)]. The paragraph on safety considerations in public interest [49 USC 40101(d)] reads:

87 Aviation Regulation

2119

. . . the Administrator shall consider the following matters, among others, as being in the public interest: 1. assigning, maintaining, and enhancing safety and security as the highest priorities in air commerce. 2. regulating air commerce in a way that best promotes safety and fulfills national defense requirements. 3. encouraging and developing civil aeronautics, including new aviation technology. 4. controlling the use of the navigable airspace and regulating civil and military operations in that airspace in the interest of the safety and efficiency of both of those operations. 5. consolidating research and development for air navigation facilities and the installation and operation of those facilities. 6. developing and operating a common system of air traffic control and navigation for military and civil aircraft. 7. providing assistance to law enforcement agencies in the enforcement of laws related to regulation of controlled substances, to the extent consistent with aviation safety.

The statutory mandate of the FAA also includes regarding safety: . . . before authorizing new air transportation services, evaluating the safety implications of those services; and preventing deterioration in established safety procedures, recognizing the clear intent, encouragement, and dedication of Congress to further the highest degree of safety in air transportation and air commerce, and to maintain the safety vigilance that has evolved in air transportation and air commerce and has come to be expected by the traveling and shipping public. [49 USC 40101(a)]

The Title 14 CFR comprises of several subchapters as follows: A Definitions (Parts 1–3) B Procedural rules (Parts 11–17) C Aircraft (Parts 21–50). Includes airworthiness certification (21–39), maintenance (43), as well as aircraft registration and marking (45–49) D Airmen (Parts 60–67) E Airspace (Parts 71–77) F Air Traffic and General Operating Rules (Parts 91–106) which includes operating rules (91–99) and special classes of vehicles (101–105) G Air Carriers and Operators for Compensation or Hire: Certification and Operations (Parts 110–139) H Schools and Other Certified Agencies (Parts 140–147) I Airports (Parts 150–169) J Navigational Facilities (Parts 170–171) K Administrative Regulations (Parts 183–193) N War Risk Insurance (Parts 198–199) The provisions of the Title 14 CFR notwithstanding, the FAA issues supplementary nonregulatory material like handbooks, orders, Advisory Circulars (AC), and Technical Standard Orders (TSO) that clearly define appropriate procedures, standards, and practices that may be followed to ensure compliance with current regulations. This material helps ensure that aircraft manufacturers and operators are able to establish the minimum level of safety and reliability required for civil operations ASTM International (2004). Nevertheless, the following sections will draw material only from the CFR since the focus of this chapter is on regulatory requirements.

2120

K. Dalamagkidis

With respect to the ACs, of particular interest are AC 00-2 and AC 00-44. The first provides a checklist of all ACs, as well as the status of other FAA publications, while the second contains a list of any changes made by the FAA to the CFR. The FAA also publishes the Aeronautical Information Manual (AIM), a regularly updated document, mainly aimed at pilots. The AIM contains the basic flight information and procedures when flying in the U.S. National Airspace System (NAS). It consists of the following chapters: 1. Air Navigation 2. Aeronautical Lighting and Other Airport Visual Aids 3. Airspace 4. Air Traffic Control 5. Air Traffic Procedures 6. Emergency Procedures 7. Safety of Flight 8. Medical Facts for Pilots 9. Aeronautical Charts and Related Publications 10. Helicopter Operations The AIM is supplemented by the International Flight Information Manual (IFIM) which provides information for planning international flights. Other handbooks published by the FAA that might be of interest include the “Airplane Flying Handbook,” the “Helicopter Flying Handbook,” and the “Pilot’s Handbook of Aeronautical Knowledge,” all aimed at pilots, as well as the “Aviation Maintenance Technician Handbook” and the “Aviation Instructor’s Handbook” for technicians and instructors, respectively. These and a number of other documents are available for download from the FAA website free of charge. Several of the aforementioned documents adopt established standards prepared by government agencies like the U.S. Department of Defense, standards development organizations, as well as other organizations, national or international. A non-exhaustive list of organizations that have been involved with the development of aerospace-related standards is provided below: • Aeronautical Radio, Incorporated (ARINC) • American Institute of Aeronautics and Astronautics (AIAA) • American National Standards Institute (ANSI) • American Society of Testing & Materials (ASTM) • American Welding Society (AWS) • Electronic Industries Alliance (EIA) • Electrostatic Discharge Association (ESDA) • European Organization for Civil Aviation Equipment (EUROCAE) • Institute of Electrical and Electronics Engineers (IEEE) • Institute of Environmental Sciences and Technology (IEST) • International Civil Aviation Organization (ICAO) • International Electrotechnical Commission (IEC) • International Organization for Standardization (ISO) • National Aeronautics and Space Administration (NASA) • National Institute of Standards and Technology (NIST)

87 Aviation Regulation

• • • •

2121

The North Atlantic Treaty Organization Standards Agency (NSA) Occupational Safety and Health Administration (OSHA) Radio Technical Commission for Aeronautics (RTCA) Society of Automotive Engineers (SAE)

87.2

Important Definitions

In this section, some definitions will be provided that are important from a regulatory viewpoint. The first concerns the definition of the term aircraft, which according to the ICAO is “. . . any machine that can derive support in the atmosphere from the reactions of the air other than the reactions of the air against the earth’s surface” [ICAO Annex 1, Annex 6 Part I]. The FAA defines an aircraft as “. . . a device that is used or intended to be used for flight in the air” [Title 14 CFR 1.1]. As such, among many other things, UAVs also fall within that definition. An aircraft can be classified as public or civil. A public aircraft is defined by the FAA as either “. . . an aircraft used only for the United States Government, . . . [or] owned and operated by the government of a State, the District of Columbia, or a territory or possession of the United States or a political subdivision of one of these governments” or “. . . an aircraft owned or operated by the armed forces or chartered to provide transportation to the armed forces. . . ,” provided some conditions are met. Civil aircraft are all aircraft that are not public [Title 14 CFR 1.1]. To operate an aircraft means “use, cause to use or authorize to use aircraft, for the purpose of air navigation including the piloting of aircraft, with or without the right of legal control (as owner, lessee, or otherwise)” [Title 14 CFR 1.1]. Aircraft often operate within controlled airspace, that is, “an airspace of defined dimensions within which air traffic control service is provided to IFR flights and to VFR flights in accordance with the airspace classification” [Title 14 CFR 1.1]. The air traffic control service (ATS) is a “a service operated by appropriate authority to promote the safe, orderly, and expeditious flow of air traffic” [Title 14 CFR 1.1]. Another relevant term is air traffic management (ATM) which refers to “the dynamic, integrated management of air traffic and airspace – safely, economically and efficiently – through the provision of facilities and seamless services in collaboration with all involved parties” (Air Transport Association (ATA) 2008). The terms accident, incident, damage, and hazard are of particular importance when specifying safety requirements. One can find varying definitions for the term accident. One of the most succinct used by the FAA is “An unplanned event or series of events that results in damages.” (Federal Aviation Administration 2000). The ICAO defines the same term as “An occurrence associated with the operation of an aircraft which takes place between the time any person boards the aircraft with the intention of flight until such time as all such persons have disembarked, in which (a) a person is fatally or seriously injured as a result of: being in the aircraft; or direct contact with any part of the aircraft, including parts which have become detached from the aircraft; or direct exposure to jet blast (except when the injuries are from natural causes, self inflicted or inflicted by other persons, or when

2122

K. Dalamagkidis

the injuries are to stowaways hiding outside the areas normally available to the passengers or crew); or (b) the aircraft sustains damage or structural failure which: adversely affects the structural strength, performance or flight characteristics of the aircraft and would normally require major repair or replacement of the affected component (except for engine failure or damage, when the damage is limited to the engine, its cowlings or accessories; or for damage limited to propellers, wing tips, antennas, tires, brakes, fairings, small dents or puncture holes in the aircraft skin); or (c) the aircraft is missing or is completely inaccessible” [ICAO Annex 13]. On the other hand, incident is defined as “An occurrence, other than an accident, associated with the operation of an aircraft which affects or could affect the safety of operation” [ICAO Annex 13]. The definition of accidents makes use of the term damage, which in turn is defined as “. . . the severity of injury, and/or the physical, and/or functional, and/or monetary loss that could result if hazard control is less than adequate” (Federal Aviation Administration 2000). The term hazard refers to conditions that may lead to accidents, regardless of their probability or the severity of the outcome. The FAA differentiates between primary hazards and contributory hazards: “A primary hazard is one that can directly and immediately results in: loss, consequence, adverse outcome, damage, fatality, system loss, degradation, loss of function, injury, etc. The primary hazard is also referred to as: catastrophe, catastrophic event, critical event, marginal event, and negligible event” (Federal Aviation Administration 2000). Initiating hazards include events and conditions that start an adverse chain of events that can lead to an accident. Primary hazards are events that directly and immediately cause an accident. Contributory hazards are any kind of conditions that can lead to an accident, including primary hazards.

87.3

Airworthiness Certification

In order for any aircraft to fly legally in the United States, it must carry an airworthiness certificate issued by the FAA [Title 14 CFR 91.203]. Airworthiness certification covers a wide spectrum of areas related to aspects of the aircraft design, construction, and operation. Presented below are some of these areas along with a selection of the various aspects investigated during certification: Flight: Performance, flight characteristics, controllability, maneuverability, and stability Structure: Loads, control surfaces, stabilizing and balancing surfaces, and fatigue evaluation Design and construction: Wings, control surfaces, control systems, landing gear, and pressurization Powerplant: Fuel system, oil system, cooling system, induction system, exhaust, and control Equipment: Instruments’ installation, electrical systems, lights, and safety equipment

87 Aviation Regulation

2123

In addition to aircraft, Airworthiness Directives (AD) exist for aircraft engines and propellers. According to FAA Order 8130.2F, there are two conditions that need be met in order for an aircraft to be considered airworthy; it must conform to its type certificate including any supplemental certificates, and it must be in a condition that ensures safe operation Federal Aviation Administration (2004). For aircraft that are not type certified, compliance with the second condition is adequate. Besides standard certification, Special Airworthiness Certificates (SAC) are also available, usually for experimental or special purpose aircraft. It should be noted that the Title 14 CFR allows the FAA administrator to prescribe additional requirements and special conditions for aircraft, aircraft engines, or propellers when due to a novel or unusual feature, current airworthiness regulations are inadequate or inappropriate [Title 14 CFR 21.16].

87.3.1 Type Certificate A type certificate is a collection of documents, drawings, specifications, datasheets and any related information needed to demonstrate compliance with the applicable paragraphs of the Title 14 CFR [Title 14 CFR 21.41]. These may also include inspection and preventive maintenance programs and instructions for continued airworthiness [Title 14 CFR 21.31]. During the application for type certificate, the FAA administrator may require an inspection and test of the aircraft [Title 14 CFR 21.33], which may also include flight tests [Title 14 CFR 21.35]. Once a type certificate has been issued, it is in effect until surrendered, suspended, or revoked [Title 14 CFR 21.51]. Nevertheless, after modifications to an aircraft, a new certificate may be required. When the extend of the changes is not significant, the type certificate can be amended [Title 14 CFR 21.91] or a supplemental certificate will be issued [Title 14 CFR 21.113].

87.3.2 Standard Certificates Standard airworthiness certificates are given to aircraft that are type certificated in any of the categories defined in [Title 14 CFR 21.175], including: • Normal, utility, acrobatic, and commuter aircraft (Title 14 CFR Part 23) • Transport aircraft (Title 14 CFR Part 25) • Normal rotorcraft (Title 14 CFR Part 27) • Transport rotorcraft (Title 14 CFR Part 29) • Manned free balloons (Title 14 CFR Part 31) In addition to the above categories, type certification is available for primary [Title 14 CFR 21.24], restricted [Title 14 CFR 21.25], U.S. Army surplus [Title 14 CFR 21.27] and imported [Title 14 CFR 21.29] aircraft, as well. An overview of the applicability requirements for each of the aforementioned categories is given in Table 87.1.

2124

K. Dalamagkidis

Table 87.1 Aircraft types with standard airworthiness certificates along with occupancy, weight, and other restrictions (compiled from information in the Title 14 CFR) Category Max. seats MTOW (kg) Notes Normal 9a 5; 670 Non-acrobatic operations Utility 9a 5; 670 Limited acrobatic operations Acrobatic 9a 5; 670 No restrictions Commuter 19a 8; 600 Non-acrobatic operations Transport N/A N/A Primary 4b 1; 225 Limited power/unpressurized cabin Restricted N/A N/A Special purpose operationsc Normal rotorcraft 9a 3; 175 d Transport rotorcraft 9a 9; 070 Manned free balloons N/A N/A a Excluding pilot seats b Includes the pilot c Includes agricultural, forest and wildlife conservation, aerial surveying, patrolling, weather control, and aerial advertising operations d Transport rotorcraft are type certificated in two categories (A and B). Rotorcraft that meet the above restriction may be certificated in the B category, while those with higher seating capacity must be certificated in the A category

87.3.3 Special Certificates In FAA Order 8130.2F and for aircraft that do not meet requirements for a standard certificate but are still capable of safe flight, SAC are available. More specifically special certificates can be given in the primary [Title 14 CFR 21.184], restricted [Title 14 CFR 21.185] and limited [Title 14 CFR 21.189] categories, for aircraft type certificated under these categories. In addition to that, SACs are available for aircraft belonging to the light-sport category and for experimental aircraft. Finally special flight permits are also available.

87.3.3.1 Light-Sport FAA Order 8130.2F also provides rules for certification of light-sport Aircraft (LSA). This category is for aircraft other than helicopters that do not exceed 600– 650 kg, having a maximum speed of not more than 120 knots and a capacity of not more than two persons. Additional requirements are made based on the presence of certain equipment. A SAC is issued for aircraft of this type after successful inspection of the aircraft and its documentation. The latter includes operating instructions and maintenance procedures and a statement from the manufacturer that the aircraft complies with the provisions of the appropriate consensus standards [Title 14 CFR 21.190]. Upon successful completion of the inspection, the FAA may amend the certificate with operational restrictions, if deemed necessary.

87 Aviation Regulation

2125

87.3.3.2 Experimental Experimental certificates are given for a variety of purposes [Title 14 CFR 21.191]: • Research and development of equipment, operating techniques, or new aircraft designs. • Showing aircraft compliance with a type certificate or a supplemental certificate after major changes. • Crew training. • Exhibitions at air shows or movies. This includes required pilot training and flight from and to the exhibition area. • Air racing, including practicing and flight from/to the area. • Market surveys, sales training and customer flight crew training. • Operating of amateur-built aircraft. • Operating of primary kit-built aircraft that have not been assembled under the supervision and control of a production certificated entity. • Operation of certain types of light-sport aircraft. Before a special certificate in this category is issued, the applicant must submit appropriate documentation. In the case of aircraft used for research and development purposes, this documentation includes the purpose of the experiment along with the number of flights, the location, and drawings/photographs of the aircraft [Title 14 CFR 21.193]. FAA Order 8130.2F makes provisions for several operational restrictions for experimental aircraft depending on their characteristics. The duration of experimental certificates is 1 year or less except for kit-built aircraft that typically do not expire [Title 14 CFR 21.181]. 87.3.3.3 Special Flight Permits These permits are given to aircraft that would not qualify for other airworthiness certificates but are capable of safe flight [Title 14 CFR 21.197]. The purpose of these permits is to allow the aircraft to fly to a different location for storage, repairs, and maintenance or to avoid areas of impending danger. The permit is issued after an application where the purpose and characteristics of the flight are detailed, and it may include limitations or special instructions from the FAA [Title 14 CFR 21.199]. Special flight permits may also be given for airworthy aircraft, to allow them to fly with excess fuel weight, beyond their certificated capacity or when flying over areas where refueling is not possible [Title 14 CFR 21.197].

87.4

Special Aircraft Categories

Although normally all aircraft need either a standard or a SAC to fly, there is a category of aircraft (classified as vehicles in the Title 14 CFR) for which this requirement is waived. The other special category concerns remote-controlled (R/C) model aircraft that also operates under few restrictions. Although not mentioned in the

2126

K. Dalamagkidis

Title 14 CFR, R/C aircraft are of interest since they present the basis of many UAS designs. It should be stressed however that R/C models are allowed to operate only for recreational purposes and that in the Federal Aviation Administration (2007) it was made clear that UAS operations cannot be based on R/C model procedures.

87.4.1 Vehicles This category of aircraft includes moored balloons, unmanned balloons, unmanned rockets defined in Title 14 CFR Part 101, and ultralights defined in Title 14 CFR Part 103. Ultralights are single-occupant, manned aircraft used for recreation or sport purposes only, with a maximum empty weight of 70 kg for unpowered and 115 kg for powered vehicles [Title 14 CFR 103.1]. Many of the requirements regarding pilot certification, operating and flight rules, vehicle registration and marking, and maintenance certification, including the requirement to carry an airworthiness certificate that are normally applicable to aircraft, are waived for this category. Nevertheless, operational restrictions may be in place. For example, the following pertain to the operation of ultralight vehicles: • Daylight operations only [Title 14 CFR 103.11]. • Yield the right-of-way to all aircraft [Title 14 CFR 103.13]. • No operations allowed over congested areas [Title 14 CFR 103.15]. • No operations allowed in Class A, B, C, and D airspace. For operations in Class E near airports, ATC authorization is required first [Title 14 CFR 103.17]. • Pilot must operate by visual reference with the surface [Title 14 CFR 103.21].

87.4.2 R/C Models Model airplanes are regulated on a voluntary basis, based on AC91-57 with few operational restrictions. In addition to that an independent organization, the Academy of Model Aeronautics (AMA) issues normal or restricted flight permits after inspection of the model, provides insurance for its members, and organizes areas to safely practice aeromodeling. It is noteworthy that in Academy of Model Aeronautics (2007), restrictions additional to the ones in FAA AC91-57 are imposed, both in design (e.g., the weight of the models and their propulsion methods) as well as in operation.

87.5

Pilot Certification

Title 14 CFR Part 61 is involved with the requirements for issuing pilot, flight instructor, and ground instructor certificates, ratings, and authorizations [Title 14 CFR 61.1]. An appropriate pilot certificate is required for a person to assume the role of pilot in command or of required crew member [Title 14 CFR 61.3]. Some

87 Aviation Regulation Table 87.2 Pilot certificates summarized from [Title 14 CFR 61.5]

2127

Category Airplane

Rotorcraft Lighter-than-air Weight-shift-control aircraft Powered lift Powered parachute

Class Single-engine land Multiengine land Single-engine sea Multiengine sea Helicopter Gyroplane Airship Balloon Weight-shift-control aircraft land Weight-shift-control aircraft sea N/A Powered parachute land Powered parachute sea

operators are also required to possess a current medical certificate issued based on procedures described in Title 14 CFR Part 67. There are several types of pilot certificates with different training and certification requirements and with different privileges for their holders [Title 14 CFR 61.5]: 1. Student pilot 2. Sport pilot 3. Recreational pilot 4. Private pilot 5. Commercial pilot 6. Airline transport pilot certificate Each pilot certificate (with the exception of a student certificate) comes with ratings for aircraft categories, classes, and types the holder may operate as well as the instrument rating for private and commercial pilots. Table 87.2 summarizes the aircraft category and class ratings. There are also instrument ratings for airplanes, helicopters, and powered lifts [Title 14 CFR 61.5]. Similar ratings are placed on flight instructor and ground instructor certificates when all the training and certification requirements are met. Title 14 CFR Part 61 also includes the level of knowledge, training, operations proficiency, and experience a pilot must possess before being issued a certificate. This includes training and testing procedures. Title 14 CFR Part 63 is involved with certification of crew members other than pilots and Title 14 CFR Part 65 with airmen certification.

87.6

Operation Rules

Operational rules for manned aircraft operating in the U.S. NAS are prescribed in Title 14 CFR Part 91, which applies to all aircraft with the exception of moored balloons, kites, unmanned rockets, unmanned free balloons, and ultralights

2128

K. Dalamagkidis

[Title 14 CFR 91.1]. Part 91 also establishes the responsibility for aircraft operators to support the continued airworthiness of each airplane [Title 14 CFR 91.1]. The person ultimately responsible for the operation of the aircraft is the pilot in command [Title 14 CFR 91.3]. The pilot is also responsible for evaluating the airworthiness of the aircraft and determining if it is in a condition safe to fly [Title 14 CFR 91.7]. After the aircraft has been deemed safe to fly and before takeoff, the pilot needs to be familiar with any information concerning the flight, such as weather reports, fuel requirements, airport characteristics, and aircraft performance characteristics [Title 14 CFR 91.103]. To minimize the risk of collisions, no person is allowed to operate an aircraft in close proximity to another [Title 14 CFR 91.111], and when the weather conditions permit, the pilot should be alert in order to see and avoid other aircraft [Title 14 CFR 91.113]. Additionally right-of-way rules are established [Title 14 CFR 91.113]. With the exception of water operations, typically the aircraft with less maneuverability has the right-of-way. This rule is superseded when an aircraft is in distress, at which time it has the right-of-way with respect to all other air traffic. In general during emergencies pilots are allowed to deviate from the requirements of Part 91, even contrary to ATC instruction, provided that Air traffic control (ATC) is notified of this deviation as soon as possible [Title 14 CFR 91.3,91.123]. In any other situation, no one is allowed to deviate from ATC clearance and instructions [Title 14 CFR 91.123]. Additional safety regulations do not permit pilots to fly below 10;000 ft or in proximity of Class B, C, and D airspace at speeds exceeding 250 and 200 knots, respectively [Title 14 CFR 91.117]. Similarly, minimum safe altitudes are established so that upon catastrophic failures, an emergency landing can take place without undue risk to people or property [Title 14 CFR 91.119].

87.6.1 Flight Rules Title 14 CFR Part 91 defines two types of flight rules: visual flight rules (VFR) and instrument flight rules (IFR). In addition to the normal operations, Title 14 CFR Part 91 includes guidelines for emergencies as well as special operations like aerobatics, towing, and parachuting.

87.6.1.1 Visual Flight Rules Under VFR the pilot is expected to control the aircraft’s trajectory and avoid other aircraft based on visual cues, although separation instruction may be provided by ATC when flying in certain classes of controlled airspace. A prerequisite to flying under VFR is the presence of enough fuel onboard, so that the aircraft can reach its first landing destination and fly for 30 min or 45 min after that during the day or night, respectively [Title 14 CFR 91.151]. Similar requirements exist on the flight altitude and weather conditions [Title 14 CFR 91.155]. The minimum weather conditions for VFR operations are summarized in Table 87.3.

87 Aviation Regulation

2129

Table 87.3 Weather minimums for VFR operations (Source: [Title 14 CFR 91.155]) Visibility Distance from clouds Airspace (statute miles) Above Below Horizontal Class A Class B Class C Class D Class E (1,200 ft and