Wireless Sensor Networks: Energy Harvesting and Management for Research and Industry (Signals and Communication Technology) 3030296989, 9783030296988

Citation preview

Signals and Communication Technology

Hossam Mahmoud Ahmad Fahmy

Wireless Sensor Networks Energy Harvesting and Management for Research and Industry

Signals and Communication Technology Series Editors Emre Celebi, Department of Computer Science, University of Central Arkansas, Conway, AR, USA Jingdong Chen, Northwestern Polytechnical University, Xi’an, China E. S. Gopi, Department of Electronics and Communication Engineering, National Institute of Technology, Tiruchirappalli, Tamil Nadu, India Amy Neustein, Linguistic Technology Systems, Fort Lee, NJ, USA H. Vincent Poor, Department of Electrical Engineering, Princeton University, Princeton, NJ, USA

This series is devoted to fundamentals and applications of modern methods of signal processing and cutting-edge communication technologies. The main topics are information and signal theory, acoustical signal processing, image processing and multimedia systems, mobile and wireless communications, and computer and communication networks. Volumes in the series address researchers in academia and industrial R&D departments. The series is application-oriented. The level of presentation of each individual volume, however, depends on the subject and can range from practical to scientific. “Signals and Communication Technology” is indexed by Scopus.

More information about this series at http://www.springer.com/series/4748

Hossam Mahmoud Ahmad Fahmy

Wireless Sensor Networks Energy Harvesting and Management for Research and Industry

123

Hossam Mahmoud Ahmad Fahmy Faculty of Engineering Department of Computer Engineering and Systems Ain Shams University Cairo, Egypt

ISSN 1860-4862 ISSN 1860-4870 (electronic) Signals and Communication Technology ISBN 978-3-030-29698-8 ISBN 978-3-030-29700-8 (eBook) https://doi.org/10.1007/978-3-030-29700-8 © Springer Nature Switzerland AG 2020 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. 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. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Dedicated to my family; parents, brothers and sister with whom I grew up warmly… wife and daughters who gave my life a caring taste… Dedication is not only for who are in our world…

Preface

Writing a book is tempting, many ideas and topics, idea after idea, and topic upon topic, what to elaborate, which to mention, the reader must find a satisfying answer, enough knowledge; overlooking and going-by are painful choices for the author, space is limited, and a tough decision is to be made, without compromising what should be transferred to the audience. Writing a scientific book is navigating, across the Nile, the Mediterranean, the Atlantic and the Indian oceans, in boat and in glass submarine, looking and searching for known and unknown species, appreciating diversified colors and variety of sizes, collecting for a near benefit and for the future. I navigated for the second time, explored, day and night, when cold and hot, whether windy or breezing, without tolerating a least chance to know and learn. Networking is a field of integration, hardware and software, protocols and standards, simulation and testbeds, wired and wireless, VLSI and communication, energy harvesting and management, an orchestrated harmony that collaborates dependably, all for the good of a connected well-performing network. That is the charm of networking, of life in a civilization that recognizes differences and goes on. This book focuses on the concepts of energy, and energy harvesting and management techniques for WSNs; a meticulous care has been accorded to the definitions, terminologies, and protocols. Definitions and terminologies are made clear without leaning on the relaxing assumption that they are already known or easily reachable, and the reader is not to be diverted from the main course. Neatly drawn figures assist in viewing and imagining the offered topics. To make energy-related topics felt and seen, the adopted technologies as well as their manufacturers are presented in detail. With such a depth, this book is intended for a wide audience, and it is meant to be helper and motivator, for the senior undergraduates, postgraduates, researchers, and practitioners; concepts and energy-related applications are laid out, research and practical issues are backed by the appropriate literature, and new trends are put under focus. For senior undergraduate students, it familiarizes with conceptual foundations and practical project implementations. Also, it is intended for graduate students making a thesis and in need of specific knowledge

vii

viii

Preface

on WSNs and the related energy harvesting and management techniques. Moreover, it is targeting researchers and practitioners interested in features and applications of WSNs, and on the available energy harvesting and management projects and testbeds. Three parts form the backbone of this book. Part I (Concepts and Energy Harvesting) includes Chap. 1 and 2, for a review of WSN concepts and in-depth presentation of the energy harvesting techniques. Part II (Energy Management Perspectives) embodies Chap. 3 and Chaps. 4–6 for a thorough analysis of the three perspectives on energy management: specifically, duty-cycling, data-driven, and mobility-based approaches. Part III (Harvesting and Management Projects and Testbeds) containing Chaps. 7–9 brings practice to theory through energy harvesting and management projects and testbeds. Part IV is a single concluding chapter. Chapter 10 ignites the launch into the wide realm of WSNs, research, and implementation of energy-focused protocols and techniques for energy harvesting and management. A longer WSN lifetime is the prime target. Exercises at the end of each chapter are not just questions and answers; they are not limited to recapitulate ideas. Their design objective is not bound to be a methodical review of the provided concepts, but rather as a motivator for lot more of searching, finding, and comparing beyond what has been presented in this book. Talking numbers, this book extends over ten chapters and embodies 188 acronyms, 238 colored figures, 41 tables, and above 650 references. With the advance of technology, writing a book is becoming easier, and information is attainable; but it is certainly tedious, and details and depth are not to be missed within a comforting accuracy. Reader trust cannot be waived. Every bit of knowledge included in this book is checked and rechecked multiple times, no accidental slips. A book, any book, is a step in a long path sought to be correct, precise as possible, nonetheless errors are non-escapable, and they are avoided iteratively, with follow-up and care. The preface is the first get-together between the author and the audience, it is the last written words, and it is lying in the ground after the end line, to restore taken breath, to enjoy relaxing after long painful efforts, mentally and physically, to relax in preparation for a new game. Bringing a book to life consumes months and months, days and nights, events after events, familial, social, and at the wide world of technology, sports, and politics. This book has seen much and recorded some. An author has his ups and downs, as everybody, but he is visible like nobody. Could he manage to hide some of his letdowns? Yes he has to, unlike anybody, for the sake of his book, his readership. A book is a whole life, maybe in the current, in the past, or in future. The author has many dreams, completing the current chapter, reaching the last chapter, agreeing on the book cover, handing the book to the publisher, receiving the manuscript for revision, talking royalties, …

Preface

ix

Writing with care and feelings can be a title for my books. Authoring is a duty, a passion, an exhausting ordeal with mostly a moral reward… If you find somebody talking to himself, tumbling, wearing a differently colored pair of shoes, don’t laugh at him, he is probably writing a book… Cairo, Egypt

Hossam Mahmoud Ahmad Fahmy email: [email protected]

Contents

Part I 1

Concepts and Energy Harvesting

Wireless Sensor Networks Essentials . . . . . . . . . . . . . . . . . . . 1.1 Sensing, Senses, Sensors . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Toward Wireless Sensor Networks . . . . . . . . . . . . . . . . . 1.3 Mobile Ad Hoc Networks (MANETs) . . . . . . . . . . . . . . 1.4 Wireless Mesh Networks (WMNs) . . . . . . . . . . . . . . . . . 1.5 Closer Perspective to WSNs . . . . . . . . . . . . . . . . . . . . . 1.5.1 Wireless Sensor Nodes . . . . . . . . . . . . . . . . . . . . 1.5.2 Architecture of WSNs . . . . . . . . . . . . . . . . . . . . 1.6 Types of WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Terrestrial WSNs . . . . . . . . . . . . . . . . . . . . . . . . 1.6.2 Underground WSNs . . . . . . . . . . . . . . . . . . . . . . 1.6.3 Underwater Acoustic Sensor Networks (UASNs) . 1.6.4 Multimedia WSNs . . . . . . . . . . . . . . . . . . . . . . . 1.6.5 Mobile WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . 1.7 Performance Metrics of WSNs . . . . . . . . . . . . . . . . . . . . 1.8 WSNs Standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9 Protocol Stack of WSNs . . . . . . . . . . . . . . . . . . . . . . . . 1.9.1 Physical Layer . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9.2 Data Link Layer . . . . . . . . . . . . . . . . . . . . . . . . 1.9.3 Network Layer . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9.4 Transport Layer . . . . . . . . . . . . . . . . . . . . . . . . . 1.9.5 Application Layer . . . . . . . . . . . . . . . . . . . . . . . 1.9.6 Cross-Layer Protocols for WSNs . . . . . . . . . . . . 1.10 Conclusion for Energetic Trip . . . . . . . . . . . . . . . . . . . . 1.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . .

3 4 5 7 8 11 11 12 13 13 14 15 16 17 19 21 23 26 27 28 29 30 32 33 35 36

xi

xii

2

Contents

Energy Harvesting in WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Energy Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Energy Harvesting Concepts and Components . . . . . . . . . . . 2.2.1 Energy Harvesting Architectures . . . . . . . . . . . . . . . . 2.2.2 Power and Energy Differentiated . . . . . . . . . . . . . . . . 2.2.3 Energy Harvesting Versus Battery-Operated Systems . 2.2.4 Storage Technologies . . . . . . . . . . . . . . . . . . . . . . . . 2.2.5 Harvesting Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.6 Conditions for Energy-Neutral Operation . . . . . . . . . . 2.2.7 Characteristics and Classifications of the Harvestable Energy Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.8 Multisupply and Autonomous Energy Harvesting . . . . 2.3 Energy Harvesting Mechanisms . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Photovoltaic Energy Harvesting . . . . . . . . . . . . . . . . 2.3.2 Energy Harvesting from Motion and Vibration . . . . . 2.3.3 Energy Harvesting from Temperature Differences . . . 2.3.4 Wind Energy Harvesting . . . . . . . . . . . . . . . . . . . . . 2.3.5 Wireless Energy Harvesting . . . . . . . . . . . . . . . . . . . 2.3.6 Biochemical Energy Harvesting . . . . . . . . . . . . . . . . 2.3.7 Acoustic Energy Harvesting . . . . . . . . . . . . . . . . . . . 2.3.8 Hybrid Energy Harvesting . . . . . . . . . . . . . . . . . . . . 2.4 MEMS for Energy Harvesters Fabrication . . . . . . . . . . . . . . 2.5 Conclusion for Enlightenment . . . . . . . . . . . . . . . . . . . . . . . 2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Part II

. . . . . . . . .

. . . . . . . . .

41 41 43 43 44 48 50 55 56

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

58 60 62 64 64 69 71 71 73 80 82 85 90 92 92

Energy Management Perspectives

3

Energy Management Techniques for WSNs . . . . . 3.1 Energy Conservation Approaches . . . . . . . . . . 3.1.1 Duty-Cycling Techniques . . . . . . . . . . 3.1.2 Data-Driven Techniques . . . . . . . . . . . 3.1.3 Mobility-Based Techniques . . . . . . . . 3.2 Conclusion for More on Energy Management . 3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . .

4

Energy Management Techniques for WSNs (1): Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Duty-Cycling Approach Taxonomy . . . . . . 4.1.1 Topology Control Protocols . . . . . . 4.1.2 Power Management Protocols . . . . .

. . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

103 103 105 106 106 107 107 107

. . . .

. . . .

. . . .

. . . .

109 109 111 133

Duty-Cycling . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Contents

xiii

4.2 Conclusion for Longer Duty-Cycling . . . . . . . . . . . . . . . . . . . . 246 4.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 5

6

Energy Management Techniques for WSNs (2): Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Data-Driven Approach Taxonomy . . . . . . . 5.1.1 Data Reduction Protocols . . . . . . . . 5.1.2 Energy-Efficient Data Acquisition . . 5.2 Conclusion for Well-Managed Lifestyle . . . 5.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

259 259 260 340 383 388 389

Energy Management Techniques for WSNs (3): Mobility-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Mobility in WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 Architecture of WSNs with Mobile Elements . . . . . . . 6.1.2 Role of Mobile Elements in WSNs . . . . . . . . . . . . . . 6.2 Mobility-Based Approach Taxonomy . . . . . . . . . . . . . . . . . . 6.2.1 Mobile Sink Protocols . . . . . . . . . . . . . . . . . . . . . . . 6.2.2 Mobile Relay Protocols . . . . . . . . . . . . . . . . . . . . . . 6.3 Conclusion for Controlled Mobility . . . . . . . . . . . . . . . . . . . 6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

399 399 400 401 405 406 448 476 483 484

Part III 7

Data-Driven . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

Harvesting and Management Projects and Testbeds

Energy Harvesting Projects for WSNs . . . . . . . . . . . . . . . . . . . . 7.1 Necessities-Driven Projects . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Energy Harvesting Projects . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 ZebraNet: Energy-Efficient Computing for Wildlife Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.2 Prometheus for Perpetual Environmentally Powered Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.3 Solar Biscuit: A Batteryless Wireless Sensor Network System for Environmental Monitoring Applications . . 7.2.4 Heliomote for Solar Energy Harvesting in Wireless Embedded Systems . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.5 Everlast: Long-Life, Super-Capacitor-Operated Wireless Sensor Node . . . . . . . . . . . . . . . . . . . . . . . 7.2.6 AmbiMax: Autonomous Energy Harvesting Platform for Multisupply Wireless Sensor Nodes . . . . . . . . . . . 7.2.7 Sunflower: Low-Power, Energy Harvesting System with Custom Multichannel Communication Interface .

. . 489 . . 489 . . 490 . . 490 . . 507 . . 514 . . 526 . . 536 . . 550 . . 558

xiv

Contents

7.2.8 Micro-Solar Power Sensor Networks for Forest Watersheds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.9 Energy Harvesting from Hybrid Indoor Ambient Light and Thermal Energy Sources . . . . . . . . . . . . . . . . . . 7.3 Conclusion for Radiance . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

9

Energy Management Projects for WSNs . . . . . . . . . . . . . . . . . . . 8.1 Energy Management Projects . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Evolution and Sustainability of a Wildlife Monitoring Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Initial System Design . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Evolution Stage 1: Improving Sensing and Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.3 Evolution Stage 2: Hardware Improvements . . . . . . . 8.2.4 Network Maintenance Costs . . . . . . . . . . . . . . . . . . . 8.2.5 Gained Experience . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Conclusion for Brightness . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . WSNs Energy Testbeds . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1 Functionalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Typical WSNs Energy Testbed . . . . . . . . . . . . . . . . . 9.2.1 PowerBench: A Scalable Testbed Infrastructure for Benchmarking Power Consumption . . . . . . 9.3 Conclusion for Brilliance . . . . . . . . . . . . . . . . . . . . . . 9.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Part IV

. . 569 . . . .

. . . .

580 599 601 602

. . 611 . . 611 . . 612 . . 614 . . . . . . .

. . . . . . .

619 626 630 633 634 635 636

. . . . . . . 639 . . . . . . . 639 . . . . . . . 641 . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

641 644 645 645

Ignition

10 Last Flare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651 Index of Abbreviations and Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . 659

About the Author

Prof. Hossam M. A. Fahmy Professor of Computer Engineering served as Chair of the Computer Engineering and Systems Department, Faculty of Engineering, Ain Shams University, Cairo, Egypt, from 2006 to 2008 and from 2010 to 2012. He participates in many academic activities in Egypt and abroad. Prof. Fahmy has published and refereed extensively in Springer, Elsevier and IEEE journals and in several refereed international conferences. His teaching and research areas are focused on computer networks, MANETs, WSNs, VANETs, fault tolerance, software, and Web engineering. He authored Wireless Sensor Networks: Concepts, Applications, Experimentation and Analysis, book published by Springer, 2016. He founded and chaired the IEEE International Conference on Computer Engineering and Systems (ICCES) from 2006 to 2008 and from 2010 to 2013. He is Senior IEEE Member, IEEE Region 8 Distinguished Visitor (2013–2015) and (2015–2018), Member of the Distinguished Visitor Committee of the IEEE Computer Society, and Member in the Cloud Computing Special Technical Community of the IEEE Computer Society. He speaks Arabic, French, and English.

xv

List of Acronyms

2-DOF ACC ADC AEA AFECA AINS AMRP AODV ARMA ARQ ASCENT ASK AWP B-MAC BER BTU CC CCA CDS CEC CLUDDA CNES CNS COP CPLD CSMA/CA CTS CUSUM DAC DBMAC

2-degree of freedom Active congestion control Analog-to-digital converter Adaptive election algorithm Adaptive fidelity energy-conserving algorithm Autonomous Intelligent Networks and Systems Average minimum reachability power Ad hoc on-demand distance vector Autoregressive–moving-average Automatic repeat request Adaptive self-configuring sensor networks topologies Amplitude-shift keying Asynchronous wakeup protocol Berkeley-MAC Bit error rate British thermal unit Control/charger Clear channel assessment Connected dominating set Cluster-based energy conservation Clustered diffusion with dynamic data aggregation Centre National d’Etudes Spatiales Center at nearest source Computer operating properly Complex programmable logic device Carrier sense multiple access with collision avoidance Clear to send Cumulative sum Digital-to-analog converter Delay bounded medium access control

xvii

xviii

DBP DCF DCO DCS DOD DOP DPM DS DSDV DSF DSP DSR DT-MSM DTN EADAT ECN EDD EDLC EEDC EGS EH EM-EH EMACS EMI ESR FAR FFT FLAMA FPA FR FRTS FV GIF GIT GLPK GMRE GPIO HCL HEED HEH I2C ID IDC IMD

List of Acronyms

Derivative-based prediction Distributed coordination function Digitally controlled oscillator Data collection and location system Depth of discharge Dilution of precision Dynamic probabilistic model Data send Highly dynamic destination-sequenced distance-vector routing Damage Sensitive Feature Digital signal processing Data success ratio Delay-tolerant mobile sink model Delay-tolerant networking Energy-aware distributed aggregation tree Explicit contention notification Enhanced directed diffusion Electric double-layer capacitor Energy-efficient data collection Electronic grade silicon Energy harvesting Electromagnetic energy harvester EYES-medium access control protocol for WSNs Electromagnetic interference Equivalent series resistance FloodNet adaptive routing Fast Fourier transform Flow-aware medium access Fast path algorithm Flame retardant Future request to send Frequency to voltage Graphics interchange file Greedy incremental tree GNU Linear Programming Kit Greedy maximum residual energy General-purpose input/output High contention level Hybrid energy-efficient distributed clustering Hybrid energy harvesting Inter-Integrated Circuit Identification Insulation-displacement connector Implantable biomedical device

List of Acronyms

ISI JPEG JTAG L-MAC LCL LCS LEACH Li-ion LiPo LPL MACAW MANET MC3 MDC MDS MEH MEMS MILP MPP MPPT MR MRE MS MSEMS MSM MSN MSPR MTS MULE NAMA NAV NiCd NiMH NOAA NP OEM ONR OOK OWFA P-MOSFET PAC/C PAMAS PANEL PCB PDF

Information Sciences Institute Joint Photographic Experts Group Joint Test Action Group Lightweight medium access protocol Low contention level Location-based clustering scheme Low-energy adaptive clustering hierarchy Lithium-ion Lithium polymer Low power listening Media access protocol for wireless LAN Mobile ad hoc network Multicamera coordination and control Mobile data collector Minimal dominating set Micro-energy harvester Microelectromechanical system Mixed integer linear programming Maximum power point Maximum power point tracking Mobile relay Mean relative error Mobile sink Macro-sensor electromechanical system Mobile sink model Maximum slot number Multiple shortest path routing More to send Mobile ubiquitous LAN extension Node activation multiple access Network allocation vector Nickel–cadmium Nickel–metal hydrid National Oceanic and Atmospheric Administration Neighbor protocol Original equipment manufacturer Office of Naval Research ON-OFF keying Optimal wakeup frequency assignment P-type metal–oxide–semiconductor field-effect transistor Power-aware computing and communications Power-aware multi-access protocol with signaling Position-based aggregator node election Printed circuit board Probability density function

xix

xx

PEGASIS PLA PLC POR PREMON PTX PRX PSM PTT PTZ PV PWM RBS RC RCA RFID RINAS RLE RLS RM ROI RTC RTS RTWAC S-LEC S-MAC SAF SB SC SEP SHM SI SLA SMA SMP SMPS SNGF SNR SOI SPIN SSM STC SWIM T-MAC TAG

List of Acronyms

Power-efficient gathering in sensor information system Piecewise linear approximation Programmable logic controller Polynomial regression Prediction-based monitoring Primary transmitter Primary receiver Power saving mode/Power saving mechanism Platform terminal transmitter Pan-tilt-zoom Photovoltaics Pulse width modulation Reference broadcast synchronization Reservoir capacitor Reservoir capacitor array Radio frequency identification Restricted input network activation scheme Run length encoding Recursive least square Random movement Region of interest Real-time clock Request to send Radio triggered wakeup with addressing capability Sequential lossless entropy compression Sensor-MAC Similarity-based adaptive framework Solar biscuit Switched capacitor Schedule exchange protocol Structural health monitoring Standard international Sealed lead acid SubMiniature version A Sensor Management Protocol Switched-mode power supply Stateless non-deterministic geographic forwarding Signal-to-noise ratio Silicon on insulator Sensor protocols for information Static sink model Standard testing condition Shared wireless infostation model Timeout-MAC Tiny aggregation service

List of Acronyms

TDFN TDMA TEG TEH TMPO TRAMA TSR TTL UASN UCLA USC VAR VEH WDT WID WLAN WMN WPAN WSF WSN WSN-ME l-TEG

Thin, dual-in-line flat package, no lead Time-division multiple access Thermoelectric power generator Thermal energy harvesting Topology management by priority ordering Traffic-adaptive medium access protocol Total solar radiation Time to live Underwater acoustic sensor network University of California, Los Angeles University of Southern California Value-added reseller Vibration-based energy harvester Watchdog timer Wireless impedance device Wireless LAN Wireless mesh network Wireless personal area network Wakeup schedule function Wireless sensor network WSN with mobile element Micro-scale thermoelectric generator

xxi

List of Figures

Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig.

1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.1

Fig. 2.2 Fig. 2.3 Fig. 2.4 Fig. 2.5 Fig. 2.6 Fig. 2.7 Fig. 2.8 Fig. 2.9 Fig. 2.10 Fig. 2.11 Fig. 2.12 Fig. 2.13

Mobile ad hoc network (Cordeiro and Agrawal 2002) . . . . . . Three-tier architecture for wireless mesh networks . . . . . . . . . Components of a sensor node (Akyildiz et al. 2002) . . . . . . . Architecture of WSNs [based on (Tilak et al. 2002)] . . . . . . . Fastest runners with different metrics . . . . . . . . . . . . . . . . . . . IEEE 802 standards with focus on IEEE 802.15 . . . . . . . . . . . Wireless standards space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Protocol stack of WSNs (Wang and Balasingham 2010) . . . . Instances of a linear wireless network (Holland et al. 2011) . . Energy harvesting architectures (Sudevalayam and Kulkarni 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Power and energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Harvesting energy from the environment . . . . . . . . . . . . . . . . Energy conversion mechanisms [based on (Basagni et al. 2013)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Piezoelectric transducer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Electromagnetic transducer [based on (Amirtharajah and Chandrakasan 1998)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . Energy harvesting from thermal differences (Vullers et al. 2010) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Resonant inductive coupling (Akhtar and Rehmani 2015) . . . Possible power harvesting from body-centered sources [based on (Starner 1996)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . Principles of Helmholtz resonator-based energy scavenger (Kinsler et al. 1999) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fabricated Helmholtz resonator-based energy scavenger (Kim et al. 2009b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A 2-DOF motion mechanism to harvest vibration energy in arbitrary directions in a plane (Zhu et al. 2011) . . . . . . . . . The MEMS packaged harvester compared to an Australian dollar (Zhu et al. 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . .

7 8 11 12 19 22 23 25 27

.. .. ..

44 46 49

.. ..

63 66

..

67

.. ..

69 73

..

75

..

78

..

79

..

81

..

81

. . . . . . . . .

xxiii

xxiv

List of Figures

Fig. 2.14 Fig. 2.15 Fig. 2.16 Fig. Fig. Fig. Fig. Fig.

3.1 4.1 4.2 4.3 4.4

Fig. 4.5 Fig. 4.6 Fig. 4.7 Fig. 4.8 Fig. 4.9 Fig. 4.10 Fig. 4.11 Fig. 4.12 Fig. Fig. Fig. Fig.

4.13 4.14 4.15 4.16

Fig. 4.17 Fig. 4.18 Fig. 4.19 Fig. 4.20 Fig. 4.21 Fig. 4.22 Fig. Fig. Fig. Fig. Fig.

4.23 4.24 4.25 4.26 4.27

Proposed 2-DOF MEMS EM-EH system (Tao et al. 2016) . . Proposed 3-DOF MEMS EM-EH (Liu et al. 2012) . . . . . . . . Redrawing the structure of the energy supply modules in a wireless sensor node . . . . . . . . . . . . . . . . . . . . . . . . . . . . Energy management techniques (Anastasi et al. 2009b) . . . . . Duty-cycling approach taxonomy (Anastasi et al. 2009) . . . . . Example of a virtual grid (Xu et al. 2001) . . . . . . . . . . . . . . . State transitions in GAF (Xu et al. 2001) . . . . . . . . . . . . . . . . Average number of hops ± standard deviation (bars) versus average number of active neighbors in range for different distances D (Zorzi and Rao 2003b) . . . . . . . . . . . Span positioning (Chen et al. 2002) . . . . . . . . . . . . . . . . . . . . Network self-configuration example (Cerpa and Estrin 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ASCENT state transitions (Cerpa and Estrin 2004) . . . . . . . . Power management approach taxonomy (Anastasi et al. 2009) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Low-power listen mode (Schurgers et al. 2002) . . . . . . . . . . . Interference due to aggressive wakeup (Schurgers et al. 2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Separate wakeup and data frequencies using two radios (Schurgers et al. 2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Separate data and wakeup along time (Schurgers et al. 2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PTW functioning (Anastasi et al. 2009) . . . . . . . . . . . . . . . . . Radio-triggered power management (Anastasi et al. 2009) . . . Network and traffic model (Keshavarzian et al. 2006) . . . . . . Fully synchronized wakeup pattern (Keshavarzian et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . Shifted even and odd wakeup pattern (Keshavarzian et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . Forward ladder wakeup pattern (Keshavarzian et al. 2006) . . . Slot assignment under the (7, 3, 1)-design (Zheng et al. 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Relationship between the wakeup schedule and the communication schedule (Zheng et al. 2003) . . . . . . . . . . . . . Forwarding candidate set (Paruchuri et al. 2004) . . . . . . . . . . Entry fields in a neighbor list maintained by each node (Paruchuri et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . TRAMA time organization (Rajendran et al. 2006) . . . . . . . . FLAMA time organization (Rajendran et al. 2005). . . . . . . . . Traffic flows (Rajendran et al. 2005) . . . . . . . . . . . . . . . . . . . Periodic listen and sleep (Ye et al. 2004) . . . . . . . . . . . . . . . . Node synchronization (Ye et al. 2004) . . . . . . . . . . . . . . . . . .

.. ..

88 89

. . . . .

91 105 110 112 112

. . . . .

. . 116 . . 119 . . 121 . . 122 . . 133 . . 137 . . 138 . . 139 . . . .

. . . .

140 141 142 143

. . 148 . . 149 . . 150 . . 158 . . 159 . . 163 . . . . . .

. . . . . .

163 169 177 180 186 186

List of Figures

Fig. 4.28 Fig. 4.29 Fig. 4.30 Fig. 4.31 Fig. 4.32 Fig. 4.33 Fig. 4.34 Fig. 4.35 Fig. 4.36 Fig. 4.37 Fig. 4.38 Fig. 4.39 Fig. 4.40 Fig. 4.41 Fig. 4.42 Fig. 4.43 Fig. 4.44 Fig. 4.45 Fig. 4.46 Fig. 4.47 Fig. 4.48 Fig. 4.49 Fig. 4.50 Fig. 4.51

Timing relationship between a receiver and multiple senders (Ye et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Adaptive listen reduces sleep latency by at least half (Ye et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ten-hop linear topology (Ye et al. 2004) . . . . . . . . . . . . . . . . T-MAC basic data exchange (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FRTS packet exchange (Van Dam and Langendoen 2003) . . . Taking priority on full buffers (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Part of the simulation network (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nodes-to-sink at a 100-Bytes message length (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nodes-to-sink options at a 20-Bytes message length (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Event-based unicast pattern at a 20-Bytes message length (Van Dam and Langendoen 2003) . . . . . . . . . . . . . . . . . . . . . D-MAC chain transmission (Lu et al. 2004) . . . . . . . . . . . . . . Data prediction to reduce sleep delay (Lu et al. 2004) . . . . . . Sleep delay due to interference between two sending nodes (Lu et al. 2004). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A random 1000 m
500 m topology with 50 nodes (Lu et al. 2004). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mean packet latency for a data gathering tree at a different number of sources (Lu et al. 2004) . . . . . . . . . . . . . . . . . . . . . Energy consumption for a data gathering tree at a different number of sources (Lu et al. 2004) . . . . . . . . . . . . . . . . . . . . . Data delivery ratio for a data gathering tree at a different number of sources (Lu et al. 2004) . . . . . . . . . . . . . . . . . . . . . Tradeoff between energy, latency, and throughput for a data gathering tree under different traffic loads (Lu et al. 2004) . . . CCA effectiveness for a typical wireless channel (Polastre et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Application operations performed when the radio is turned ON (Polastre et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Measured throughput of each protocol with no duty-cycle under a contended channel (Polastre et al. 2004) . . . . . . . . . . Measured power consumption for maintaining a given throughput in a ten-node network (Polastre et al. 2004) . . . . . Effective energy consumption per byte (Polastre et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . End-to-end latency (Polastre et al. 2004) . . . . . . . . . . . . . . . .

xxv

. . 189 . . 192 . . 194 . . 198 . . 201 . . 202 . . 203 . . 203 . . 204 . . 205 . . 209 . . 212 . . 213 . . 214 . . 216 . . 216 . . 217 . . 217 . . 220 . . 221 . . 225 . . 226 . . 227 . . 228

xxvi

List of Figures

Fig. 4.52 Fig. 4.53 Fig. 4.54 Fig. 4.55 Fig. 4.56 Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig.

5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10

Fig. Fig. Fig. Fig. Fig. Fig.

5.11 5.12 5.13 5.14 5.15 5.16

Fig. 5.17 Fig. 5.18 Fig. Fig. Fig. Fig. Fig. Fig.

5.19 5.20 5.21 5.22 5.23 5.24

Fig. 5.25 Fig. 5.26 Fig. 5.27

Effect of latency on power consumption (Polastre et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Time frame rule (Rhee et al. 2008) . . . . . . . . . . . . . . . . . . . . . . . Throughput comparison thru the one-hop MICA2 benchmark (Rhee et al. 2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Throughput comparison thru the two-hop MICA2 benchmark (Rhee et al. 2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Throughput comparison thru the multihop MICA2 benchmark (Rhee et al. 2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Data-driven approach taxonomy (Anastasi et al. 2009) . . . . . . . . Data reduction approach taxonomy (Anastasi et al. 2009) . . . . . Data aggregation in a WSN (Ozdemir and Xiao 2009) . . . . . . . . Tree-based data aggregation (Ozdemir and Xiao 2009) . . . . . . . Directed diffusion (Ozdemir and Xiao 2009) . . . . . . . . . . . . . . . Cluster-based data aggregation (Ozdemir and Xiao 2009) . . . . . LEACH clustering protocol (Fasolo et al. 2007) . . . . . . . . . . . . . Aggregation paths over a ring structure (Fasolo et al. 2007). . . . . . Tributaries’ and deltas’ protocol (Fasolo et al. 2007) . . . . . . . . . Block diagram of the compressor/uncompressor schemes (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . . . . Data prediction approach taxonomy (Anastasi et al. 2009) . . . . . Ken approach (Chu et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . Two-dimensional Gaussian linear model (Chu et al. 2006) . . . . . Disjoint-Cliques model (Chu et al. 2006) . . . . . . . . . . . . . . . . . . Average model (Chu et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . Experimentation outcomes from Intel Research Lab (Chu et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Experimentation outcomes from Botanical Gardens (Chu et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dual prediction scheme with error threshold emax = 0.5 °C (Le Borgne et al. 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sensor radio modes (Han et al. 2004) . . . . . . . . . . . . . . . . . . . . . Data collection process (Han et al. 2004) . . . . . . . . . . . . . . . . . . Active–listening model (AL) (Han et al. 2004) . . . . . . . . . . . . . . Active–sleeping model (AS) (Han et al. 2004) . . . . . . . . . . . . . . Active–listening–sleeping model (ALS) (Han et al. 2004) . . . . . Energy consumption and query response time comparison (Han et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Impact of Ta adaptation on system performance (Han et al. 2004). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Impact of range size adaptation on system performance (Han et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Range of pv ð^pv Þ versus the number of readings to be predicted (Δ) (Goel et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

229 234 240 241 242 260 260 261 263 265 267 268 271 273 278 286 287 292 293 294 296 297 310 318 319 321 323 326 329 330 330 337

List of Figures

Fig. 5.28 Fig. 5.29 Fig. 5.30 Fig. 5.31 Fig. 5.32 Fig. 5.33 Fig. 5.34 Fig. 5.35 Fig. 5.36 Fig. 5.37 Fig. 5.38 Fig. 5.39 Fig. 5.40 Fig. 5.41

Fig. 5.42 Fig. Fig. Fig. Fig. Fig.

5.43 5.44 5.45 5.46 5.47

Fig. 5.48 Fig. 6.1 Fig. 6.2 Fig. 6.3 Fig. 6.4 Fig. 6.5 Fig. 6.6

Energy saving versus model accuracy (Goel et al. 2006) . . . . Ips versus BER and m/d (Goel et al. 2006) . . . . . . . . . . . . . . . Energy-efficient data acquisition taxonomy (based on Anastasi et al. 2009) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Frequency change recognition (Alippi et al. 2007) . . . . . . . . . Sampling rate versus message loss rate (Alippi et al. 2007) . . MRE for low-frequency capacitance versus message loss rate (Alippi et al. 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MRE for high-frequency capacitance versus message loss rate (Alippi et al. 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MRE for temperature versus message loss rate (Alippi et al. 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Original and reconstructed high-frequency capacitance (Alippi et al. 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A scenario of coordination and control of multiple PTZ cameras (Natarajan et al. 2015). . . . . . . . . . . . . . . . . . . . . . . . Multi-camera systems architectures (Natarajan et al. 2015) . . . Functionalities of camera nodes (Natarajan et al. 2015) . . . . . Key features of WSNs for SHM (Kijewski-Correa and Su 2009) . . . . . . . . . . . . . . . . . . . . . . . Multiscale approach for SHM. a Multi-scale network applied to a beam (Kijewski-Correa et al. 2005). b Picturization of multiscale WSN [based on (Kijewski-Correa and Su 2009)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Operation of the multi-scale network [based on (Kijewski-Correa et al. 2005)] . . . . . . . . . . . . . . . . . . . . . . . . Model-based querying (Deshpande et al. 2004) . . . . . . . . . . . Placement of WSN nodes in the tunnel (Raza et al. 2012) . . . Value and time tolerances (Raza et al. 2012) . . . . . . . . . . . . . Derivative-based prediction (Raza et al. 2012) . . . . . . . . . . . . Comparison between DBP and PLA, SAF, POR (Raza et al. 2012) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . General framework for sensor energy management (Alippi et al. 2009) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Architecture of a WSN-MSE with relocatable nodes (Di Francesco et al. 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . Architecture of WSN-MEs with MDCs (Di Francesco et al. 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . Architecture of a WSN-ME with mobile peers (Di Francesco et al. 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mobility-based approach taxonomy (Anastasi et al. 2009b) . . Shortest paths from a sensor to the sink (Wang et al. 2005) . . Dataflows received and transmitted a node i (Wang et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xxvii

. . 338 . . 338 . . 340 . . 344 . . 346 . . 346 . . 347 . . 347 . . 348 . . 353 . . 354 . . 355 . . 357

. . 358 . . . . .

. . . . .

359 362 371 372 373

. . 374 . . 387 . . 402 . . 403 . . 404 . . 405 . . 408 . . 411

xxviii

Fig. 6.7 Fig. 6.8 Fig. 6.9 Fig. 6.10 Fig. 6.11 Fig. 6.12 Fig. 6.13 Fig. 6.14 Fig. Fig. Fig. Fig. Fig. Fig.

6.15 6.16 6.17 6.18 6.19 6.20

Fig. 6.21 Fig. 6.22 Fig. Fig. Fig. Fig.

6.23 6.24 6.25 6.26

Fig. 6.27 Fig. 6.28 Fig. 6.29 Fig. 6.30 Fig. 6.31 Fig. 6.32 Fig. 6.33 Fig. 6.34 Fig. 6.35

List of Figures

Network size versus lifetime (Wang et al. 2005) . . . . . . . . . . Improvement ratio in network lifetime when the sink is mobile (Wang et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sensor nodes sink communication (Papadimitriou and Georgiadis 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sink node placement (Papadimitriou and Georgiadis 2006) . . Average lifetime for various networks sizes (Papadimitriou and Georgiadis 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Average sink sojourn times for various network sizes (Papadimitriou and Georgiadis 2006) . . . . . . . . . . . . . . . . . . . Typical WSN scenarios (Basagni et al. 2008) . . . . . . . . . . . . . Sink optimum routes obtained by constraints Eqs. 6.33–6.36 (Basagni et al. 2008). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Neighboring sink sites (Basagni et al. 2008). . . . . . . . . . . . . . Average network lifetime versus tmin (Basagni et al. 2008) . . Average data latency versus tmin (Basagni et al. 2008) . . . . . . Average overhead per node (Basagni et al. 2008). . . . . . . . . . SSM, MSM, and DT-MSM instances (Yun and Xia 2010) . . Lifetimes of MSM and DT-MSM for various radii of sink coverage (Yun and Xia 2010) . . . . . . . . . . . . . . . . . . . . . . . . . Lifetime versus the number of sink locations (Yun and Xia 2010) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lifetime versus the number of sink locations (Yun and Xia 2010) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lifetime versus transmission range (Yun and Xia 2010) . . . . Three-tier MULE architecture [based on (Jain et al. 2006)] . . Queue model for MULE architecture (Jain et al. 2006) . . . . . Amount of time a sensor is in contact with a MULE (Jain et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Effect of scaling l on performance metrics (Jain et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Effect of scaling, K, the amount of data transferred between a MULE and a sensor (Jain et al. 2006) . . . . . . . . . . . . . . . . . . Effect of the mobility models (Jain et al. 2006) . . . . . . . . . . . Energy consumption of MULE and ad hoc network models (Jain et al. 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mobile relay at work (Wang et al. 2008a) . . . . . . . . . . . . . . . Dividing the nodes in concentric circles around the sink (Wang et al. 2008a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Traffic distribution for a randomly deployed network (Wang et al. 2008a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Network lifetime for nodes randomly deployed on a circular region (Wang et al. 2008a) . . . . . . . . . . . . . . . . . . . . . . . . . . . Average lifetime improvement (Wang et al. 2008a) . . . . . . . .

. . 413 . . 414 . . 416 . . 420 . . 421 . . 422 . . 425 . . . . . .

. . . . . .

429 433 434 435 436 439

. . 444 . . 445 . . . .

. . . .

446 447 450 453

. . 456 . . 459 . . 460 . . 461 . . 463 . . 465 . . 467 . . 471 . . 472 . . 473

List of Figures

Fig. 6.36 Fig. 6.37 Fig. 6.38 Fig. 6.39 Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig.

6.40 6.41 6.42 6.43 7.1 7.2 7.3 7.4

Fig. 7.5 Fig. 7.6 Fig. 7.7 Fig. 7.8 Fig. Fig. Fig. Fig.

7.9 7.10 7.11 7.12

Fig. 7.13 Fig. 7.14 Fig. 7.15 Fig. 7.16 Fig. 7.17 Fig. 7.18 Fig. 7.19 Fig. 7.20 Fig. 7.21 Fig. 7.22 Fig. 7.23

Route dilation compared (Wang et al. 2008a) . . . . . . . . . . . . . Compared network lifetime for different approaches (Wang et al. 2008a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lifetime improvement of adding one mobile relay versus adding more energy to static sensors (Wang et al. 2008a) . . . Adding static sinks to improve network lifetime (Wang et al. 2008a). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Coverage issue in mobile WSNs (Zhu et al. 2014) . . . . . . . . . Localization issue in mobile WSNs (Zhu et al. 2014). . . . . . . Target tracking issue in mobile WSNs (Zhu et al. 2014) . . . . Data gathering issue in mobile WSNs (Zhu et al. 2014) . . . . . ZebraNet structure (Zhang et al. 2004) . . . . . . . . . . . . . . . . . . Version 3 of ZebraNet architecture (Zhang et al. 2004) . . . . . The 200
300
1:2500 ZebraNet node (Zhang et al. 2004) . . . . ZebraNet power consumption during a periodic data sampling (Zhang et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Solar cells module (Zhang et al. 2004) . . . . . . . . . . . . . . . . . . Collared plains zebra at Sweetwaters Game Reserve (Zhang et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Prometheus architecture (Jiang et al. 2005) . . . . . . . . . . . . . . . Perpetual Prometheus self-sustaining Telos mote (Jiang et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Charging circuit block diagram (Jiang et al. 2005) . . . . . . . . . Three basic modes in SB [based on (Minami et al. 2005)] . . . Communication timing of a SB node (Minami et al. 2005) . . Multihop communication in ordinary mode [based on (Minami et al. 2005)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Randomization algorithm (Minami et al. 2005) . . . . . . . . . . . Communication in emergency mode (Minami et al. 2005) . . . Block diagram of the SB node [based on (Minami et al. 2005)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Activity of the SB node (Minami et al. 2005) . . . . . . . . . . . . SB nodes placement and experimental setup (Minami et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Measured IV characteristics of the Solar World 4-4.0-100 solar panel (Raghunathan et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . Harvesting-aware power management . . . . . . . . . . . . . . . . . . . Coordinated energy harvesting framework for a distributed system (Raghunathan et al. 2005) . . . . . . . . . . . . . . . . . . . . . . Heliomote solar harvesting sensor node (Raghunathan et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . Top and bottom sides of Heliomote PCB (Raghunathan et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Everlast block diagram (Simjee and Chou 2006) . . . . . . . . . .

xxix

. . 473 . . 474 . . 475 . . . . . . . .

. . . . . . . .

475 479 480 480 481 490 491 492

. . 497 . . 500 . . 505 . . 507 . . . .

. . . .

512 513 517 518

. . 519 . . 520 . . 521 . . 522 . . 523 . . 525 . . 527 . . 531 . . 532 . . 533 . . 533 . . 539

xxx

Fig. Fig. Fig. Fig. Fig.

List of Figures

7.24 7.25 7.26 7.27 7.28

Fig. 7.29 Fig. 7.30 Fig. 7.31 Fig. 7.32 Fig. 7.33 Fig. 7.34 Fig. 7.35 Fig. 7.36 Fig. 7.37 Fig. 7.38 Fig. 7.39 Fig. 7.40 Fig. 7.41 Fig. 7.42 Fig. 7.43 Fig. 7.44 Fig. 7.45 Fig. 7.46 Fig. 7.47 Fig. 7.48 Fig. 7.49 Fig. 7.50

Everlast prototype board (Simjee and Chou 2006) . . . . . . . . . PFM regulator (Simjee and Chou 2006) . . . . . . . . . . . . . . . . . Average PFM regulator efficiency (Simjee and Chou 2006) . . PFM controller and DC circuitry (Simjee and Chou 2006) . . . Direct capacitor charging versus PFM-regulated charging (Simjee and Chou 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Voltage and current tracking constants calculated over one day (Simjee and Chou 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MPPT comparison: DC sweep versus Voc method (Simjee and Chou 2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-day stress test (Simjee and Chou 2006) . . . . . . . . . . . . . Components of AmbiMax platform (Park and Chou 2006a) . . Architecture of an energy harvesting subsystem in AmbiMax platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . AmbiMax board (Park and Chou 2006a) . . . . . . . . . . . . . . . . MPPT using a switching regulator and hysteresis comparator (Park and Chou 2006a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hysteresis comparator configuration of MPP tracker (Park and Chou 2006a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Control and charger (Park and Chou 2006a) . . . . . . . . . . . . . . AmbiMax prototype (Park and Chou 2006a) . . . . . . . . . . . . . Sunflower architecture (Stanley-Marbell and Marculescu 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sunflower PCB (Stanley-Marbell and Marculescu 2007) . . . . Average measured voltage drop per PIN diode (Stanley-Marbell and Marculescu 2007) . . . . . . . . . . . . . . . . . Voltage at output of TPS61070 regulator, the super-capacitor (Stanley-Marbell and Marculescu 2007) . . . . . . . . . . . . . . . . . Charging the super-capacitor with a single PIN photodiode (Stanley-Marbell and Marculescu 2007) . . . . . . . . . . . . . . . . . Micro-solar system architecture and parameters (Taneja et al. 2008). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . HydroWatch and HydroSolar (Taneja et al. 2008) . . . . . . . . . Equivalent electrical circuit for a PV module (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV curves of solar panel at different lux conditions (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PR experimental curves of solar panel at different lux conditions (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . Equivalent electrical circuit of the thermal energy harvester (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV curves of TEG at different thermal gradients (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

539 540 542 543

. . 546 . . 547 . . 547 . . 549 . . 550 . . 551 . . 551 . . 553 . . 553 . . 555 . . 557 . . 560 . . 561 . . 568 . . 568 . . 569 . . 571 . . 575 . . 580 . . 583 . . 584 . . 585 . . 586

List of Figures

Fig. 7.51 Fig. 7.52 Fig. 7.53

Fig. 7.54

Fig. 7.55

Fig. 7.56

Fig. 7.57

Fig. 7.58 Fig. 7.59

Fig. 7.60

Fig. 8.1 Fig. 8.2 Fig. 8.3 Fig. 8.4 Fig. 9.1

PR experimental curves of TEG at different thermal gradients (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equivalent electrical circuit of the proposed HEH system (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Experimental and simulated electrical power harvested from parallel solar and thermal energy sources (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV and PR curves of HEH system at fixed solar irradiance of 380 lx (3 W/m2) and different thermal conditions of 5–10 K (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV and PR curves of HEH system at fixed solar irradiance of 1010 lx (3 W/m2) and different thermal conditions of 5–10 K (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV and PR curves of HEH system at fixed thermal condition of ΔT = 5 K and varying solar irradiances of 380–1010 lx (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PV and PR curves of HEH system at fixed thermal condition of ΔT = 10 K and varying solar irradiances of 380–1010 lx (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Functional block diagram of HEH system (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Efficiency of HEH boost converter at fixed voltage reference-based MPPT and varying temperature difference (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Efficiency of HEH boost converter at fixed voltage reference-based MPPT and varying solar irradiance (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . European badger “Meles meles” (Hillman 2016) . . . . . . . . . . Heterogeneous wildlife monitoring network (Dyo et al. 2010) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Badger detection node and RFID tag potted in epoxy and collar mounted (Dyo et al. 2010) . . . . . . . . . . . . . . . . . . . . . . Second design version (Dyo et al. 2010) . . . . . . . . . . . . . . . . PowerBench architecture (Haratcherev et al. 2008) . . . . . . . . .

xxxi

. . 587 . . 588

. . 589

. . 591

. . 592

. . 593

. . 594 . . 595

. . 598

. . 598 . . 612 . . 613 . . 616 . . 627 . . 643

List of Tables

Table 1.1 Table 1.2 Table 2.1 Table 2.2 Table 2.3 Table 2.4 Table 2.5 Table 2.6 Table 2.7 Table Table Table Table Table Table

4.1 4.2 4.3 4.4 4.5 4.6

Table 4.7 Table 4.8 Table 4.9 Table 4.10 Table 5.1

Sensor and mesh nodes characteristics . . . . . . . . . . . . . . . . . . . ISM bands defined by ITU-R . . . . . . . . . . . . . . . . . . . . . . . . . . Rechargeable battery technologies compared [based on (Taneja et al. 2008)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Characteristics of energy sources (Sudevalayam and Kulkarni 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Energy densities for motion and vibration transducers (Roundy 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Motion and vibration transducers compared (Roundy 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Human energy expenditures for selected activities (Starner 1996) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Characteristics and generation of indoor energy sources (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Performance of energy harvesters under indoor and outdoor conditions (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . Performance analysis of Span (Chen et al. 2002) . . . . . . . . . . . Characterization of radio power (Schurgers et al. 2002) . . . . . . Typical current draw values (Crossbow 2002) . . . . . . . . . . . . . Transceiver TR1001 data (van Hoesel and Havinga 2004) . . . . Control message contents (van Hoesel and Havinga 2004) . . . . . Time and current consumption to satisfy primitive operations of the monitoring application (Polastre et al. 2004) . . . . . . . . . Parameters for a monitoring application running B-MAC (Polastre et al. 2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Default settings of Z-MAC parameters (Rhee et al. 2008) . . . . Average energy consumption during the setup operations in the multihop MICA2 testbed (Rhee et al. 2008). . . . . . . . . . . . Duty-cycling techniques classified . . . . . . . . . . . . . . . . . . . . . . Dictionary used (Marcelloni and Vecchio 2009) . . . . . . . . . . .

10 22 52 59 68 68 74 83 84 118 137 164 174 174 222 224 239 243 248 279

xxxiii

xxxiv

List of Tables

Table 5.2 Table 5.3 Table 5.4 Table 5.5

Table 5.6 Table 5.7 Table 5.8 Table 5.9 Table Table Table Table

5.10 5.11 6.1 7.1

Table 7.2 Table 7.3 Table 7.4 Table 7.5 Table 7.6 Table Table Table Table

7.7 8.1 8.2 8.3

Number of samples of the four smooth datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . Statistical characteristics of the four smooth datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . Compression ratios obtained by LEC on the four smooth datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . Number of packets to deliver the uncompressed and compressed versions of the four datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Compression ratios obtained by S-LZW on the four smooth datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . Complexity of LEC against S-LZW (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Statistical characteristics of three non-smooth datasets (Marcelloni and Vecchio 2009) . . . . . . . . . . . . . . . . . . . . . . Compression ratios for LEC against S-LZW for three non-smooth datasets (Marcelloni and Vecchio 2009) . . . . . . Symbols used (Han et al. 2004) . . . . . . . . . . . . . . . . . . . . . . Data-driven techniques classified . . . . . . . . . . . . . . . . . . . . . Mobility-based techniques classified . . . . . . . . . . . . . . . . . . . Conditions for Prometheus perpetual operation (Jiang et al. 2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sample power modes (Simjee and Chou 2006) . . . . . . . . . . Comparison of Sunflower to contemporary sensor platforms (Stanley-Marbell and Marculescu 2007) . . . . . . . . . . . . . . . . Operating voltage ranges and power budgets of the active components (Stanley-Marbell and Marculescu 2007) . . . . . . Node lifetime using energy storage elements without recharge (Taneja et al. 2008) . . . . . . . . . . . . . . . . . . . . . . . . Technical characteristics of solar panel used (Tan and Panda 2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Energy harvesting projects compared . . . . . . . . . . . . . . . . . . Routing table (Dyo et al. 2010) . . . . . . . . . . . . . . . . . . . . . . Comparative design evolution (Dyo et al. 2010) . . . . . . . . . Compared average cost to maintain each stage for four weeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . 280 . . 281 . . 282

. . 283 . . 283 . . 284 . . 284 . . . .

. . . .

284 320 378 482

. . 514 . . 548 . . 562 . . 563 . . 576 . . . .

. . . .

583 601 624 626

. . 632

Part I

Concepts and Energy Harvesting

Chapter 1

Wireless Sensor Networks Essentials

Good beginnings lead to happy endings, most of the time …

Beginnings are usually uneasy, sometimes stiff, the acquaintance with newness does not go without tensity, the first year in school, in college, at work, the first year of marriage, the early months of retirement, and even when first time using a new gadget. Some fear change, a new TV, watch, mobile phone, software, color, and brand. Befriending own habits, as human nature, grows with years; juniors are usually more receptive and adaptive. The first fifteen minutes of a movie are decisive, to stay or leave right away. The first chapter is the hardest; it introduces the author, the book, and the topic. Writing is not dumping words and machinely composing sentences, it is a live dialog between the author and the audience, and they see each other in their minds, while writing and while reading issues, debates, controversies, questions and answers, noise, smiles, brain storming, head scratching. This chapter bears his task with willingness, enthusiasm, and good will. Significant developments in scalable standards are now pacing adoption and presenting wireless sensor networks (WSNs) in applications welcomed at IT, industry, home, work, …, everywhere. Wireless sensors can be deployed quickly in an ad hoc fashion and used to report environmental changes, ensure the efficiency of industrial processes in an oil refinery, determine how much power the blade servers in a data center are using, or tell if the refrigerator is still as energy-efficient as when it was purchased. In the 15 years that WSNs have been around, improvements in their architecture and protocols have continued to push applications to the mainstream. Semiconductor technology continues to follow Moore’s law, providing smaller, more powerful, and cheaper wireless devices. There are now established and reliable low-power standards supporting the multiplicity of WSNs applications. The Internet, the largest known network, has extended into the world of low-power embedded WSNs devices. Stepwise, WSNs energy considerations and perspectives are to be introduced with depth and focus.

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_1

3

4

1.1

1 Wireless Sensor Networks Essentials

Sensing, Senses, Sensors

Sensing is what distinguishes the living from stones to rocks. Alive creatures have several levels and ways of sensing, without sensing there is no communication with the outside word, and there is no life. Lecturing on zoology or botany is not an objective, but a quick reminder on senses of the living is recalled (Birds and Blooms 2013). Many animals see the world completely differently to humans. Being able to see helps animals locate food, move around, find mates, and avoid predators, whether they live at the bottom of the ocean or soar high in the sky. Eyesight is important for most animals and nearly all animals can see; 95% of all species have eyes. Some animals live in complete darkness in caves or underground, where they cannot see anything; their eyes often no longer work, but they have developed an extra-sensitive sense of touch to feel their way around. However, only two animal groups have evolved the ability to hear: vertebrates like mammals, birds, and reptiles and arthropods such as insects, spiders, and crabs. No other animals can hear. Some animals have a remarkable sense of hearing, finely tuned to where and how they live; many animals hear sounds that humans cannot. Human senses of smell and taste are feeble compared to those of many other animals, a keen sense of smell allows animals to find food and mates, as well as to stay out of danger, and it can stop an animal wandering into a rival’s territory or help it find its way. Animals communicate using visual signals, sounds, touch, smells, and taste. Vision, touch, and taste work well over short distances, but sounds travel much farther and scent marks can last long after the animal has moved on. Sometimes, the aim is to deceiving, blending into the background, and pretending to be a twig or playing dead; animals give out all sorts of false information to avoid danger or help catch their next meal. Their tricks and deceptions vary from camouflage and mimicry to distracting, startling, scaring, and confusing others (National Museum Scotland 2019). An insect’s acute sense of smell enables it to find mates, locate food, avoid predators, and even gather in groups. Insects have sense organs for taste, touch, smell, hearing, and sight. Some insects have sense organs for temperature and humidity as well as stresses and movements of their body parts. Some insects rely on chemical cues to find their way to and from a nest, or to space themselves appropriately in a habitat with limited resources. Insects, you may have noticed, do not have noses. So, how are they able to sense the faintest of scents in the wind? Antennae sometimes are called “feelers.” However, antennae as primarily “smellers” are the insect’s noses because they are covered with many organs of smell. These organs help the insect to find food, a mate, and places to lay eggs. Insects even can decide which direction to fly by using their sense of smell (National Museum of Natural History 2019). How do fish sense movement? Fish have the five senses that people have, but have a sixth sense that is more than a sense of touch. Fish have a row of special cells inside a special canal along the surface of the fish’s skin. This is called the

1.1 Sensing, Senses, Sensors

5

“lateral line” which allows them to detect water vibrations. This sixth sense allows fish to detect movement around them and changes in water flow. Detecting movement helps fish find prey or escape from predators. Detecting changes in water flow help fish chose where to swim (petMED 2019). What about birds? They depend less on the senses of smell and taste than people do. The odors of food, prey, enemies, or mates quickly disperse in the wind. Birds possess olfactory glands, but they are not well developed in most species, including the songbirds in our backyards. The same is true for taste, which is related to smell. While humans have 9000 taste buds, songbirds have fewer than 50. That means the birds we feed around must locate their food by sight or touch, two senses that are highly developed in birds (Birds and Blooms 2013). Plants, unlike animals, do not have ears, eyes, or tongues to help them feel and acquire information from their environment. But without being helpless, they do sense their environment in other ways and respond accordingly. Plants can detect various wavelengths and use colors to tell them what the environment is like. When a plant grows in the shadow of another, it will send a shoot straight up toward the light source, and it has also been shown that plants know when it is day and when it is night. Leaf pores on plants open up to allow photosynthesis during the daytime and close at night to reduce water loss. Plants also respond to ultraviolet light by producing a substance that is essentially a sunscreen so that they do not get sunburned. Plants can sense weather changes and temperatures as well. Plants have specific regulators, plant hormones, minerals, and ions that are involved in cell signaling and are important in environmental sensing. In fact, without these, the plants will not grow properly (Urbina 2018). Reminding of human senses is easy, the use of eye contacts, the eye attraction to what is beautiful, the love of perfumes, the appreciation of beautiful music, the relieving touch of softness, and the tantalizing taste of sweeties. It is all senses. Human interaction with the environment is an eternal task that grows and expands with the expansion of ambitions, with technology. This book is interested in presenting wireless sensor networks (WSNs) in comprehensive details that are far beyond what birds, insects, and mammals can. As an opening start, this chapter will present a thorough survey of WSNs.

1.2

Toward Wireless Sensor Networks

With the recent technological advances in wireless communications, processor, memory, radio, low power, highly integrated digital electronics, and micro-electromechanical systems (MEMS), it has become possible to significantly develop tiny and small size, low power, and low-cost multifunctional sensor nodes (Warneke and Pister 2002). A WSN is a network that is made of tens to thousands of these sensor nodes, which are densely deployed in an unattended environment with the capabilities of sensing, wireless communications, and computations (i.e., collecting and disseminating environmental data) (Akyildiz et al. 2002). These

6

1 Wireless Sensor Networks Essentials

nodes are capable of wireless communications, sensing, and computation (software, hardware, algorithms). So, it is obvious that a WSN is the result of the combination of sensor techniques, embedded techniques, distributed information processing, and communication mechanisms. Functionally, smart sensor nodes are low-power devices equipped with one or more sensors, a processor, memory, power supply, a radio interface, and some additional components that will be detailed later. A variety of mechanical, thermal, biological, chemical, optical, and magnetic sensors may be attached to the sensor node to measure properties of the environment. Since the sensor nodes have limited memory and are typically deployed in difficult-to-access locations, a radio interface is implemented for wireless communication to transfer the data to a basestation (e.g., a laptop, a personal handheld device, or an access point to a fixed infrastructure). Battery is the main power source in a sensor node, also a secondary power supply that harvests power from the environment such as solar panels may be added to the node depending on the appropriateness for the environment where the sensor will be deployed (Yick et al. 2008). Regarding their practicality and low cost, WSNs have great potential for many applications in scenarios such as military target tracking and surveillance (Yick et al. 2005), natural disaster relief (Castillo-Effen et al. 2004), biomedical health monitoring (Gao et al. 2005), and hazardous environment exploration and seismic sensing (Wener-Allen et al. 2006). In military target tracking and surveillance, a WSN can assist in intrusion detection and identification. Specific examples include spatially correlated and coordinated troop and tank movements. With natural disasters, sensor nodes can sense and detect the environment to forecast disasters before they occur. In biomedical applications, surgical implants of sensors can help monitor a patient’s health. For seismic sensing, ad hoc deployment of sensors along the volcanic area can detect the development of earthquakes and eruptions. Energy is the driver and concern of living beings that have the need to eat and drink, and of modern technologies that need gas, winds, and sun. Noteworthy, one of the most important WSN limitations is energy conservation; therefore, the main WSN’s focus is on power conservation through appropriate optimization of communication and operation management. Several analyses of energy-efficient use for sensor networks have been realized, and several algorithms that lead to efficient transport layer protocols have been proposed. What is the size of a WSN and where to place nodes? The environment plays a key role in determining the size of the WSN network, the deployment scheme, and the network topology. The network size varies with the monitored environment. For indoor environments, fewer nodes are required to form a network in a limited space; whereas, outdoor environments may require more nodes to cover a larger area. An ad hoc deployment is preferred over preplanned deployment when the environment is inaccessible by humans or when the network is composed of hundreds to thousands of nodes. Obstructions can also limit communication between nodes, which in turn affects the network connectivity, or topology. The position of sensor nodes is not usually predetermined, although the application can provide some

1.2 Toward Wireless Sensor Networks

7

guidelines and insights that can lead to the construction of an optimal design that satisfies application requirements and meets wireless network limitations. To go from here and there, a better route is to be selected, and several routing, power management, and data dissemination protocols have been designed for WSNs, depending on both their architecture and the applications they are intended to support. WSN protocols support the proliferation of WSNs and efficiently make them an integral constituent of daily life. To make WSNs practically useful and functioning, these protocols are designed to overcome the unique constraints of small memory, tiny size, limited energy, and to fulfill standards of scalability, adaptivity, fault tolerance, low, latency, and robustness. In the coming section, an overview of MANETs is provided as a step that leads to WSNs.

1.3

Mobile Ad Hoc Networks (MANETs)

At first, it is needed to strengthen up basics; a mobile ad hoc network (MANET) is one that comes together as needed, not necessarily with the support of an existing Internet infrastructure or any fixed station, it is an autonomous system of mobile hosts serving as routers and connected by wireless links (Cordeiro and Agrawal 2002). This contrasts the single-hop cellular network that supports the need for wireless communication by installing basestations as access points, such that the communication between wireless nodes relies on the wired backbone and the fixed basestations. In a MANET, there is no infrastructure, and the network topology changes unpredictably since nodes are free to move. As for the mode of operation, ad hoc networks are peer-to-peer multihop mobile wireless networks where information packets are transmitted in a store and forward manner from source to destination via intermediate nodes as shown in Fig. 1.1. Topology changes as the nodes move, for instance, as node MH2 changes its point of attachment from MH3 to MH4 other nodes must follow the new route to forward packets to MH2. It is to be clear that not all nodes are within radio reach of each other; otherwise there

MH2 MH2 MH4

MH3

Asymmetric link

MH5 Symmetric link MH1

MH6

Fig. 1.1 Mobile ad hoc network (Cordeiro and Agrawal 2002)

MH7

8

1 Wireless Sensor Networks Essentials

would not be any routing problem. Bidirectional links between nodes indicate that they are within radio range of each other, for instance, MH1 and MH3. Unidirectional links indicate that a node may transmit while the other cannot, for instance, MH4 can send to MH7, while MH7 cannot. The following sections go further in the WSNs journey.

1.4

Wireless Mesh Networks (WMNs)

Mesh network architectures have been conceived by both industry and academia. A wireless mesh network is a fully wireless network that employs multihop communications to forward traffic to and from wired Internet entry points. Different from flat ad hoc networks, a wireless mesh network (WMN) introduces a hierarchy in the network architecture by the implementation of dedicated nodes (wireless routers) communicating among each other and providing wireless transport services to data traveling from users to other users or to access points (access points are special wireless routers with a high-bandwidth wired connection to the Internet backbone). As shown in Fig. 1.2, the network of wireless routers forms a wireless backbone tightly integrated into the mesh network, which provides multihop connectivity between nomadic users and wired gateways. The meshing among wireless routers and access points creates a wireless backhaul communication system, which provides each mobile user with a low-cost, high-bandwidth, and easy multihop

Internet Wired/Wireless connections Access points

Wireless routers

Wireless connections

Nomadic users

Fig. 1.2 Three-tier architecture for wireless mesh networks

1.4 Wireless Mesh Networks (WMNs)

9

interconnection service with a limited number of Internet entry points, and with other wireless mobile users. Backhaul is used to indicate the service of forwarding traffic from the user originator node to an access point from which it can be distributed over the external network, the Internet in this case. The mesh network architecture addresses the emerging market requirements for building wireless networks that are highly scalable and cost-effective, offering a solution for the easy deployment of high-speed ubiquitous wireless Internet. Mesh networking has more than a benefit (Bruno et al. 2005): • Reduction of installation costs. Currently, one of the major efforts to provide wireless Internet, beyond the boundaries of indoor WLANs, is through the deployment of WiFi hot spots. Basically, a hot spot is an area that is served by a single WLAN or a network of WLANs, where wireless clients access the Internet through an 802.11-based access point. The downside of this solution is a tolerable increase in the infrastructure costs, because a cabled connection to the wired backbone is needed for every access point in the hot spot. As a consequence, the hot spot architecture is costly, unscalable, and slow to deploy. On the other hand, building a mesh wireless backbone enormously reduces the infrastructural costs because the mesh network needs only a few access points connected to the wired backbone. • Large-scale deployment. In the recently standardized WLAN technologies (i.e., 802.11a and 802.11g), increased data rates have been achieved by using more spectrally efficient modulation schemes. However, for a specific transmit power, shifting toward more efficient modulation techniques reduces coverage, i.e., the further from the access point the lower the data rate available. Moreover, for a fixed total coverage area, more access points should be installed to cover small-sized cells. Obviously, this miniaturization of WLANs cells further hinders the scalability of this technology, especially in outdoor environments. On the other hand, multihop communications offer long-distance communications via hopping through intermediate nodes. Since intermediate links are short, these transmissions could be at high data rates, resulting in increased throughput compared to direct communications. The wireless backbone can realize a high degree of spatial reuse through wireless links covering longer distance at a higher speed than conventional WLAN technologies. • Reliability. The wireless backbone provides redundant paths between each pair of endpoints, significantly increasing communications reliability eliminating single points of failure and potential bottleneck links within the mesh. Network fault tolerance is increased against potential problems such as node crash, path failure due to temporary obstacles or external radio interference, by the existence of multiple possible destinations (i.e., any of the exit points toward the wired Internet), and alternative routes to these destinations. • Self-management. The adoption of peer-to-peer networking to build a wireless distribution system provides all the advantages of ad hoc networking, such as self-configuration and self-healingness. Consequently, network setup is automatic and transparent to users. For instance, when adding additional nodes in the

10

1 Wireless Sensor Networks Essentials

mesh, these nodes use their meshing functionalities to automatically discover all possible wireless routers and determine the optimal paths to the wired network. In addition, the existing wireless routers reorganize, taking into account the new available routes. Thus, the network can easily be expanded, because the network self-reconfigures to assimilate the new elements. With the differences between WSN and WMN, many similarities coexist: • The goal of any WSN and WMN, is to create and maintain network connectivity as easy as possible, in order to get as many data, as fast, easy, secure as needed from source to destination node(s), while consuming the least possible number of resources, such as the wireless spectrum, node energy, node memory, node processing power, and financial budget. • Multihop networks are created, which requires some form of node addressing and a routing protocol. Many popular WSN and WMN technologies share the limited 2400–2500 MHz ISM band of the wireless spectrum. Table 1.1 compares sensor and mesh nodes (Bouckaert et al. 2010).

Table 1.1 Sensor and mesh nodes characteristics Sensor nodes General

Target form factor Antenna Power consumption Power Price

Network

RAM/ROM Processing power Bandwidth Interface(s) Max packet size IP capabilities Sleeping schemes Delay per hop Mobility

Mesh nodes 3

Small or tiny O (mm )

Larger O (cm3)

Integrated O (mW)

External O (W)

Small battery or energy harvesting Relatively cheap (a few dollars or less) KBytes Very limited

Unlimited due to external power source Relatively expensive (50–$500 and up) Mbytes Relatively high

Low (a few Mbps and frequently less) Single, often proprietary Small O (bytes)

Relatively high (several Mbps) Single or multiple, often standardized Larger O (kbytes)

Limited or none Often used

IP capable Rarely used

O (ms) to several seconds None to highly mobile

O (ms) Most often limited or none

1.5 Closer Perspective to WSNs

1.5 1.5.1

11

Closer Perspective to WSNs Wireless Sensor Nodes

To get closer to how a WSN is built, an insight into a sensor node is to come first. Specifically, a sensor node is made up of basic components as shown in Fig. 1.3: • Sensing units. Sensing units are usually composed of two subunits, sensors and analog to digital converters (ADCs). The analog signals produced by the sensors based on the observed phenomenon are converted to digital signals by the ADC and then fed into the processing unit. • Processing unit. The processing unit is generally associated with a small storage unit and manages the procedures that make the sensor node collaborates with the other nodes to carry out the assigned sensing tasks. • Transceiver unit. A transceiver unit connects the node to the network. • Power unit. Power units may be supported by a power scavenging unit such as solar cells. • Application-dependent additional components such as a location finding system, a power generator, and a mobilizer. Most of the sensor network routing techniques and sensing tasks require the knowledge of location with high accuracy, thus it is common that a sensor node has a location finding system. A mobilizer may sometimes be needed to move sensor nodes when it is required to carry out the assigned tasks. All of these subunits may need to fit into a matchbox-sized module whose size may be smaller than even a cubic centimeter, which is light enough to remain suspended in the air. Added to the size, there are also some other stringent specifications of sensor nodes (Khan et al. 1999):

Location finding system

Sensing unit Sensor

ADC

Mobilizer

Processing unit Processor

Transceiver

Storage

Power unit

Fig. 1.3 Components of a sensor node (Akyildiz et al. 2002)

Power generator

12

1 Wireless Sensor Networks Essentials

• Consume extremely low power. • Operate in high volumetric densities. • Have low production cost, can be easily replaced, and the malfunction of any does not halt other sensors. • Are autonomous and operate unattended. • Are adaptive to the environment.

1.5.2

Architecture of WSNs

The term architecture has been adopted to describe the activity of designing any kind of system, and it is the complex or carefully designed structure of something; one of its common uses is in describing information technology, such as computer architecture and network architecture. The architecture of WSNs is built up of main entities as shown in Fig. 1.4: • The sensor nodes that form the sensor network. Their main objectives are making discrete, local measurement about the phenomenon surrounding these sensors, forming a wireless network by communicating over a wireless medium, and collecting data and routing data back to the user via a sink (basestation). • The sink (basestation) communicates with the user via Internet or satellite communication. It is located near the sensor field or well-equipped nodes of the sensor network. Collected data from the sensor field are routed back to the sink by a multihop infrastructureless architecture through the sink.

Monitored area Sensor node WSN

Sink (basestation)

Target User

Fig. 1.4 Architecture of WSNs [based on (Tilak et al. 2002)]

1.5 Closer Perspective to WSNs

13

• The phenomenon, which is an entity of interest to the user, is sensed and analyzed by the sensor nodes. • The user who is interested in obtaining information about a specific phenomenon to measure/monitor its behavior. Although many protocols and algorithms have been proposed for traditional wireless ad hoc networks, they are not well suited for the unique features and application requirements of sensor networks, as detailed in this section. For further illustration, the differences between WSNs and MANETs are outlined below (Akyildiz et al. 2002): • The number of sensor nodes in WSNs can be several orders of magnitude higher than the nodes in MANETs. • Sensor nodes are densely deployed. • Sensor nodes are prone to failures. • The topology of a sensor network changes very frequently. • Unlike a node in MANETs, a sensor node may not have a unique global IP address due to the numerous numbers of sensors and the resulting high overhead. • Sensor nodes as deployed in high numbers are extremely cheap and considerably tiny, unlike MANET nodes (e.g., PDAs, laptops, etc.). • The communication paradigm used in WSNs is broadcasting; whereas, MANETs are based on point-to-point communications. • The topology of a WSN changes very frequently. • Limited energy and bandwidth conservation are the main concern in WSN protocols design, which is not really worrisome in MANETs.

1.6

Types of WSNs

WSNs can be deployed on ground, underground, and underwater. Five functional types can be distinguished: specifically, terrestrial, underground, underwater, multimedia, and mobile WSNs (Yick et al. 2005). What follows provides the details of each type.

1.6.1

Terrestrial WSNs

Terrestrial WSNs are deployed in a given area (Yick et al. 2008). There are two ways to deploy sensor nodes on WSNs:

14

1 Wireless Sensor Networks Essentials

• In unstructured WSN, which contains a dense collection of sensor nodes, sensor nodes may be deployed in an ad hoc manner into the field; once deployed, the network is left unattended to perform monitoring and reporting functions. In an unstructured WSN, network maintenance such as managing connectivity and detecting failures is difficult since there are so many nodes. • In structured WSN, all or some of the sensor nodes are deployed in a preplanned manner. The advantage of a structured network is that fewer nodes can be deployed with lower network maintenance and management cost. Fewer nodes are beneficially deployed since they are placed at specific locations to provide coverage while ad hoc deployment can have uncovered regions. Sensor nodes are deployed on the sensor field within reach of the transmission range of each other and at densities that may be as high as 20 nodes/m3. Densely deploying hundreds or thousands of sensor nodes over a field requires maintenance of topology along three phases: • Predeployment and deployment phase. Sensor nodes may either be thrown in the deployment field as a mass from an airplane or an artillery shell, or placed one by one by a human or a robot. • Post deployment phase. After deployment, topology changes due to change in sensor nodes position, reachability (that may be affected by jamming, noise, moving obstacles, etc.), remaining energy, malfunctioning, and task details. • Redeployment of additional nodes. Additional sensor nodes can be redeployed to replace malfunctioning nodes or to account for changes in task dynamics. In a terrestrial WSN, reliable communication in a dense environment is a must. Sensor nodes must be able to effectively communicate data back to the basestation. While battery power is limited and may not be rechargeable, terrestrial sensor nodes, however, can be equipped with a secondary power source such as solar cells, and it is important for sensor nodes to conserve energy. For a terrestrial WSN, energy can be conserved with multihop optimal routing, short transmission range, in-network data aggregation, eliminating data redundancy, minimizing delays, and using low duty-cycle operations.

1.6.2

Underground WSNs

Underground WSNs consist of a number of sensor nodes buried underground or in a cave or mine used to monitor underground conditions (Li and Liu 2007, 2009; Li and Akyildiz 2007). Additional sink nodes are located above ground to relay information from the sensor nodes to the basestation. An underground WSN is more expensive than a terrestrial WSN in terms of equipment, deployment, and maintenance. Underground sensor nodes are expensive because appropriate equipment parts must be selected to ensure reliable communication through soil, rocks, water, and other mineral contents. The underground environment makes

1.6 Types of WSNs

15

wireless communication a challenge due to signal losses and high levels of attenuation. Unlike terrestrial WSNs, the deployment of an underground WSN requires careful planning and energy and cost considerations. Energy is an important concern in underground WSNs. Like terrestrial WSN, underground sensor nodes are equipped with a limited battery power, and once deployed into the ground, it is difficult to recharge or replace a sensor node’s battery. As usual, a key objective is to conserve energy in order to increase the network lifetime, which can be achieved by implementing efficient communication protocol.

1.6.3

Underwater Acoustic Sensor Networks (UASNs)

Underwater acoustic sensor networks (UASNs) technology provides new opportunities to explore the oceans, and consequently it improves understanding of the environmental issues, such as the climate change, the life of ocean animals, and the variations in the population of coral reefs. Additionally, UASNs can enhance the underwater warfare capabilities of the naval forces since they can be used for surveillance, submarine detection, mine countermeasure missions, and unmanned operations in the enemy fields. Furthermore, monitoring the oil rigs with UASNs can help taking preventive actions for the disasters such as the rig explosion that took place in the Gulf of Mexico in 2010. Last but not the least, earthquake and tsunami forewarning systems can also benefit from the UASN technology (Erol-Kantarci et al. 2011). Ocean monitoring systems have been used for the past several decades, where traditional oceanographic data collection systems utilize individual and disconnected underwater equipment. Generally, this equipment collects data from their surroundings and sends these data to an onshore station or a vessel by means of satellite communications or underwater cables. In UASNs, this equipment is replaced by relatively small and less expensive underwater sensor nodes that house various sensors onboard, e.g., salinity, temperature, pressure, and current speed sensors. The underwater sensor nodes are networked, unlike the traditional equipment, and they communicate underwater via acoustics. In underwater, radio signals attenuate rapidly; hence, they can only travel to short distances while optical signals scatter and cannot travel far in adverse conditions, as well. On the other hand, acoustic signals attenuate less, and they are able to travel further distances than radio signals and optical signals. Consequently, acoustic communication emerges as a convenient choice for underwater communications. However, it has several challenges (Heidemann et al. 2006): • The bandwidth of the acoustic channel is low, hence the data rates are lower than they are in terrestrial WSNs. Data rates can be increased by using short-range communications, which means more sensor nodes, will be required to attain a certain level of connectivity and coverage. In this respect, large-scale UASN poses additional challenges for communication and networking protocols.

16

1 Wireless Sensor Networks Essentials

• The acoustic channel has low link quality, which is mostly due to the multipath propagation and the time variability of the medium. • Furthermore, the speed of sound is slow (approximately 1500 m/s) yielding large propagation delay. • In mobile UASNs, the relative motion of the transmitter or the receiver may create the Doppler effect. • UASNs are also energy limited similar to other WSNs. Due to the above challenges, UASNs rooms research studies in novel medium access, network, transport, localization, synchronization protocols, and architectures (Jornet et al. 2008; Vuran and Akyildiz 2008; Lee et al. 2010; Ahna et al. 2011). The design of network and management protocols is closely related to the network architecture, and various UASN architectures have been proposed in the literature. Moreover, localization has been widely addressed since it is a fundamental task used in tagging the collected data, tracking underwater nodes, detecting the location of an underwater target, and coordinating the motion of a group of nodes. Furthermore, location information can be used to optimize the medium access and routing protocols (Chandrasekhar et al. 2006; Erol-Kantarci et al. 2011; Zhou et al. 2011). Underwater sensor nodes must be able to self-configure and adapt to harsh ocean environment, and they are equipped with a limited battery, which cannot be replaced or recharged. The issue of energy conservation for underwater WSNs involves developing efficient underwater communication and networking techniques.

1.6.4

Multimedia WSNs

Multimedia WSNs have been proposed to enable monitoring and tracking of events in the form of multimedia such as video, audio, and imaging (Akyildiz et al. 2007). Multimedia WSNs consist of a number of low-cost sensor nodes equipped with cameras and microphones. These sensor nodes interconnect with each other over a wireless connection for data retrieval, processing, correlation, and compression. They are deployed in a preplanned manner into the environment to guarantee coverage. Challenges in multimedia WSN include as follows: • • • • •

High-bandwidth demand. High-energy consumption. Quality of service (QoS) provisioning. Data processing and compressing techniques. Cross-layer design.

Multimedia content such as a video stream requires high bandwidth in order of the content to be delivered quickly; consequently, high data rate leads to

1.6 Types of WSNs

17

high-energy consumption. Thus, transmission techniques that support high bandwidth and low-energy consumption have to be developed. QoS provisioning is a challenging task in a multimedia WSN due to the variable delay and variable channel capacity. It is important that a certain level of QoS must be achieved for reliable content delivery. In-network processing, filtering, and compression can significantly improve network performance in terms of filtering and extracting redundant information and merging contents. Similarly, cross-layer interaction among protocol layers can improve the processing and delivering of data.

1.6.5

Mobile WSNs

Mobile WSNs consist of a collection of sensor nodes that can move on their own and interact with the physical environment (Di Francesco et al. 2011). There are several comparative issues between mobile and static sensor nodes as follows: • Like static nodes, mobile nodes have the ability to sense, compute, and communicate. • Contrarily, mobile nodes have the ability to reposition and organize themselves in the network. A mobile WSN can start off with some initial deployment, and nodes can then spread out to gather information. Information gathered by a mobile node can be communicated to another mobile node when they are within range of each other. • Another key difference is data distribution. In a static WSN, data can be distributed using fixed routing or flooding while dynamic routing is used in a mobile WSN. Mobility in WSNs is useful for several reasons, as presented in what follows (Anastasi et al. 2009): • Connectivity. As nodes are mobile, a dense WSN architecture is not a pressing requirement. Mobile elements can cope with isolated regions, such that the constraints on network connectivity and on nodes (re)deployment can be relaxed. Hence, a sparse WSN architecture becomes a feasible option. • Cost. Since fewer nodes can be deployed, the network cost is reduced in a mobile WSN. Although adding mobility features to the nodes might be expensive, it may be possible to exploit mobile elements, which are already present in the sensing area (e.g., trains, buses, shuttles or cars) and attach sensors to them. • Reliability. Since traditional (static) WSNs are dense and the communication paradigm is often multihop ad hoc, reliability is compromised by interference and collisions; moreover, message loss increases with the increase in number of hops. Mobile elements, instead, can visit nodes in the network and collect data

18

1 Wireless Sensor Networks Essentials

directly through single-hop transmissions; this reduces not only contention and collisions, but also the message loss. • Energy efficiency. The traffic pattern inherent to WSNs is convergecast, i.e., messages are generated from sensor nodes and are collected by the sink. As a consequence, nodes closer to the sink are more overloaded than others and subject to premature energy depletion. This issue is known as the funneling effect, since the neighbors of the sink represent the bottleneck of traffic. Mobile elements can help reduce the funneling effect, as they can visit different regions in the network and spread the energy consumption more uniformly, even in the case of a dense WSN architecture. However, mobility in WSNs also introduces significant challenges, which do not arise in static WSNs, as illustrated below: • Contact detection. Since communication is possible only when the nodes are in the transmission range of each other, it is necessary to detect the presence of a mobile node correctly and efficiently. This is especially true when the duration of contacts is short. • Mobility-aware power management. In some cases, it is possible to exploit the knowledge on the mobility pattern to further optimize the detection of mobile elements. In fact, if visiting times are known or can be predicted with certain accuracy, sensor nodes can be awake only when they expect the mobile element to be in their transmission range. • Reliable data transfer. As available contacts might be scarce and short, there is a need to maximize the number of messages correctly transferred to the sink. In addition, since nodes move during data transfer, message exchange must be mobility-aware. • Mobility control. When the motion of mobile elements can be controlled, a policy for visiting nodes in the network has to be defined. That is, the path and the speed or sojourn time of mobile nodes have to be defined in order to improve (maximize) the network performance. • Challenges also include deployment, localization, navigation and control, coverage, maintenance, and data processing. Mobile WSN applications include environment monitoring, target tracking, search and rescue, and real-time monitoring of hazardous material. Mobile sensor nodes can move to areas of events after deployment to provide the required coverage. In military surveillance and tracking, they can collaborate and make decisions based on the target. Mobile sensor nodes can achieve a higher degree of coverage and connectivity compared to static sensor nodes. In the presence of obstacles in the field, mobile sensor nodes can plan ahead and move appropriately to unobstructed regions to increase target exposure.

1.7 Performance Metrics of WSNs

1.7

19

Performance Metrics of WSNs

Metric is the standard of measurement, and it varies with the measured environment. Time delay is a widely used metric, it is the time needed to obtain a response after applying certain input, and its units are coarsely seconds, but specifically at which scale? In an electronic environment, time delay units are microseconds and less; in electromechanical environment, they are milliseconds or more; in pure mechanical systems, they are seconds and above. In athletic run sports, speed is the metric, and its unit scale varies with distances, from the 100 m race till the marathon. Generally, speed varies with who is running and where, and a professional human runner spends 2:15 h in a 42.195 km marathon, while a cheetah that is three times faster just needs 25 min to reach the end point; the metric is time, the same, but for humans it is measured in hours, while it is in minutes for the cheetah (Fig. 1.5). One of the metrics for goods is weight, and its unit is kilograms or pounds, but for coal it is a multiplicity of kilograms for home use and tons for the industry. On the other hand, gold weight is calculated in grams or ounces for personal use and in kilograms for gold traders. Lifetime a metric related to living being existence, and it is left for the reader to have some thoughts about it, at least for mind relief. Back to WSNs, for a WSN to perform satisfactorily, some metrics are also defined, measured, and interpreted far from confusion. Several metrics, close to WSN’s characteristics as introduced in the previous sections of this chapter, evaluate sensor network performance. Specifically: • Network lifetime. It is a measure of energy efficiency, and as sensor nodes are battery operated, WSNs protocols must be energy-efficient to maximize system lifetime. System lifetime can be measured by generic parameters such as the time until half of the nodes die, or by application directed metrics, such as when the network stops providing the application with the desired information about the environment; it is also calculated as the time until message loss rate exceeds a given threshold.

(a) Usain Bolt hits 9:58 sec for 100m Fig. 1.5 Fastest runners with different metrics

(b) Cheetah fastest runner on earth

20

1 Wireless Sensor Networks Essentials

• Energy consumption. It is the sum of used energy by all WSN nodes. The consumed energy of a node is the sum of the energy used for communication, including transmitting, receiving, and idling. Assuming each transmission consumes an energy unit, the total energy consumption is equivalent to the total number of packets sent in the network. • Latency. It is the end-to-end delay that implies the average time between sending a packet from the source and the time for successfully receiving the message at the destination. Measurement takes into account the queuing and the propagation delay of the packets. The observer is interested in getting information about the environment within a given delay. The precise units of latency are application-dependent. • Accuracy. It is the freedom from mistake or error, correctness, conformity to truth, and exactness. Obtaining accurate information is the primary objective of the observer. There is a tradeoff between accuracy, latency, and energy efficiency. A WSN should be adaptive such that its performance achieves the desired accuracy and delay with minimal energy expenditure. For example, the WSN task, the application, can either request more frequent data dissemination from the same sensor nodes, or it can direct data dissemination from more sensor nodes with the same frequency. • Fault tolerance. Sensors may fail due to surrounding physical conditions or when their energy runs out. It may be impractical to replace existing sensors; in response, the WSN must be fault-tolerant such that non-serious failures are hidden from the application in a way that does not hinder it. Fault tolerance may be achieved through data replication, as in the SPIN protocol (Xiao et al. 2006). However, data replication itself requires energy; there is a tradeoff between data replication and energy efficiency, generally, data replication should be application-specific, higher priority data according to the application that might be replicated for fault tolerance. • Scalability. As a prime factor, it is WSN adaptability to increased workload that is to include more sensor nodes than what was anticipated during network design. A scalable network is one that can be expanded in terms of the number of sensors, complexity of the network topology, data quality, e.g., sampling rate, sensor sensitivity, and amount of data while the cost of the expansion installation and operational cost, communication time, processing time, power, and reliability is no worse than a linear, or nearly linear, function of the number of sensors (Pakzad et al. 2008). WSN scalability needs to consider an integrated view of the hardware and software. For hardware, scalability involves sensitivity and range of MEMS sensors, communication bandwidth of the radio, and power usage. The software issues include reliability of command dissemination and data transfer, management of large volume of data, and scalable algorithms for analyzing the data. The combined hardware–software issues include high-frequency sampling and the tradeoffs between onboard computations compared with wireless communication between nodes.

1.7 Performance Metrics of WSNs

21

• Network throughput. It is a common metric for all networks. The end-to-end throughput measures the number of packets per second received at the destination. • Success rate. It is also a common metric. It is the total number of packets received at the destinations verses the total number of packets sent from the source.

1.8

WSNs Standards

A standard is a required or agreed level of quality or attainment. There are standards for health, industry, and education. The International Organization for Standardization, known as ISO, is an international standard-setting body composed of representatives from various national standards organizations. Founded on February 23, 1947, long time before WSNs were born, ISO promotes worldwide proprietary, industrial, and commercial standards. The WSN’s standards are tightly coupled to the ISM frequency bands that are recalled in the next paragraph. The Industrial, scientific, and medical (ISM) radio bands were first established in 1947 by the International Telecommunications Union (ITU) in Atlantic City. The ISM bands are defined by the ITU-R in 5.138, 5.150, and 5.280 of the Radio Regulations (ITU 1947). Individual countries use of the designated bands may differ due to variations in national radio regulations. ISM are radio bands (portions of the radio spectrum) reserved internationally for the use of radio frequency (RF) energy for industrial, scientific, and medical purposes other than telecommunications (Table 1.2). Examples of applications in these bands include radio frequency process heating, microwave ovens, and medical diathermy machines. The powerful emissions of these devices can create electromagnetic interference and disrupt radio communication using the same frequency, so these devices were limited to certain bands of frequencies. Wireless sensor standards have been developed with a key design requirement for low-power consumption. The standards define the functions and protocols necessary for sensor nodes to interface with a variety of networks. The IEEE 802.15 is a working group focusing on wireless personal area networks (WPANs); it has seven different approved standards and several ongoing standards discussions that are in different phases of the standardization process (IEEE 2019). All 802.15.x approved standards propose PHY and MAC layers; they do not provide network, transport, or application layers, implying that this task is left for other parties. ZigBee is a company alliance that constructs network and application layers to 802.15.4 devices. Figure 1.6 lists the IEEE 802 standards with a focus on IEEE 802.15. As instances, the 802.15.1 is a standard of the lower transport layers of the Bluetooth stack that contains a MAC and a PHY layer specifications. Task Group 2 has delivered 802.15.2 as a recommended practice for coexistence of WPAN devices with other radio equipment in unlicensed frequency bands. Also, Task Group 3 of 802.15 presented a standard in 2003 that was intended for high-rate

22

1 Wireless Sensor Networks Essentials

Table 1.2 ISM bands defined by ITU-R Frequency range

Bandwidth

Center frequency

Availability

00.000 kHz

150 kHz

150 kHz

75 kHz

6.765 MHz 13.553 MHz 26.957 MHz

6.795 MHz 13.567 MHz 27.283 MHz

30 kHz 14 kHz 326 kHz

6.780 MHz 13.560 MHz 27.120 MHz

40.660 MHz 433.050 MHz

40.700 MHz 434.790 MHz

40 kHz 1.74 MHz

40.680 MHz 433.920 MHz

866.000 MHz

868.000 MHz

2 MHz

867.000 MHz

902.000 MHz

928.000 MHz

26 MHz

915.000 MHz

2.400 GHz

2.4835 GHz

83.5 MHz

2.441 GHz

5.725 GHz

5.875 GHz

150 MHz

5.800 GHz

Region 1 low power, narrow band Subject to local acceptance Radio frequency identification Citizen Band (CB) radio models Radio models Region 1 and subject to local acceptance Region 1. Very narrow band, few channels Region 2 only (with some exceptions) Region 1, Region 2, and Region 3 Region 3 has extended the upper range, additional * 150 MHz

24.000 GHz 24.250 GHz 250 MHz 24.125 GHz 61.000 GHz 61.500 GHz 500 MHz 61.250 GHz Subject to local acceptance 122.000 GHz 123.000 GHz 1 GHz 122.500 GHz Subject to local acceptance 244.000 GHz 246.000 GHz 2 GHz 245.000 GHz Subject to local acceptance Region 1 comprises Europe, Africa, the Middle East west of the Arabian Gulf including Iraq, the former Soviet Union, and Mongolia Region 2 covers the Americas, Greenland, and some of the Eastern Pacific Islands Region 3 contains most of non-former-Soviet Union Asia, east of and including Iran, and most of Oceania

IEEE 802 Local and Metropolitan Area Networks Standard Committee (LMSC) IEEE 802.2 IEEE 802.3 IEEE IEEE IEEE 802.16 IEEE 802.20 Logical Link (Ethernet) 802.11 802.15 Broadband Mobile Control (LLC) Wireless Wireless wireless broadband LANs PANs access wireless access (WLANs) (WPANs) IEEE 802.15.1 IEEE IEEE (WPAN/Bluetooth) 802.15.2 802.15.3 (Coexistence) (High rate WPANs)

IEEE 802.15.4 (Low rate WPANs)

IEEE IEEE 802.15.5 802.15.6 (Mesh (BANs) networking)

Fig. 1.6 IEEE 802 standards with focus on IEEE 802.15

IEEE 802.15.7 Visible Light Communication (VLC)

1.8 WSNs Standards

23

Satellite WRAN IEEE 802.22

Wireless range

WWAN (>10-50 km)

3G

4G

5G

WiMax IEEE 802.16

WMAN (1-5 km) WiFi IEEE 802.11 WLAN (100-500 m)

WPAN (10 m)

0.01

Bluetooth IEEE 802.15.1

High rate WPAN IEEE 802.15.3

ZigBee IEEE 802.15.4

0.1

UWB IEEE 802.15.4a

1

10

100

1,000

10,000

Data rate (Mbps)

Fig. 1.7 Wireless standards space

WPAN with application areas such as multimedia and digital imaging. High rate in this context is transfer rates of 11, 22, 33, 44, and 55 Mbps. Task Group 3 had two subworking groups called 802.15.3a and 802.15.3b, where the former was supposed to present a new PHY based on ultra-wide band (UWB) radio technique, and the latter came up in 2005 with an amendment to the MAC sublayer. In subgroup 3a, two different proposals of UWB techniques were discussed as a new PHY layer, but two industry alliances could not come to a consensus on which one to adopt. Consequently, IEEE decided to postpone further meetings in this subgroup, and there is no UWB PHY standard yet to high-rate WPAN. Task Group 4 of 802.15 has developed a standard intended for low data transfer rates of WPANs as opposite to the high transfer rates of 802.15.3. In Fig. 1.7, the wireless standards space, including IEEE 802.15 standards, is portrayed based on data rate and wireless equipment range.

1.9

Protocol Stack of WSNs

A protocol, etiquette, code of conduct, is a set of rules that govern a certain behavior, in social or diplomatic activities, at work, when driving, etc. Socially, there is a dress for night parties, and there is a way to put off a coat, to sit, eat, and speak. Diplomatic activities are framed in strict protocol rules that determine who comes first, who is next, who will be to the right, and who speaks; deviating from

24

1 Wireless Sensor Networks Essentials

such rules is a serious breach of job duties. A protocol is also the draft of a treaty or agreement. At work, there are limitations to what can be said in public, to what can be worn, and to where to eat or smoke. When driving, there are rules to follow a lane or to change lanes, to surpass, to honk, and to speed limits. Protocol rules may be imposed by administrative regulations, or by social habits, either way they are followed, and monitored, and a person is appreciated with regard to how far he clings to protocol guidelines. In communication networks, protocols govern, determine the functioning specifications and guidelines, and guarantee how networks fulfill their intended use. A WSN is an ad hoc arrangement of multifunctional sensor nodes in a sensor field, disseminated to gather information regarding some phenomenon. Sensor nodes can be densely distributed over a large and maybe remote area and collaborate their efforts to the benefit of the network to the extent that even if a number of nodes malfunction, the network will continue to function. There are two main layouts for WSNs. The first is a star layout where the nodes communicate, in a single hop, directly to the sink whenever possible and peer-to-peer communication is minimal. In the second, information is routed back to the sink via data passing between nodes. This multihop communication is expected to consume less power than single-hop communication because nodes in the sensor field are densely distributed and are relatively close to each other. As previously stated, WSNs differ from traditional ad hoc networks in a few very significant ways: • Power awareness. Because nodes are placed in remote, hard to reach places, it is not feasible to replace dead batteries. All protocols must be designed to minimize energy consumption and preserve the life of the network. • Sensor nodes lack global identifications (IDs), so that the networks lack the usual infrastructure. Attribute-based naming and clustering are used instead. Querying WSNs is done by asking for information regarding a specific attribute of the phenomenon, or asking for statistics about a specific area of the sensor field. This requires protocols that can handle requests for a specific type of information, as well as data-centric routing and data aggregation. • Position of the nodes may not be engineered or predetermined, and therefore, must provide data routes that are self-organizing. A protocol stack for WSNs must support their typical features and singularities. According to Akyildiz et al. (2002), the sensor network protocol stack is much like the traditional protocol stack, with the following layers: application, transport, network, data link, and physical. The physical layer is responsible for frequency selection, carrier frequency generation, signal detection, modulation, and data encryption. The data link layer is responsible for the multiplexing of data streams, data frame detection, medium access, and error control. It ensures reliable point-to-point and point-to-multipoint connections in a communication network.

1.9 Protocol Stack of WSNs

25

The network layer takes care of routing the data supplied by the transport layer. The network layer design in WSNs must consider the power efficiency, data-centric communication, data aggregation, etc. The transport layer helps to maintain the dataflow and may be important if WSNs are planned to be accessed through the Internet or other external networks. Depending on the sensing tasks, different types of application software can be set up and used on the application layer. WSNs must also be aware of several management planes in order to function efficiently, specifically, mobility, power, task, quality of service (QoS), and security management planes. Among them, the functions of task, mobility, and power management planes have been elaborated in (Akyildiz et al. 2002; Wang and Balasingham 2010). The protocol stack and the associated planes used by the sink, cluster head, and sensor nodes are shown in Fig. 1.8. The power management plane is responsible for minimizing power consumption and may turn off functionality in order to preserve energy. The mobility management plane detects and registers movement of nodes so that a data route to the sink is always maintained. The task management plane balances and schedules the sensing tasks assigned to the sensing field, and thus, only the necessary nodes are assigned with sensing tasks and the remainder are able to focus on routing and data aggregation. QoS management in WSNs (Howitt et al. 2006) can be very important if there is a real-time requirement with regard to the data services. QoS management also deals with fault tolerance, error control, and performance optimization in terms of certain QoS metrics. Security management is the process of managing, monitoring, and controlling the security-related behavior of a network. The primary function of security management is in controlling access points to critical or sensitive data. Security management also includes the seamless integration of different security function modules, including encryption, authentication, and intrusion detection.

Fig. 1.8 Protocol stack of WSNs (Wang and Balasingham 2010)

26

1.9.1

1 Wireless Sensor Networks Essentials

Physical Layer

In many WSNs, the number and location of nodes make recharging or replacing the batteries infeasible. For this reason, energy consumption is a universal design issue for WSNs. Much work has been done to minimize energy dissipation at all levels of system design, from the hardware to the protocols to the algorithms. Hence, to the network, it is important to appropriately set parameters of the protocols in the network stack. At the physical layer, the parameters open to the network designer which include modulation scheme, transmit power, and hop distance. The optimal values of these parameters will depend on the channel model. When a wireless transmission is received, it can be decoded with a certain probability of error, based on the ratio of the signal power to the noise power of the channel (i.e., the SNR). As the energy used in transmission increases, the probability of error goes down, and thus, the number of retransmissions goes down. Thus, there exists an optimal tradeoff between the expected number of retransmissions and the transmit power to minimize the total energy dissipated to receive the data (Holland et al. 2011). At the physical layer, there are two main components that contribute to energy loss in a wireless transmission, the loss due to the channel, and the fixed energy cost to run the transmission and reception circuitry (Heinzelman et al. 2002). The loss in the channel increases as a power of the hop distance, while the fixed circuitry energy cost increases linearly with the number of hops. This implies that there is an optimal hop distance where the minimum amount of energy is expended to send a packet across a multihop network. Similarly, there is a tradeoff between the transmit power and the probability of error. In this tradeoff, there are two parameters that a network designer can change to optimize the energy consumed, transmit power, and hop distance. The third option for physical layer parameter selection is much broader than the other two. The coding/modulation of the system determines the probability of transmission success, and changes in the probability of a successful transmission lead to changes in the optimal values for the other physical layer parameters (Wang et al. 2001). To illustrate these physical layer tradeoffs, consider the linear network shown in Fig. 1.9. In this network, a node must send data back to the basestation. The first physical layer consideration is hop distance. In the first case (Network 1), the hop distance is very small, which translates to low per-hop energy dissipation. Because the transmit energy must be proportional to dn where n  2 and d is the distance between the transmitter and receiver, the total transmit energy to get the data to the basestation will be much less using the multihop approach than a direct transmission (Heinzelman et al. 2002). However, in this network, the main factor in the energy dissipation of the transmission is the large number of hops. The fixed energy cost to route through each intermediate hop will cause the total energy dissipation to be high. In the second case (Network 2), the hop distance is very large. With so few hops, there is little drain of energy on the network due to the fixed energy cost. However, there is a large energy drain on the nodes due to the high-energy cost to transmit

1.9 Protocol Stack of WSNs

27

Network 1

Network 2

Network 3 (a) Network 1 has a short hop distance (b) Network 2 has a long hop distance (c) Network 3 has the optimal hop distance Fig. 1.9 Instances of a linear wireless network (Holland et al. 2011)

data over the long individual hop distances. With a large path loss factor, the total energy in this case will far exceed the total energy in the case of short hops. Thus, it is clear that a balance must be struck, as shown in Network 3, so that the total energy consumed in the network is at a minimum.

1.9.2

Data Link Layer

The responsibilities of the data link layer are the multiplexing of data streams, data frame detection, medium access (MAC), and error control. A WSN must have a specialized MAC protocol to address the issues of power conservation and data-centric routing. The MAC protocol must meet two goals. The first is to create a network infrastructure, which includes establishing communication links between may be thousands of nodes and providing the network self-organizing capabilities. The second goal is to fairly and efficiently share communication resources between all the nodes. Existing MAC protocols fail to meet these two goals because power conservation is only a secondary concern in their development. Also, WSNs have no central controlling agent and a much larger number of nodes than traditional ad hoc networks. Any MAC protocol for WSNs must also take into account the ever-changing topology of the sensor network due to node failure and redistribution. Since sensor nodes are usually operated by batteries and left unattended after deployment, power saving is a critical issue in WSNs. Many research efforts in recent years have focused on developing power saving schemes for WSNs. These schemes include power saving hardware design, power saving topology design (Salhieh et al. 2001; Chakrabarti et al. 2003), power-efficient MAC layer protocols (Ye et al. 2002; Zheng et al. 2005; Rajendran et al. 2006; Pang et al. 2012), and network layer routing protocols (Sohrabi et al. 2000; Akkaya and Younis 2005).

28

1 Wireless Sensor Networks Essentials

Designing power-efficient MAC protocols is one of the techniques that prolong the lifetime of the network. In addition to energy efficiency, latency and throughput are also important features for consideration in MAC protocol design for WSNs. Commercial standards like IEEE 802.11 have a power management scheme for ad hoc networks, wherein the nodes remain in idle listening state at low traffic to conserve power; significant power is wasted even in the idle listening mode. Hence, IEEE 802.11 is not suitable for sensor networks. A properly designed MAC protocol allows the nodes to access the channel in a way that saves energy and support QoS.

1.9.3

Network Layer

The network layer in a WSN must be designed with typical considerations in mind, ever existing power efficiency, WSNs are data-centric networks, and WSNs have attribute-based addressing and location awareness. The link layer handles how two nodes talk to each other, while the network layer is responsible for deciding which node to talk to. The simplest design is flooding. When using flooding, each node receiving data repeats it by broadcasting the data to every neighbor unless the max hop lifetime of the data has been reached or the receiving node is the destination. The major support for flooding is the simplicity. It requires no costly topology maintenance or complex route discovery. The shortcomings, however, are substantial as follows: • Implosion. It occurs when two nodes (A and B) share multiple (n) neighbors. Node A will broadcast data to all n of these neighbors. Node B will then receive a copy of the data from each of them. • Overlap, when two nodes share the same sensing region. If a stimulus occurs within this overlap, both nodes will report it. • The last and most crucial problem is resource blindness. Flooding does not take into account available energy resources. Gossiping is an enhancement to flooding. In gossiping, when a node receives data, it randomly chooses a neighbor and sends the data to it. Gossiping avoids the problem of implosion, but does not address the other two concerns and contributes to the latency of the network. A step up from flooding and gossiping is ideal dissemination. In this algorithm, data are sent along a shortest path route from the originating node. Such approach guarantees that every node will receive every piece of information exactly once. No energy is wasted in sending or receiving redundant data. However, the overhead involved in keeping track of the shortest paths is substantial. Also, ideal dissemination does not take into account that some node may not need a particular piece of information; nor does it allow for resource awareness.

1.9 Protocol Stack of WSNs

29

A little more sophisticated family of protocols is sensor protocols for information via negotiation (SPIN). The SPIN family addresses the deficiencies of classic flooding by negotiation and resource adaptation. With more sophisticated and energy-aware techniques for data dissemination, it reduces the amount of energy expended, solves the problems of implosion, overlap, and resource blindness, and ensures that only interested nodes will expend energy to receive data (Kulik et al. 2002; Rehena et al. 2011). Negotiation helps to overcome the problems of implosion and overlap and ensures only useful and desired information is disseminated. In order of negotiation to work, nodes must describe the data to be sent using meta-data. In order of SPIN to be efficient, the meta-data must be significantly shorter than the data being described. Also, meta-data describing two distinguishable pieces of data must be different. Likewise, if two pieces of data are indistinguishable, they will share the same meta-data. The format of the meta-data is not specified by SPIN, but rather application-specific. SPIN-2 is an implementation of SPIN that employs a low-energy threshold. When energy is abundant, the node functions as normal. However, when the resource manager detects that a node power supply is reaching the low-energy threshold, the node will not participate in later stages of the protocol. This prolongs the life of the node and allows it to perform only high priority functions. SPIN is a more sophisticated and energy-aware schema for data dissemination. It reduces the amount of energy expended, solves the problems of implosion, overlap, and resource blindness, and ensures that only interested nodes will expend energy to receive data.

1.9.4

Transport Layer

Transport control protocol for WSNs should account for several concerns (Wang et al. 2005) as follows: • Congestion control and reliability. The more data streams flow from sensor nodes to sinks in WSNs, the more congestion might occur around sinks. Also, there are some high-bandwidth data streams produced by multimedia sensors. Therefore, it is necessary to design effective congestion detection, congestion avoidance, and congestion control mechanisms for WSNs. Although MAC protocol can recover packets loss from bit-error, it has no way to handle packets loss from buffer overflow. Then, the transport protocol for WSNs should have mechanism for packets loss recovery such as ACK and selective ACK as used in TCP protocol so as to guarantee reliability. Reliability under WSNs may have different meaning from traditional networks that generally guarantee correct transmission of every packet. For some application, WSNs only need to correctly receive packets from a certain area, not

30

•

•

•

•

1 Wireless Sensor Networks Essentials

from every sensor nodes in this area, or may be contempt with some ratio of successful transmission from a sensor node. These modified reliability concept motivates the design of different transport control protocols. It would be better to use hop-by-hop mechanism for congestion control and loss recovery since it can reduce packet dropping and conserve energy. The hop-by-hop mechanism can simultaneously lower buffer requirement at intermediate nodes, which suits the limited memory sensor nodes. Simplifying initial connecting process or use connectionless protocol so as to speed up start and guarantee throughput and lower transmission delay. Most of the applications in WSNs are reactive that is passively monitor and wait for event occurring before reporting to sink. These applications may have only few packets for each reporting, and the simple and short initial setup process is more effective and efficient. Avoiding packets dropping as possible to lessen energy wastage. In order to avoid packet dropping, the transport protocol can use active congestion control at the cost of a lower link utility. The active congestion control (ACC) can triggers congestion avoidance before congestion occurs. An example of ACC is to make sender (or intermediate nodes) reduce sending (or forwarding) rate when the buffer size of their downstream neighbors overruns a threshold. Guaranteeing fairness for different sensor nodes so that each sensor node can achieve a fair throughput. Otherwise, the loaded sensor nodes cannot properly report events in their area, which leads to erroneous monitoring, tracking, and control. Enabling cross-layer interaction. If a routing algorithm can notify route failure to the transport protocol, the transport protocol will know that packet loss is not from congestion but from route failure, and consequently, the sender will regulate its current sending rate to guarantee high throughput and low delay.

1.9.5

Application Layer

To address application layer protocols, it is primordial to address some functions that are to be implemented, specifically, data fusion and management, clock synchronization, and positioning. A WSN is intended to be deployed in environments where sensors can be exposed to circumstances that might interfere with provided measurements. Such circumstances include strong variations of pressure, temperature, radiation, and electromagnetic noise. Thus, measurements may be imprecise in such scenarios. Data fusion is used to overcome sensor failures, technological limitations, and spatial and temporal coverage problems. Data fusion is generally defined as the use of techniques that combine data from multiple sources and gather this information in order to achieve inferences, which will be more efficient and potentially more accurate than if they were achieved by means of a single source. The term efficient, in this case, can mean more reliable delivery of accurate

1.9 Protocol Stack of WSNs

31

information, more complete, and more dependable. The data fusion can be implemented in both centralized and distributed systems. In a centralized system, all raw sensor data would be sent to one node, and the data fusion would all occur at the same location. In a distributed system, the different fusion modules would be implemented on distributed components (Abdelgawad and Bayoumi 2012). Communications in WSNs are data-centric, with the objective of delivering collected data in a timely fashion. Also, such networks are resource-constrained, in terms of sensor nodes’ processing power, communication bandwidth, storage space, and energy. This gives rise to new face-offs in information processing and data management in WSNs. In-network data processing techniques, from simple reporting to more complicated collective communications, such as data aggregation, broadcast, multicast, and gossip, are challenging. On the other hand, data collected by sensors can intrinsically be viewed as signals. By exploiting signal processing techniques, collective communications can be done in more energy-efficient ways. Several works deal with data management, Xu et al. (2009) investigate in-network query processing strategies for K nearest neighbor (KNN) queries in location-aware WSNs. Also, Brayne et al. (2008) propose an adaptive query processing mechanism to dynamically adjust query processing in WSNs. Moreover, Akcan and Brönnimann (2007) develop a distributed, weighted sampling algorithm to sample sensing data to reduce energy consumption. By exploring the adaptive model selection algorithms, Le Borgne et al. (2007) derive an adaptive, lightweight, and on-line algorithm for prediction sensing data. Sensed data are of limited usage if it is not accompanied by the coordinates of the sensor position and a time stamp, and this is a primary motive for clock synchronization in WSNs. Data fusion is a prime function that depends also on clock synchronization. For instance, a vehicle going through acoustic sensors can be detected, throughout its path, by different sensor nodes at different moments. A fusion node receiving the raw information from the sensor nodes can refine it by estimating the speed and the direction of the sensed vehicle. For this application, among others, synchronized time stamps together with position information are essential. Also, WSNs are expected to have very small form factors and be cheap such that they can be deployed in very large numbers. Once deployed, WSNs are usually unattended, so battery replacement is impractical, but since they are typically expected to work for extended periods of time, there is no better way to conserve energy but to put the nodes to sleep and to wakeup at the same time to be able to exchange information. Clock synchronization in WSNs is the subject of extensive work (Elson and Römer 2003; Sundararaman et al. 2005; Sun et al. 2006; Sommer and Wattenhofer 2009; Wu et al. 2011). Positioning. Knowledge of the position of the sensing nodes in a WSN is an essential part of many sensor network operations and applications. Sensors reporting monitored data need to also report the location where the information is sensed, and hence, sensors need to be aware of their position. In addition, many network protocols, such as routing, require location information in order to provide the specific protocol service. WSNs may be deployed in hostile environments

32

1 Wireless Sensor Networks Essentials

where malicious adversaries attempt to spoof the locations of the sensors by attacking the localization process. For example, an attacker may alter the distance estimations of a sensor to several reference points, or replay beacons from one part of the network to some distant part of the network, thus providing false localization information. Hence, there is a need to ensure that the location estimation is performed in a robust way, even in the presence of attacks. Furthermore, adversaries can compromise the untethered sensor devices and force them to report a false location to the data collection points. Therefore, a secure positioning system must have a mechanism to verify the location claim of any sensor. Positioning in WSN is a topic of extensive research, leading to numerous positioning systems that provide an estimation of the sensor location (Lazos et al. 2005; Akkaya et al. 2007; Kim et al. 2007; Tennina et al. 2008, 2009; Younis and Akkaya 2008). System administrators interact with WSNs using sensor management protocol (SMP). Unlike many other networks, WSNs consist of nodes that do not have global IDs, and they are usually infrastructureless. Therefore, SMP needs to access the nodes by using attribute-based naming and location-based addressing. SMP is a management protocol that provides the software operation needed to perform several administrative tasks (Akyildiz et al. 2002): • Introducing to the sensor nodes the rules related to data aggregation, attribute-based naming, and clustering. • Exchanging data related to the location finding algorithms. • Time synchronization of the sensor nodes. • Moving sensor nodes. • Turning sensor nodes on and off. • Querying the sensor network configuration and the status of nodes and reconfiguring the sensor network. • Authentication, key distribution, and security in data communications.

1.9.6

Cross-Layer Protocols for WSNs

The severe energy constraints of battery-powered sensor nodes necessitate energy-efficient communication protocols in order to fulfill the application objectives of WSNs. It is much more resource-efficient, according to some research, to have a unified scheme which melts common protocol layer functionalities into a cross-layer module for resource-constrained sensor nodes. A unified cross-layer communication protocol, for efficient and reliable event communication, considers the effects on WSNs of replacing transport, routing, medium access functionalities, and physical layers (wireless channel). A unified cross-layering is such that both the information and the functionalities of traditional communication layers are melted in a single protocol. The objective of

1.9 Protocol Stack of WSNs

33

the proposed cross-layer protocol is highly reliable communication with minimal energy consumption, adaptive communication decisions, and local congestion avoidance. Protocol operation is governed by the concept of initiative determination. Based on this concept, the cross-layer protocol performs received-based contention, local congestion control, and distributed duty-cycle operation in order to realize efficient and reliable communication in WSN. Performance evaluation reveals that the proposed cross-layer protocol significantly improves the communication efficiency and outperforms the traditional layered protocol architectures (Akyildiz et al. 2006).

1.10

Conclusion for Energetic Trip

Sensing is life, WSNs are acquiring snowballing interests in research and industry, and they are infiltrated in day-to-day use. Owing to their requirement of low device complexity as well as slight energy consumption, proper standards are devised to ensure impeccable communication and meaningful sensing. Concerns that WSNs have been unreliable and difficult to use are lessening. But to put a WSN together, a potential user or developer has to be adept in multiple disciplines, hardware, embedded software programming, RF, and enterprise integration, which creates a gap between the application concept and deployed network. What is constantly needed is a way to abstract the complexity of setting up and commissioning a WSN from the ongoing management and data mining of the sensor data itself. As much as WSNs are made easily accessible over conventional IP-based networks, their potential user base will become far vaster and more diverse. A key attribute of WSNs, and the reason they represent the future of intelligent embedded devices, is their ability to be deployed in diverse and varied physical world environments. With no computer-based map of sensor locations, users may be left alone to remember (or guess) where their sensors had been deployed. Sensor network applications, that bind the physical to the logical positioning, allow users to upload an existing floor plan, map, or image into the WSN user interface and then survey an individual sensor node positioned on that map. Once the nodes begin to monitor and collect data on a particular space, thing, or interaction, the map provides context, meaning, and the ability to easily manage the WSN deployment. Once the network is formed, the sensor nodes start collecting data. Collected data need to be accessible, either in a database or directly to an application for display or analysis. This is where the WSNs have taken an experience from the enterprise IT world. Modern enterprise applications communicate and share information using the Web services model, which provides a convenient and scalable way for WSNs to pass collected data to an end user application or remote database. Sensor data can be accessed from a corporate IT network using Web services that build Web pages and API calls to collect data from the WSNs and return them in a well-formed XML to the requestor.

34

1 Wireless Sensor Networks Essentials

There is no shortage of current and potential applications for WSNs, and as a result a wide array of sensor and actuator devices has come on the market. Accommodating that variety of devices, however, is not a trivial challenge. Users should look for an embedded operating system that supports a wide range of leading hardware platforms, while preserving the full capabilities of each. The leading open-source embedded operating system designed for wireless sensor networks, TinyOS, is such an operating system. The OS should include a simple driver framework to support the incorporation of new sensors across multiple platforms. External expansion ports and drivers should also be available to add new kinds of sensors after installation. Accommodation for analog sensors of different types (resistive or inductive) as well as digital sensors (contact switches) is crucial. This makes sensor nodes and WSNs ideal for proof-of-concept and pilot networks where functionality and return on investment (ROI) must be proved before finalizing industrial design and appropriate enclosure in the deployment environment. Protocols are the rules of communication. Several considerations must be taken when developing protocols for WSNs. Traditional thinking, where the focus is on quality of service, is somehow revised. In WSNs, QoSis compromised to conserve energy and preserve the life of the network. WSNs are a kind of “totalitarian” system, every one is for the good of all, no individualism, and the whole network must survive even at the expense of falling sensors. Concern must be accorded at every level of the protocol stack to conserve energy and to allow individual nodes to reconfigure the network and modify their set of tasks according to the resources available. The protocol stack for WSNs consists of five standard protocol layers trimmed to satisfy typical sensors features, namely application layer, transport layer, network layer, data link layer, and physical layer. These layers address network dynamics and energy efficiency. Functions such as localization, coverage, storage, synchronization, security, and data aggregation and compression are network services that enable proper sensors functioning. Implementation of WSNs protocols at different layers in the protocol stack aims at minimizing energy consumption, and end-to-end delay, and maintaining system efficiency. Traditional networking protocols are not designed to meet these WSNs requirements, hence, new energy-efficient protocols have been proposed for all layers of the protocol stack. These protocols employ cross-layer optimization by supporting interactions across the protocol layers. Specifically, protocol state information at a particular layer is shared across all the layers to meet the specific requirements of the WSN. As sensor nodes operate on limited battery power, energy usage is a very important concern in a WSN; there has been significant research focus that revolves around harvesting and energy conservation by minimizing energy consumption. When a sensor node is depleted of energy, it will fade out and disengage from the network, which may significantly impact the performance of the application. Sensor network lifetime depends on the number of active nodes and network connectivity, so energy must be used efficiently in order to maximize the network lifetime.

1.10

Conclusion for Energetic Trip

35

Energy conservation in a WSN maximizes network lifetime and is addressed through efficient reliable wireless communication, smart sensor placement to achieve adequate coverage, security and efficient storage management, and data aggregation and data compression. Such approaches satisfy both the energy constraint andprovideQoS. For reliable communication, services such as congestioncontrol, active buffer monitoring, acknowledgments, and packet loss recovery are necessary to guarantee packet delivery. Communication strength depends on the placement of sensor nodes. Sparse sensor placement may result in long-range transmission and higher energy usage, while dense sensor placement may result in short-range transmission and less energy consumption. Coverage is interrelated to sensor placement. The total number of sensors in the network and their placement determine the degree of network coverage. Depending on the application, a higher degree of coverage may be required to increase the accuracy of the sensed data. One for all, and all for all, that is the main objective of all layers in the WSNs protocol stack.

1.11 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21.

Exercises

What are the components of a wireless sensor node? Detail the specs of a sensor node. How are sensor nodes deployed in a terrain? What are the deployment phases? Define MANETs and explain symmetric and asymmetric links. Describe the architecture of WSNs. Determine the differences between MANETs and WSNs. Detail the characteristics of WMNs. Compare between WSNs and WMNs. What are the types of WSNs? Illustrate the characteristics of UASNs. How is WSNs mobility useful? Define protocol and bring out some instances. What are the considerations and concerns of the WSNs protocol stack? Elaborate on the physical layer for WSNs typical features. How is the data link layer for WSNs different? Explain how is the network layer in WSNs different. What is positioning and clock synchronization? How is data fusion crucial in WSNs? What is the importance of data aggregation for WSNs? Determine the functions of the transport layer in WSNs. How does the typical usage of WSNs affect the application layer?

36

1 Wireless Sensor Networks Essentials

References Abdelgawad, A., and M. Bayoumi. 2012. Data Fusion in WSN. In Resource-Aware Data Fusion Algorithms for Wireless Sensor Networks, vol. 118, 17–35. Springer-Verlag. Ahna, Joon, Affan Syedb, Bhaskar Krishnamacharia, and John Heidemannb. 2011. “Design and Analysis of a Propagation Delay tolerant ALOHA Protocol for Underwater Networks.” Ad Hoc Networks Journal 9 (5): 752–766 (Elsevier). Akcan, H., and H. Brönnimann. 2007. “A New Deterministic Data Aggregation Method for Wireless Sensor Networks.” Signal Processing 87 (12): 2965–2977 (Elsevier). Akkaya, K., and M. Younis. 2005. “A Survey on Routing Protocols for Wireless Sensor Networks.” Ad Hoc Networks 3 (3): 325–349 (Elsevier). Akkaya, K., M. Younis, and W. Youssef. 2007. “Positioning of Base Stations in Wireless Sensor Networks.” Communications Magazine 45 (4): 96–102 (IEEE). Akyildiz, I.F., M.C. Vuran, and O.B. Akan. 2006. “A Cross-Layer Protocol for Wireless Sensor Networks.” In 40th Annual Conference on Information Sciences and Systems, 1102–1107. Princeton, NJ: IEEE. Akyildiz, I.F., T. Melodia, and K.R. Chowdury. 2007. “Wireless Multimedia Sensor Networks: A Survey.” IEEE Wireless Communication 14 (6): 32–39 (IEEE). Akyildiz, I.F., W. Su, Y. Sankarasubramaniam, and E. Cayirci. 2002a. “A Survey on Sensor Networks.” Communications Magazine 40 (8): 102–114 (IEEE). Akyildiz, I.F., W. Su, Y. Sankarasubramaniam, and E. Cayirci. 2002b. “Wireless Sensor Networks: A Survey.” Computer Networks 38 (4): 393–422 (Elsevier). Anastasi, G., M. Conti, M. Di Francesco. 2009. “Reliable and Energy-Efficient Data Collection in Sparse Sensor Networks with Mobile Elements.” Performance Evaluation 66 (12) (December 2009): 791–810 (Elsevier). Birds and Blooms. 2013. Can Birds Smell or Taste? http://www.birdsandblooms.com/Birds/ Summer/Can-Birds-Smell-or-Taste/. Accessed 22 Aug 2013. Bouckaert, Stefan, Eli De Poorter, Benôit Latré, Jeroen Hoebeke, Ingrid Moerman, and Piet Demeester. 2010. “Strategies and Challenges for Interconnecting Wireless Mesh and Wireless Sensor Networks.” Wireless Personal Communications 53 (3) (May 2010): 443–463 (Springer). Brayne, A., A. Lopes, D. Meira, R. Vasconcelos, and R. Menezes. 2008. “An Adaptive in-Network Aggregation Operator for Query Processing in Wireless Sensor Networks.” Journal of Systems and Software 81 (3): 328–342 (Elsevier). Bruno, R., M. Conti, and E. Gregori. 2005. “Mesh Networks: Commodity Multihop Ad Hoc Networks.” Communications Magazine, Mar: 123–131 (IEEE). Castillo-Effen, M., D.H. Quintela, R. Jordan, W. Westhoff, and W. Moreno. 2004. “Wireless Sensor Networks for Flash-Food Alerting.” In The 5th IEEE International Caracas Conference on Devices, Circuits, and Systems, Dominican Republic, 142–146. Caracas: IEEE. Chakrabarti, A., A. Sabharwal, and B. Aazhang. 2003. “Using Predictable Observer Mobility for Power Efficient Design of Sensor Networks.” In Information Processing in Sensor Networks, ed. F. Zhao and L. Guibas, 129–145. Springer-Verlag. Chandrasekhar, V., W.K. Seah, Y.S. Choo, and H.V. Ee. 2006. “Localization in Underwater Sensor Networks: Survey and Challenges.” In The First International Workshop on Underwater Networks (WuWNet), 33–40. Los Angeles, CA: ACM. Cordeiro, C.M., and D.P. Agrawal. 2002. “Mobile Ad Hoc Networking.” In The 20th Brazilian Symposium on Computer Networks, 125–186. Di Francesco, M., S.K. Das, and G. Anastasi. 2011. “Data Collection in Wireless Sensor Networks with Mobile Elements: A Survey.” ACM Transactions on Sensor Networks (TOSN) (ACM) 8 (1) (August 2011). Elson, J., and K. Römer. 2003. “Wireless Sensor Networks: A New Regime for Time Synchronization.” ACM SIGCOMM Computer Communication Review (ACM) 33 (1): 149–154.

References

37

Erol-Kantarci, Melike, Hussein T. Mouftah, and Sema Oktug. 2011. “A Survey of Architectures and Localization Techniques for Underwater Acoustic Sensor Networks.” Communications Surveys and Tutorials 13 (3): 487–502 (IEEE). Gao, T., D. Greenspan, M. Welsh, R.R. Juang, and A. Alm. 2005. “Vital Signs Monitoring and Patient Tracking Over a Wireless Network.” In The 27th IEEE EMBS Annual International Conference, 102–105. Shanghai: IEEE. Heidemann, J., W. Ye, J. Wills, A. Syed, and Y. Li. 2006. “Research Challenges and Applications for Underwater Sensor Networking.” In Wireless Communications and Networking Conference (WCNC), 228–235. Las Vegas, USA: IEEE. Heinzelman, W., A.P. Chandrakasan, and H. Balakrishnan. 2002. “An Application Specific Protocol Architecture for Wireless Microsensor Networks.” IEEE Transactions on Wireless Communication 1 (4): 660–670 (IEEE). Holland, M., T. Wang, B. Tavli, A. Seyedi, and W. Heinzelman. 2011. “Optimizing Physical Layer Parameters for Wireless Sensor Networks.” ACM Transactions on Sensor Networks (TOSN) (ACM) 7 (4) (February 2011): 28:1–28:20. Howitt, I, W.W. Manges, P.T. Kuruganti, G. Allgood, J.A. Gutierrez, and J.M. Conrad. 2006. “Wireless Industrial Sensor Networks: Framework for QoS Assessment and QoS Management.” ISA Transactions 45 (3) (July 2006): 347–359 (Elsevier). IEEE. 2019. IEEE 802.15 Working Group for WPAN. January 1. http://www.ieee802.org/15/. Accessed 10 June 2019. ITU. 1947. ITU. http://www.itu.int/dms_pub/itu-s/oth/02/01/S020100002B4813PDFE.pdf. Accessed 7 Sept 2013. Jornet, M., M. Stojanovic, and M. Zorzi. 2008. “Focused Beam Routing Protocol for Underwater Acoustic Networks.” In The 3rd International Workshop on Underwater Networks (WUWNET), 75–82. San Francisco, CA: ACM. Khan, J.M., R.H. Katz, and K.S.J. Pister. 1999. “Next Century Challenges: Mobile Networking for “Smart Dust””. In The 5th annual ACM/IEEE international Conference on Mobile Computing and Networking (MobiCom),271–278. Seattle, WA: ACM. Kim, S., J.G. Ko, J. Yoon, and H. Lee. 2007. “Multiple-Objective Metric for Placing Multiple Base Stations in Wireless Sensor Networks.” In The 2nd International Symposium on Wireless Pervasive Computing (ISWPC). San Juan, Puerto Rico: IEEE. Kulik, J., W.R. Heinzelman, and H. Balakrishnan. 2002. “Negotiation–Based Protocols for Disseminating Information in Wireless Sensor Networks.” Wireless Networks (ACM) 8 (2/3) (March–May 2002): 169–185. Lazos, L., R. Poovendran, and S. Čapkun. 2005. “ROPE: Robust Position Estimation in Wireless Sensor Networks.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN). Los Angeles, CA: ACM. Le Borgne, Y., S. Santini, and G. Bontempi. 2007. “Adaptive Model Selection for Time Series Prediction in Wireless Sensor Networks.” Signal Processing 87 (12): 3010–3020 (Elsevier). Lee, U., P. Wang, Y. Noh, L. Vieira, M. Gerla, and J.H. Cui. 2010. “Pressure Routing for Underwater Sensor Networks.” In The 29th Conference on Information Communications (INFOCOM), 1676–1684. San Diego, CA: IEEE. Li, L., C.V. Mehmet, and I.F. Akyildiz. 2007. “Characteristics of Underground Channel for Wireless Underground Sensor Networks.” In The 6th Annual Mediterranean Ad Hoc Networking WorkShop, 92–99. Corfu, Greece. Li, M., and Y. Liu. 2007. “Underground Structure Monitoring with Wireless Sensor Networks.” In The 6th International Conference on Information Processing in Sensor Networks (IPSN), 69– 78. Cambridge, MA: ACM/IEEE. Li, Mo, and Yunhao Liu. 2009. “Underground Coal Mine Monitoring with Wireless Sensor Networks.” Transactions on Sensor Networks (ACM) 5 (2) (March 2009). National Museum of Natural History. O. Orkin Insect Zoo. National Museum of Natural History. January 1, 2019. https://naturalhistory.si.edu/exhibits/o-orkin-insect-zoo. Accessed 9 June 2019.

38

1 Wireless Sensor Networks Essentials

National Museum Scotland. Animal Senses. January 1, 2019. http://www.nms.ac.uk/our_ museums/national_museum/explore_the_galleries/natural_world/animal_senses.aspx. Accessed 9 June 2019. Pakzad, S.N., G.L. Fenves, S. Kim, and D.E. Culler. 2008. “Design and Implementation of Scalable Wireless Sensor Network for Structural Monitoring.” Journal of Infrastructure Systems (ASCE) 14: 89–101. Pang, B.M., H.S. Shi, and Y.X. Li. 2012. “An Energy-Efficient MAC Protocol for Wireless Sensor Network.” In Future Wireless Networks and Information Systems, vol. 1. ed. Y. Zhang 163– 170. Springer-Verlag. petMED. How Fish Sense and ‘Feel’. January 1, 2019. https://www.petmd.com/fish/care/evr_fi_ fish_senses (accessed June 9, 2019). Rajendran, V., K. Obraczka, and J.J. Garcia-Luna-Aceves. 2006. “Energy-Efficient, Collision-Free Medium Access Control for Wireless Sensor Networks.” Wireless Networks (ACM) 12 (1) (Feb 2006). Rehena, Z., S. Roy, and N. Mukherjee. 2011. “A Modified Spin for Wireless Sensor Networks.” In The 3rd International Conference on Communication Systems and Networks (COMSNETS), 1– 4. Bangalore, India: IEEE. Salhieh, A., J. Weinmann, M. Kochhal, and L. Schwiebert. 2001. “Power Efficient Topologies for Wireless Sensor Networks.” In International Conference on Parallel Processing, 156–163. Valencia, Spain: IEEE. Sohrabi, K., J. Gao, V. Ailawadhi, and G.J. Pottie. 2000. Protocols for Self-Organization of a Wireless Sensor Network. Personal Communications 7 (5): 16–27 (IEEE). Sommer, P., and R. Wattenhofer. 2009. “Gradient Clock Synchronization in Wireless Sensor Networks.” In The International Conference on Information Processing in Sensor Networks (IPSN), 37–48. San Francisco, CA: ACM. Sun, K., P. Ning, and C. Wang. 2006. “Secure and Resilient Clock Synchronization in Wireless Sensor Networks.” IEEE Journal on Selected Areas in Communications 24 (2): 395–408 (IEEE). Sundararaman, B., U. Buy, and A.D. Kshemkalyani. 2005. “Clock Synchronization for Wireless Sensor Networks: A Survey.” Ad Hoc Networks 3 (3): 281–323 (Elsevier). Tennina, S., M. Di Renzo, F. Graziosi, and F. Santucci. 2009. “ESD: A Novel Optimisation Algorithm for Positioning Estimation of WSNs in GPS-Denied Environments–from Simulation to Experimentation.” International Journal of Sensor Networks (Inderscience Publishers) 6 (3/ 4): 131–156. Tennina, S., M. Di Renzo, F. Graziosi, and F. Santucci. 2008. “Locating Zigbee® Nodes Using the Ti®S Cc2431 Location Engine: A Testbed Platform and New Solutions for Positioning Estimation of WSNs in Dynamic Indoor Environments.” In The First International Workshop on Mobile Entity Localization and Tracking in GPS-Less Environments (MELT), 37–42. (ACM), September. Tilak, S., N. Abu-Ghazaleh, and W. Heinzelman. 2002. “A Taxonomy of Wireless Micro-Sensor Network Models.” ACM SIGMOBILE Mobile Computing and Communications Review (ACM) 6 (2) (April 2002): 28–36. Urbina, R. Encouraging Curiosity and Why it’s Essential. Horizon Prep. Aug 28, 2018. http:// www.horizonprep.org/blog/encouraging-curiosity-and-why-its-essential. Accessed 9 June 2019. Vuran, M.C., and I.F. Akyildiz. 2008. “Cross-Layer Packet Size Optimization for Wireless Terrestrial, Underwater, and Underground Sensor Networks.” In The 27th Conference on Computer Communications (INFOCOM). Phoenix, AZ: IEEE. Wang, A., S. Cho, C. Sodini, and A. Chandrakasan. 2001. “Energy Efficient Modulation and MAC for Asymmetric RF Microsensor Systems.” In The International Symposium on Low Power Electronics and Design (ISLPED), 106–111. Huntington Beach, CA: ACM. Wang, C., K. Sohraby, Yueming Hu, Bo Li, and W. Tang. 2005. “Issues of Transport Control Protocols for Wireless Sensor Networks.” In International Conference on Communications, Circuits and Systems, 422–426. Honk Kong, China: IEEE.

References

39

Wang, Q., and I. Balasingham. 2010. “Wireless Sensor Networks—An Introduction.” In Wireless Sensor Networks: Application-Centric Design, ed. Y.K. Tan, 1–13. InTech. Warneke, B., and K.S.J. Pister. 2002. “MEMS for Distributed Wireless Sensor Networks.” In The 9th International Conference on Electronics, Circuits and Systems, 291–294. Dubrovnik: IEEE. Wener-Allen, G. et al. 2006. “Deploying a Wireless Sensor Network on an Active Volcano.” Internet Computing , Mar/April 2006: 18–25 (IEEE). Wu, Y.C., Q. Chaudhari, and E. Serpedin. 2011. “Clock Synchronization of Wireless Sensor Networks.” Signal Processing Magazine 28 (1): 124–138 (IEEE). Xiao, D., M. Wei, and Y. Zhou. 2006. “Secure-SPIN: Secure Sensor Protocol for Information via Negotiation for Wireless Sensor Networks.” In The First IEEE Conference on Industrial Electronics and Applications, 1–4. Singapore, Singapore: IEEE. Xu, B., F. Vafaee, and O. Wolfson. 2009. “In-Network Query Processing in Mobile P2P Databases.” In The 17th International Conference on Advances in Geographic Information Systems (GIS), 207–216. Seattle, WA: ACM. Ye, W., J. Heidemann, and D. Estrin. 2002. “An Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In The 21st Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM)I, 1567–1576. New York: IEEE. Yick, J., B. Mukherjee, and D. Ghosal. 2005. “Analysis of a Prediction-Based Mobility Adaptive Tracking Algorithm.” In The 2nd International Conference on Broadband Networks (BROADNETS), 753–760. Boston: IEEE. Yick, Jennifer, Biswanath Mukherjee, and Dipak Ghosal. 2008. “Wireless Sensor Network Survey.” Computer Networks , 52: 2292–2330 (Elsevier). Younis, M., and K. Akkaya. 2008. “Strategies and Techniques for Node Placement in Wireless Sensor Networks: A Survey.” Ad Hoc Networks 6 (4): 621–655 (Elsevier). Zheng, T., S. Radhakrishnan, and V. Sarangan. 2005. “PMAC: An Adaptive Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In The 19th International Parallel and Distributed Processing Symposium. Denver, CO: IEEE. Zhou, Zhong, Zheng Peng, Jun-Hong Cui, and Zhijie Shi. 2011. “Scalable Localization with Mobility Prediction for Underwater Sensor Networks.” IEEE Transactions on Mobile Computing 10 (3): 335–348 (IEEE).

Chapter 2

Energy Harvesting in WSNs

Energy is life … Energy harvesting is technology.

2.1

Energy Constraints

The deployment of WSNs in a multiplicity of diverse real-world applications is effectively proving how they did turn from scientific vision to reality. As elaborated throughout this book and will be, a multitude of systems have already been deployed. Such systems have enabled the collection of spatio-temporal data at significant granularities and inaugurated new research trends, thus restructuring the way in which industry products may be tailored to applications and field experiments. At the same time, with the outset of new sensor deployments, there is a pressing need to maintain wireless sensor nodes over prolonged deployment periods. Low-effort maintenance and self-reconfiguration have in fact been the idealistic selling points of WSNs. Network maintenance involve a number of tasks, such as changing batteries, replacing faulty nodes, and collecting data from special-purpose storage or gateway nodes. The maintenance costs must not exceed user expectations and budget to make a WSN deployment feasible and acceptable. A WSN application has several overlapping design dimensions: specifically, sensing environment, sensor node processing power, communication and storage capabilities, cost and size of each node, type of power source, topology, protocols for data dissemination and communication, and management tools. Using battery-powered sensor nodes is dominant in almost all applications. Untethered sensor nodes used in these deployments facilitate mobility and deployment in hard-to-reach locations (Fahmy 2016). A major limitation of untethered nodes is finite battery capacity as nodes operate only as long as the battery lasts. Finite node lifetime implies restricted existence of the applications or additional cost and complexity of regularly replacing batteries. Nodes could possibly use large batteries for longer lifetimes, but with increased

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_2

41

42

2 Energy Harvesting in WSNs

size, weight, and cost; they may also opt to use low-power hardware like a low-power processor and radio, but with lesser computation ability and lower transmission ranges. Several solution techniques have been proposed to maximize the lifetime of battery-powered sensor nodes. Some of these include energy-aware MAC protocols, such as SMAC (Ye et al. 2002), BMAC (Polastre et al. 2004), and XMAC (Buettner et al. 2006). There are also protocols for power-aware storage and for routing and data dissemination (Heinzelman et al. 2000; Intanagonwiwat et al. 2003). Duty-cycling strategies are proposed in Dutta et al. (2005), adaptive sensing rate in (Liu 2006), tiered system architectures in Kulkarni et al. (2005), and redundant placement of nodes in Kumar et al. (2004). While all the above techniques optimize and adapt energy usage to maximize the lifetime of a sensor node, the lifetime remains bounded and finite. These techniques help prolonging the application lifetime and/or the time interval between battery replacements, but do not preclude energy-related inhibitions. As a rule of life, all goodies cannot be obtained concurrently; understandably, the performance parameters of a WSN cannot be optimized simultaneously. Hence, higher battery capacity implies increased cost, low duty-cycle entails decreased sensing reliability, higher transmission range involves higher power requirement, and lower transmission range denotes transmission paths with more number of hops resulting in energy usage at more number of nodes. Energy harvesting is an alternative technique that has been applied to address the problem of limited node lifetime. Energy harvesting refers to harnessing energy from the surrounding nature or other energy sources such as body heat, foot strike, finger strokes, and subsequently converting it to electrical energy; the harnessed electrical energy powers the sensor nodes. If the harvested energy source is large and periodically or continuously available, a sensor node can be powered almost perpetually. Further, based on the periodicity and magnitude of harvestable energy, system parameters of a node can be tuned to increase node and network performance. That is, since a node is energy-limited till the next harvesting opportunity (recharge cycle), it can optimize its energy usage to maximize performance during that interval. For instance, a node can increase its sampling frequency or its duty-cycle to increase sensing reliability or increase transmission power to decrease length of routing paths. As a result, energy harvesting techniques have the potential to address the tradeoff between performance parameters and lifetime of sensor nodes. The challenge lies in estimating the periodicity and magnitude of the harvestable source and deciding which parameters to tune while simultaneously avoiding premature energy depletion before the next recharge cycle.

2.2 Energy Harvesting Concepts and Components

2.2

43

Energy Harvesting Concepts and Components

As earlier said, energy harvesting or scavenging energy is sniffing and collecting energy then converting it from one form to the other. Regarding sensor nodes, energy from external sources can be harvested to power the nodes and consequently increase their lifetime and capability. Knowing the energy-usage profile of a node, energy harvesting techniques could meet partly or all of its energy needs. A widespread and popular technique of energy harvesting is converting solar energy to electrical energy. As elaborated later in this section, solar energy is uncontrollable, as the intensity of direct sunlight cannot be controlled, but it is a predictable energy source with daily and seasonal patterns. Other techniques of energy harvesting convert mechanical energy or wind energy to electrical energy. For example, mechanical stress applied to piezoelectric materials, or to a rotating arm connected to a generator, can produce electrical energy. Since the amount of energy used for conversion can be varied, such techniques can be viewed as controllable energy sources. This section detailed all aspects of energy harvesting, typically, architecture, theories, sources, techniques, and storage technologies. These aspects are further highlighted in the context of typical energy harvesting projects, as fully detailed in Part III of this book.

2.2.1

Energy Harvesting Architectures

Energy harvesting systems can be divided into two architectures, specifically, harvest-use where energy is harvested just in time for use, and harvest-store-use where energy is harvested whenever possible and stored for future use (Sudevalayam and Kulkarni 2011). As shown in Fig. 2.1, we have: • Harvest-use architecture. This harvesting system directly powers the sensor node; thus, for the node to be operational, the power output of the harvesting system has to be continuously above the minimum operating point. If sufficient energy is not available, the node will be disabled. Sudden variations in harvesting capacity close to the minimum power point will cause the sensor node to oscillate in ON and OFF states. A harvest-use system can be built to use mechanical energy sources like pushing keys or buttons, walking, pedaling, etc. For example, the push of a key or button can be used to deform a piezoelectric material1 (WhatIs.com 2012), thereby generating electrical energy to send a short wireless message (Paradiso and

1

Piezoelectricity, also called the piezoelectric effect, is the ability of certain materials to generate an alternating current (AC) voltage when subjected to mechanical stress or vibration, or to vibrate when subjected to an AC voltage, or both. The most common piezoelectric material is quartz.

44

2 Energy Harvesting in WSNs Direct from source Harvesting system

Single/Double stage storage Harvesting system

Primary storage

Sensor node

(a) Harvest-use

Sensor node

Secondary storage

(b) Harvest-store-use

Fig. 2.1 Energy harvesting architectures (Sudevalayam and Kulkarni 2011)

Feldmeier 2001). Similarly, piezoelectric materials purposefully placed within a shoe may deform to different extents while walking and running (Paradiso and Starner 2005). The harvested energy can be used to transmit RFID signals and used to track the shoe-wearer (Shenck and Paradiso 2001). • Harvest-store-use architecture. It consists of a storage component that stores harvested energy and also powers the sensor node. Energy storage is useful when the available harvested energy is more than its current usage. Energy can be stockpiled in storage until enough collected for system operation; it is used later when either harvesting opportunity does not exist or energy usage of the sensor node has to be increased to improve capability and performance parameters. As an example, a harvest-store-use system can use uncontrolled but predictable energy sources like solar energy (Taneja et al. 2008). During daytime, energy is used for WSN operation and also stored for later usage. During night, the stored energy is conservatively consumed to power the WSN.

2.2.2

Power and Energy Differentiated

There is a customary confusion between energy and power to the extent of considering them alike, just synonyms. Certainly, energy and power are interrelated; it is not that tiring yet to distinguish between them. While energy is the ability to do work, power is its measurement that calculates the time by which the energy has been used; hence, energy is what one delivers and power is the rate at which it is delivered. As energy is the capability to do something, it is used for moving the car, or heating the home, or lighting the night, or even flying an airplane. Energy appears

2.2 Energy Harvesting Concepts and Components

45

in many forms and is often expressed in multiple units; specifically, Joule, Watt-hour (Wh), BTU2 (Rouse 2005), Newton meters, and Calories. There are different forms of energy; these include kinetic, potential, electric, thermal, gravitational, electromagnetic, sound, light, and elastic. For instance, one watt of electrical power maintained for one hour equals one Wh of energy, a thousand of these are a kWatt-hour (kWh). To be noted that a thousand watts for one hour, or one watt for a thousand hours, both equal one kWh. They are the same amount of energy (Lewis 2015). While energy measures the total quantity of work done, it does not specify how fast it is done. Power is the rate of energy per unit of time; it is the capacity of energy that is being used. Simply, power is defined as the rate of doing work; it finds its use in mechanical applications, heat applications, electrical applications, and several other areas. To typify, as energy is measured in joules, a ton of wood might have 18 billion (109) joules of energy stored in it (Whatisnuclear.com 2016). Power is measured in watts, which are just joules per second. So if one ton of wood were burned in a week, the furnace would be putting out a power of 18 GJ/week, which converts to 29,761.9 W (29.761 kJ/s). If burned in a month, and the furnace would be running at a power of 6944.4 W (6.9 kJ/s). In the end, no matter how quickly wood is burned, the used energy is 18 GJ. An illustration is also found in electricity. The standard unit of electrical power is watt, which is defined as a current of one Ampere, pushed by a voltage of one Volt. Basically, Volt
Ampere = Watt. If the standard wall socket delivers 220 V (120 V in the USA), for a plugged in light bulb with a current of 0.5 A flowing through it, the power used by the bulb is 220
(0.5) = 110 W (60 W in the USA). How much energy is the bulb using? It depends on how long it is left burning. A 60 W bulb burning for one hour will consume 60 Wh of energy. Six bulbs burning for ten hours would consume 6
60
10 = 3600 Wh, or 3.6 kWh. A thousand households would consume 3600 kWh (3.6 MWh). Another clarification that portrays power and energy is based on water towers. Water in the tower is energy and the flow of water out of the tower is power (Fig. 2.2). Energy can be stored; like water, it can also flow. When energy flows, it can do work such as moving stuff or lighting a house. The speed at which energy flows is power. The same amount of energy can be released at high power, which will occur quickly, or at low power which will take more time. The power and energy calculations can be recalled to be (The Engineering toolbox 2016):

2

The joule is the standard unit of energy in electronics and general scientific applications. One joule is defined as the amount of energy exerted when a force of one Newton is applied over a displacement of one meter. One joule is the equivalent of one watt of power radiated or dissipated for one second (W/s). In some applications, the British thermal unit (BTU) is used to express energy. One BTU is equivalent to approximately 1055 J.

46

2 Energy Harvesting in WSNs

(a)

Power = 6*60 W

Energy = 6*60*10 Wh

(b) Water tank Stored energy

Water pump

Water turbine

Legend: Power ∗ Time = Energy Power is the rate of using energy

Fig. 2.2 Power and energy

• Hydropower. The theoretically power available from falling water can be expressed as: W ¼q
q
g
h where, W is the energy (J). q is the density (kg/m3) (*1000 kg/m3 for water). q is the water flow (m3/s).

ð2:1Þ

2.2 Energy Harvesting Concepts and Components

47

g is the acceleration of gravity (9.81 m/s2). h is the falling height, head (m). The theoretically power available from a flow of 1 m3/s water falling 100 m can be calculated as: 


 P ¼ 1000 kg/m3
1 m3 =s
9:81 m/s2
ð100 mÞ ¼ 981; 000 W ¼ 981 kW • Energy from hydropower. The potential theoretical energy in a volume of elevated water (tank) is: W ¼q
V
g
h

ð2:2Þ

where W is the energy (J). q is the density (kg/m3) (*1000 kg/m3 for water). V is the volume of water (m3). g is the acceleration of gravity (9.81 m/s2). h is the falling height, head (m). Thus, for or a water volume of 10 m3 and elevated 10 m above a turbine, the potential energy in the water volume can be calculated as:


 W ¼ 1000 kg/m3
10 m3
9:81 m/s2
ð10 mÞ ¼ 981;000 J ¼ 981 kJ ðkWsÞ ¼ 0:27 kWh • Hydraulic pump power. The ideal hydraulic power to drive a pump depends on the mass flow rate, the liquid density, and the differential height. It can be calculated in kW as follows:
 Ph ¼ q
q
g
h= 3:6  106

ð2:3Þ

where Ph is the hydraulic power (kW). q is the water flow (m3/s). q is the density (kg/m3) (*1000 kg/m3 for water). g is the acceleration of gravity (9.81 m/s2). h is the falling height, head (m). Various devices can be used to convert one form of energy into another, while power cannot be converted or transformed. For instance, a battery converts chemical to electric energy, chemical explosion converts chemical energy to kinetic

48

2 Energy Harvesting in WSNs

and thermal energy, and so on. As previously stated, another difference between power and energy is that energy can be stored whereas power cannot. While energy comes with a time component, power is an instantaneous quantity; power cannot vary but remains constant, meanwhile energy accumulates predictably.

2.2.3

Energy Harvesting Versus Battery-Operated Systems

Wireless and embedded systems are usually powered using batteries. For applications where the system is expected to function for extended durations, energy becomes a strict bottleneck. Research efforts have been generously spent on the efficient use of battery energy. More recently, harvesting energy from the environment emerged as an alternative that has been explored to supplement or even replace batteries. An energy harvesting node is a system, which draws portion or all of its energy from the environment. Different from energy stored in the battery, a harvested energy is potentially infinite, though there may be a limit on the rate at which it can be used. For instance, a desk calculator using a solar cell is an example of a harvesting node. A network of harvesting nodes will be denoted as a harvesting network. Each node in such a network uses similar or dissimilar harvesting technologies, and some nodes may not be capable of harvesting energy at all. Several considerations characterize an energy harvesting source, when compared to using a battery. Specifically: • Rather than having a limit on the maximum energy, an energy harvesting source has a limit on the maximum rate at which the energy can be used. • The harvested energy availability typically varies with time in a non-deterministic manner, while a deterministic metric, such as residual battery, suffices to characterize the energy availability. • In networked systems with multiple harvesting nodes, different nodes may have different harvesting opportunities. In a distributed application, it is thus important to adjust the workload allocation with the energy availability at the harvesting nodes. In a battery-powered device, the specific power management design goals are to minimize the energy consumption (Sinha and Chandrakasan 2001) or to maximize the lifetime achieved (Shah and Rabaey 2002; Younis et al. 2002), while meeting compulsory performance constraints. In an energy harvesting node, one mode of usage is to treat the harvested energy as an addendum to the battery energy with a conceivable power management objective to maximize the lifetime. However, another usage mode is possible using the harvested energy at an appropriate rate such that the system continues to operate unfailingly. This mode is called

2.2 Energy Harvesting Concepts and Components

49

energy-neutral operation; a harvesting node is said to achieve energy-neutral operation if a desired performance level can be supported endlessly in case of no hardware failure (Kansal et al. 2007). For energy-neutral operation, the power management design considerations are widely unlike those of maximizing lifetime. Two design concerns are apparent: • Energy-neutral operation. It focuses on how to operate such that the energy used is always less than the energy harvested. The system may have multiple distributed components each harvesting its own energy. Thus, the performance not only depends on the spatio-temporal profile of the available energy, but also on how this energy is used to deliver network-wide performance guarantees. • Maximum performance. While ensuring energy-neutral operation, there should be a maximum performance level that can be supported in a given harvesting environment. This depends on the harvested energy at multiple distributed components. A simplistic approach may be to develop a harvesting technology whose minimum energy output at any instant suffices to supply the maximum power required by the load. This, however, has several disadvantages, such as high cost, and the infeasibility in many situations. As such, when harvesting solar energy, the minimum energy output for any solar cell would be zero at night, which cannot provide the power required by the load. A better approach is to add a power management system, between the harvesting source and the load, to satisfy the energy consumption profile from the available generation profile (Kansal et al. 2007). Three main blocks are to be discerned (Fig. 2.3).

Energy (Wh)

Harvesting system

Energy (Wh)

Harvesting sources

Fig. 2.3 Harvesting energy from the environment

Power (W)

Load

50

2 Energy Harvesting in WSNs

• Harvesting source. It is any available harvesting technology, such as a solar cell, a wind turbine, a piezoelectric harvester, or other transducer, which extracts energy from the environment. The energy output varies with time, depending on environmental conditions, which are typically outside the control of the designer. Figure 2.3 shows two possible harvesting sources with variable power output with time, a solar cell and a wind turbine. In a distributed system, multiple such harvesting sources may be present at multiple nodes at different locations. • Load. This refers to the energy-consuming activity being supported. A load, such as a sensor node, may consist of multiple subsystems such that energy consumption may be variable for its different modes of operation. For instance, the activity may involve sampling a sensor, transmitting the sensed value, and receiving an acknowledgment. In a harvesting network, the load may be an application layer activity, which requires the expenditure of energy at multiple nodes in the system, such as routing a data packet from one location to another. • Harvesting system. It is a system designed purposely to support a variable energy demand for a load from an energy harvesting source. Such variable energy is supplied when the instantaneous power supply levels from the harvesting source are not exactly matching the consumption levels of the load. In a harvesting network, this may also comprise collaboration among the power management systems of the constituent nodes to supply the available energy to distributed loads. Two approaches may somehow fulfill the load requirements from a variable supply: • Using an intermediate energy buffer in the harvesting system, such as a battery or an ultra-capacitor3 (NESSCAP 2014). • Based on load availability, the load consumption profile may be modified. In practice, neither of these approaches alone may be sufficient, since the load cannot be arbitrarily modified and energy storage technologies have non-ideal behavior that causes energy loss.

2.2.4

Storage Technologies

Storage technology plays an important role in energy harvesting systems; consequently, the choice of the storage component while perceiving its recharge

3

Ultra-capacitor, also known as super-capacitor, super-condenser, electric double-layer capacitor (EDLC) or pseudo-capacitor, is an electric capacitor that has an unusually high energy density when compared to common capacitors, typically on the order of thousands of times greater than a high-capacity electrolytic capacitor. Compared to batteries, ultra-capacitors are capable of more than ten times the power and more than thousand times the cycle life.

2.2 Energy Harvesting Concepts and Components

51

technology is of crucial significance. Rechargeable batteries and super-capacitors (Burke 2000) are the two main storage technologies of use as stressed out in the following sections. 2.2.4.1

Batteries

Rechargeable batteries, as a common choice of energy storage, can be built upon any one of several technologies (chemical compositions). A rechargeable battery is a storage cell that can be charged by reversing the internal chemical reaction. The widely popular rechargeable technologies are sealed lead acid (SLA), nickel cadmium (NiCd), nickel metal hydrid (NiMH) and Li-ion (Li-ion). These battery technologies are characterized along several directions; specifically, energy density4 (Battery University 2015), power density5 (Battery University 2015), battery cycle life6 (Woodbank Communications 2005), self-discharge rate7 (Woodbank Communications 2005), charge–discharge (roundtrip) efficiency8 (U.S. Department of Energy 2010), and number of deep recharge cycles9 (Woodbank Communications 2005). Table 2.1 shows characteristic values of the aforementioned parameters across the different rechargeable battery technologies that are revealed with typical models (calculations are revised according to the batteries datasheets). Several facts are clearly drawn: • Li-ion batteries have high output voltage, highest weight energy density and volume energy density, highest charge–discharge efficiency, and low self-discharge rate. • Though NiMH batteries have better weight energy density and volume energy density than NiCd batteries, NiCd batteries have higher number of deep recharge cycles. 4

Specific energy, or gravimetric energy density, defines battery capacity in weight (Wh/kg); energy density, or volumetric energy density, reflects volume (Wh/L). Consumer products requiring long run times at moderate load are optimized for high specific energy, also known as capacity measured in ampere-hours (Ah). 5 Specific power, or gravimetric power density, indicates loading capability. Batteries for power tools are made for high specific power; they come with moderate specific energy and very low internal resistance. 6 The number of complete charge–discharge cycles a battery can perform before its nominal capacity falls below 80% of its initial rated capacity. Key factors affecting cycle life are time t and the number N of charge–discharge cycles completed. 7 The self-discharge rate is a measure of how quickly a cell will lose its energy while sitting on the shelf due to unwanted chemical actions within the cell. The rate depends on the cell chemistry and the temperature. 8 The roundtrip efficiency is the ratio of total energy storage system output (discharge) divided by total energy input (charge) as measured at the interconnection point. 9 Cycle life decreases with increased depth of discharge (DOD); many cell chemistries will not tolerate deep discharge and cells may be permanently damaged if fully discharged.

Type Lead Acid NiCd NiMH Make Panasonic Sanyo Energizer KR-1100AAUb NH15-2500c Model No. LC-R061R3Pa Characteristics of a single storage element Nominal voltage (V) 6.0 1.2 1.2 Capacity 1300 mAh 1100 mAh 2500 mAh Energy (Wh) 7.8 1.32 3.0 Weight (g) 300 24 28.1 3 116.4 7.7 8.3 Volume (cm ) Weight energy density (Wh/Kg) 26 55 100 Volume energy density (Wh/L) 67 171.4 282 Self-discharge (per month) 3–20% 10% 30% Charge–discharge efficiency (%) 70–92 70–90 66 Memory effect No Yes No Charging method Trickle Trickle/ Pulse Trickle/ Pulse Recharge cycles 500–800 1500 1000 Legend a Panasonic LC-R061R3P (Panasonic 2005) b Sanyo KR-1100AAU (Sanyo 2008) c Energizer NH15-2500 (Energizer 2004) d Ultralife batteries UBP053048 (Ultralife Batteries 2006) e Ultralife batteries UBC433475 (Ultralife Batteries 2005) f Electric double layer capacitor: BOOSTCAP Ultra-capacitor (Maxwell Technologies 2013)

Table 2.1 Rechargeable battery technologies compared [based on (Taneja et al. 2008)] Li-polymer Ultralife UBC433475e 3.7 930 mAh 3.4 22 12.8 156 296 average loss before entering the Test state

Active neighbors < NT and • DL > LT, or • DL < LT and help message received

After Tp

Passive

Sleep After Ts

Fig. 4.7 ASCENT state transitions (Cerpa and Estrin 2004)

– When Tt expires, the node enters the active state. – If before Tt expires: The number of active neighbors is above the neighbor threshold, NT, or The average data loss rate, DL, is higher than the average loss rate before entering in the test state. Then, the new node moves into the passive state. • If multiple nodes make a transition to the test state, then the node ID is used in the announcement message as a tie-breaking mechanism (higher IDs win). The logic behind the test state is to probe the network to see if the addition of a new node may actually improve connectivity. • When a node enters the passive state, it sets up a timer Tp and sends new passive node announcement messages. This information is used by active nodes to make an estimate of the total density of nodes in the neighborhood; active nodes transmit this density estimate to any new passive node in the neighborhood. Depending on Tp, two transitions are possible: – When Tp expires, the node enters the sleep state. – If before Tp expires: The number of neighbors is below NT and Either DL is higher than the loss threshold, LT, or DL is below LT but the node received a help message from an active neighbor, and it makes a transition to the test state. While operating, ASCENT nodes are granted various eligibilities depending on their states: • While in passive state, nodes have their radio ON and are able to overhear all packets transmitted by their active neighbors. No routing or data packets are

4.1 Duty-Cycling Approach Taxonomy

• •

• • •

123

forwarded in this state since this is a listen-only state. The reasoning behind the passive state is to gather information regarding the state of the network without causing interference with other nodes. On the contrary, active neighbors forward data and routing (control) messages until they run out of energy. Active nodes can also send help messages when they find an unacceptable level of local data loss. As soon as a node joins the network, it starts monitoring the network conditions and also signals its presence as an active node through a neighbor announcement message. This process continues until the number of active nodes is such that the message loss rate experienced by the sink is below a predefined application-dependent threshold. Nodes in the passive state and test state continuously update the number of active neighbors and data loss rate values. Energy is consumed in the passive state since the radio is still ON when not receiving packets. A node that enters the sleep state turns the radio OFF, sets a timer Ts, and goes to sleep. When Ts expires, the node moves into passive state. Finally, a node in active state continues forwarding data and routing packets until it runs out of energy. If the data loss rate is greater than LT, the active node sends help messages.

Based on its design objectives and performance analysis, ASCENT is independent of the routing protocol. Moreover, it limits the packet loss due to collisions because the node density is explicitly taken into account as a parameter, i.e., in the form of a neighbor threshold value. Also, ASCENT has good scalability properties. On the other hand, energy saving does not increase proportionally with the node density because it mainly depends on passive–sleep cycle rather than on the number of active nodes.

Naps In Godfrey and Ratajczak (2004), Naps is advocated to be a different approach that models the network as a random graph, and exploits the percolation theory1 (Grimmett 1999) to characterize the network connectivity when a duty-cycle is enforced on nodes. Naps is a randomized topology management scheme that does not rely on geographic location information; it provides flexibility in the target density of waking nodes and sends only a periodic heartbeat message between waking neighbors. Naps is thus implementable even on modest hardware. In wireless ad hoc networks, nodes are distributed over a geographic region and communicate only with nodes that are within a limited radius; long-range communication is made possible by multihop forwarding messages toward their targets. Percolation is the process of a liquid slowly passing through a filter; it is how coffee is usually made. In statistical physics and mathematics, percolation theory describes the behavior of connected clusters in a random graph.

1

124

4 Energy Management Techniques for WSNs …

Knowingly, nodes are typically battery-powered, which makes energy a scarce resource. Furthermore, on most hardware, a substantial portion of energy is consumed by the radio, irrespective of whether it is sending, receiving, or merely listening for packets (Chen et al. 2002; Raghunathan et al. 2002; Schurgers et al. 2002). In high-density deployments, there is no need for all nodes to forward traffic so as to maintain connectivity in the network. In such networks, it is required to select a fraction of nodes to turn their radios or other components OFF (“nap”) for a calculated interval to have their energy consumption reduced, thus prolonging the useful lifetime of the entire system. Such a technique can also increase the capacity of the network (Gupta and Kumar 2000) by producing a “waking” lower-density connected backbone that assumes all forwarding responsibilities, to which all “napping” nodes communicate only infrequently or with reduced transmission power. With such considerations in mind when designing Naps, it was desired to select a nearly maximal set of napping nodes, under the constraint that nearly all waking nodes are in a single connected component, and nearly every napping node can transmit to a waking node. This is a significantly alleviated form of the NP-complete connected dominating set (CDS) problem (Garey and Johnson 1979), which requires selecting the smallest connected backbone such that all nodes are either in or adjacent to the backbone. These requirements might be sufficient since connectivity may be compromised anyway due to node and network unreliability or caused by an initially disconnected placement of nodes. Worth reminding, algorithms that turn OFF radios by exploiting redundancy have been classified as topology management schemes, precisely Span (section “Span”) and STEM (section “Sparse Topology and Energy Management (STEM)”), as well in Raghunathan et al. (2002) and Xu et al. (2003). Differentiating the particularities of the essential such algorithms, it is perceptible that: • GAF is similar to Naps in its goals and provides the application with an effectively static topology (section “Geographical Adaptive Fidelity (GAF)”). However, it relies on geographic information, limiting its use to hardware and environments where such information can be reliably obtained. Comparatively, it has been demonstrated that Naps provides good performance even without such features. • Differently, Span is designed to select a connected backbone for the purposes of energy conservation and does not assume geographic location information (section “Span”). Span produces a not necessarily minimal connected dominating set (CDS). However, it is comparatively complex and requires that every node periodically receives topology information from each of its neighbors in the original graph, so that the fraction of time that a node radio is idle or receiving, as opposed to sleeping, will increase as the initial density of the network increases. This feature limits Span improvement to just a factor of two in system lifetime even at high initial densities. Additionally, Span has no flexibility in the density of waking nodes, limiting its use for applications other than connectivity. • Topology management by priority ordering (TMPO) constructs and maintains a backbone topology based on a minimal dominating set (MDS) of the network (Bao and Garcia-Luna-Aceves 2003). According to this algorithm, each node

4.1 Duty-Cycling Approach Taxonomy

125

determines the membership in the MDS for itself and its one-hop neighbors based on two-hop neighbor information that is disseminated among neighboring nodes. The algorithm then ensures that the members of the MDS are connected into a connected dominating set (CDS), which can be used to form the backbone infrastructure of the communication network for such purposes as routing. TMPO shares many of the traits of span, though with greater emphasis on performance in a mobile setting. • STEM turns OFF nodes more aggressively than the previous schemes by exploiting the fact that many applications only generate traffic sporadically (section “Sparse Topology and Energy Management (STEM)”). STEM, a topology management technique, trades power savings for path setup latency in sensor networks. It emulates a paging channel by having a separate radio operating at a lower duty-cycle. Upon receiving a wakeup message, the node turns ON the primary radio, which takes care of the regular data transmissions. This wakeup message can take the form of a beacon packet in STEM-B or simply a tone in STEM-T. Topology management is specifically geared toward those scenarios where the network spends most of its time waiting for events to happen, without forwarding traffic. Similar to the adaptive fidelity energyconserving algorithm (AFECA) (Xu, Heidemann, and Estrin 2000), STEM trades energy conservation for increased routing latency. Naps as a key algorithm for energy management entails several worth spotlighting outcomes: • Naps helps “thinning” an ad hoc network by selecting nodes to nap and wake at calculated intervals to achieve an expected target density of waking nodes, without requiring that nodes know the initial density. Intuitively, the fraction of time that a node should be awake is inversely proportional to its degree and proportional to the target density. The algorithm persistently changes the subset of waking nodes to allow all nodes to periodically nap. Naps is remarkably simple as it does not require location information, nor clock synchronization, and only requires a node to wake and broadcast a heartbeat to its waking neighbors every time period T, a tunable parameter of the algorithm. • The waking subgraphs produced by Naps exhibit a phase transition. If the initial and target densities are above a critical threshold, then after running Naps on an infinite graph with the initial density, any particular node is in an infinite connected component of waking nodes with nonzero probability. This general phenomenon is called percolation and has been widely studied (Grimmett 1999). In the finite setting, the expected fraction of waking nodes in the largest connected component approaches the percolation probability as the area is increased (Penrose 2003). This is in contrast to the logarithmic density required to ensure complete connectivity (Gupta and Kumar 1999). • Empirically shown in Naps analysis, the waking subgraphs are almost completely connected even for relatively small values of the target density. Considering a network of at least 500 nodes with an initial density of 12/p and a

126

4 Energy Management Techniques for WSNs …

target density of 6/p, such that a node anticipates 6 waking and 12 total neighbors. It is expected that more than 98% of the nodes might be either waking in the largest connected component, or napping and adjacent to that component. • Naps can be used for any application that requires selecting a set of nodes with a specified target density and which can adapt to dynamic changes in the topology. Noteworthy, such adaptive algorithms have been developed for routing (Karp and Kung 2000; Rao et al. 2003); adaptivity is basically a requirement for most practical applications suffering from unreliability and mobility. Naps does not require napping nodes to receive messages allowing them to leave their radio in the sleep state; also, the number of messages received is proportional to the target density as opposed to the initial density. This is a significant improvement over some previous algorithms and is one of the primary reasons that Naps shows significant energy conservation even at arbitrarily high initial densities, especially when the energy consumption of a napping node is small. • In an obstructed environment, obstructions negatively impact the connectivity of both the underlying graph and the waking subgraphs. • Finally, while the analysis assumes that nodes are distributed according to an infinite Poisson point process, it was shown through simulation that the proposed Naps algorithm is applicable even when the network is finite, the underlying distribution is not uniform, and when nodes are mobile. In Naps, time is split into time periods with duration T. A node v operates on the following tactic: • Each node v initially waits a random amount of time tv uniformly distributed into the range [0, T) and thereafter operates in time periods of length T. • At the start of each period, node v broadcasts a HELLO message; it then listens to HELLO messages from other nodes. • If for a predetermined number threshold c of neighbors, fewer than c messages are received before the end of the period, node v remains awake during the entire period. Otherwise, node v naps from the time the cth message is received until the end of the period. Note that for t  T, the graphs occurring at times t, t + T, t + 2T, … are equivalent. Each node v adaptively estimates its local density using an estimate of its degree, deg(v); node v remains awake for an expected fraction c=ðdegðvÞ þ 1Þ of each period. Hence as the density increases, nodes will have more neighbors and thus may nap for longer periods, keeping the density of waking neighbors nearly constant. If every node has nearly the same degree, then c nodes are expected to remain awake among the deg(v) + 1 nodes in each node v radius. This yields an overall density of approximately c=p. Conclusively speaking, Naps is a decentralized topology management protocol based on a periodic sleep/wakeup scheme; it has favorable connectivity and is flexible and robust. Because Naps provides only probabilistic guarantees, more complex algorithms may be necessary for some applications. In practice, however,

4.1 Duty-Cycling Approach Taxonomy

127

node and communication unreliability might force applications to cope with probabilistic guarantees, so Naps may provide the simple and viable alternative.

Uncoordinated Power Saving Mechanisms with Latency Considerations A specific consideration of WSNs, made of a large number of nodes with limited processing and communication capabilities, and destined for time-critical monitoring applications, is presented in (Dousse, Mannersalo, and Thiran 2004). Each node performs some sensing of a particular confined area and sends the result to a data-collecting node (sink) in a multihop fashion, using other nodes as relays. These WSNs have to provide both a good coverage of the area to be monitored and an acceptable connectivity of the network. Connectivity can be somewhat relaxed to the less restrictive requirement that the sink be multihop connected to a set of nodes that span the entire monitored domain. Referring to general-purpose wireless ad hoc networks, connectivity is the requirement that the network be fully connected. Factually, the widely variable quality of the wireless channel and the limited battery lifetime of a sensor make it more expensive to ensure full connectivity of the network than to ensure that only the sink is connected to sensors well scattered over the entire monitored area. The question of having one node (the sink) connected to a large number of sensors well dispersed throughout the domain is central to percolation theory (See Footnote 1) (Grimmett 1999). The percolation probability is the probability that an arbitrary node, in particular the sink node, belongs to a cluster of infinite size. The main result of percolation theory is that there exists a finite positive value of the connectivity range, or equivalently of the node spatial density, under which the percolation probability is zero (subcritical phase) and above which it is nonzero (supercritical phase). If connectivity can be made less restrictive in WSNs than in many other ad hoc networks, energy consumption is often a much more critical variable, because of the limited battery that can be put in a sensor, as well as due to the cost of replacing a node that has failed in a harsh environment. Energy is consumed by sensors in their sensing, processing, and communicating tasks. Sensing has to be done at a periodicity dictated by the monitored event; the energy it consumes can be reduced thru increasing the number of sensors to make the area covered by a single sensor small. Processing and communication energy consumption depend on the hardware and on how data are aggregated and the medium accessed; this consumption offers the largest potential for reduction. Actually, nodes spend a considerable amount of energy when listening to their neighbors, and as long as none of these neighbors transmits any data, such energy is wasted. Data collection and medium access control (MAC) schemes need therefore to incorporate energy saving as a primary goal. In particular, most devised energy-saving MAC schemes for WSNs introduce a sleeping mode for nodes, during which no energy is spent (Sinha and Chandrakasan 2001; Ye et al. 2002).

128

4 Energy Management Techniques for WSNs …

For a node to go on sleep mode, two approaches are to be weighted against each other: • Using a time-division multiple access (TDMA) scheme. However, this requires nodes to synchronize with each other quite tightly, which can be a quite complex task in large networks with random node locations and imperfect (drifting) clocks. • Letting the nodes set their wakeup and sleeping times in a decentralized fashion. Such scheme reduces complexity; nonetheless, it increases the delay (latency) to transfer information between the sink and a distant node. Pushing the decentralization to an extreme where nodes go to sleep independently from each other eliminates the complexity of having synchronized clusters of nodes, but at the same time raises concerns about an increase of the network latency (Dousse, Mannersalo, and Thiran 2004). More importantly, it will not only increase the average latency itself, but there will also be an increase in the variance of this latency. For some applications, such as spatial data collection for statistical purposes, this is certainly acceptable, but it is not suitable for many time-critical applications. In a typical time-critical scenario, a WSN monitors an area and sends an alarm when an abnormal event is sensed, such as an intrusion and a rapidly changing variable. Even if some fixed amount of latency can be tolerated, a highly variable latency due to the random position of the nodes, the random radio range, the non-synchronized, or even random sleeping and active periods is much more problematic. Nevertheless, it is possible to let nodes go into sleep, without any coordination between their schedules, and yet have rigorous bounds on the latency. The protocol presented in this section provides a positive solution to this concern and presents analytical bounds on the latency of a sensor with random independent and identically distributed active and sleeping periods. Dynamic percolation theory used to obtain these bounds is based on the following assumptions: • Nodes are randomly scattered on the plane according to a homogeneous Poisson process. • Nodes can send data in one hop to a neighbor within some prescribed, deterministic, connectivity range. The assumption of circular direct connectivity area is debatable though (Ganesan et al. 2002). In real life, the connectivity range is far from being a circle in many circumstances; it varies widely depending on many factors, such as interferences with other nodes, background noise, time-varying channel, hardware defects, and directional antennas. A more realistic model is to set a threshold on the signal-to-noise and interference ratio at the receiver, as in Ganesan et al. (2002) which defines a direct connectivity area around a node by contour plots, whose shapes are highly variable. Interestingly, percolation does stand for models taking interferences into account (Dousse et al. 2005), as long as they hold the Poisson Boolean model. Moreover, when nodes have long sleeping periods, interferences from other nodes become less critical. Also, regarding the high level of directionality in the reception sets, which is not captured by an

4.1 Duty-Cycling Approach Taxonomy

129

isotropic model like the static Boolean model, anisotropy2 (Dictionary.com 2017) does help to make the network percolate and hence increases connectivity. The only aspect that is not included in the static Boolean model is the time variability of the wireless channel; it can, however, be included in a dynamic Boolean model, like the blinking Poisson Boolean model (Dousse et al. 2004). • Nodes switch between an active mode, ON state, and a sleeping mode, OFF state, independently from each other. As long as the nodes do not sense any event or do not receive any data from their neighbors, the durations of the active and sleeping periods are independent and identically distributed random variables with OFF-periods being constant in a periodic ON/OFF schedule or exponentially distributed according to a memoryless sleeping schedule. ON periods are identically distributed but not necessarily exponentially. • When a node senses an event, or when it receives data (message) from one of its neighbors, it stays active and broadcasts this message to all its neighbors within direct reach, until it is certain that the message has reached, with high probability, all its neighbors, or the sink. Assuming that nodes broadcast any data they sense or receive to all their neighbors makes the routing simple and enables to compute analytically the bounds, which is also convenient if incoming events rarely occur as in an intrusion detection scenario. • The sensing range is smaller than or equal to the connectivity range. The sensing range is randomly distributed between a minimal nonzero value and a maximum equal to the connectivity range. This means that there will be supercritical amount of devices for message transmissions. In order to save batteries, only a fraction of sensors listens to the transmission channel. In this protocol, several functional remarks are obvious: • It is not assumed that there is a path from any sensor to the sink. On the contrary, even if there is no connectivity of all sensors at all times because many of them are sleeping and only a few are active, it is possible to transfer data from any sensor to the sink in a bounded time with probability one, without any coordination between the sensors. • The blinking Poisson Boolean model suits sensor networks where nodes switch between a sleeping and an active phase. Even though their switching ON/OFF schedules are not coordinated, their positions are random, and the durations TON and TOFF are such that the number of active nodes at any particular time is considerably low such that the network is always disconnected. • It has been proved that any message (alarm) generated by a sensor will reach the sink in a time proportional to the distance between the sensor and the sink. The rate of this linear growth does not depend on the random locations of the nodes, but only on the parameters k (node density), r (connectivity range), TON and TOFF (the average active and sleeping durations).

2

In physics, it means of unequal physical properties along different axes. In Botany, it means of different dimensions along different axes.

130

4 Energy Management Techniques for WSNs …

• Only considered is the subcritical phase where the network is disconnected. • Similar to Naps (section “Naps”), this protocol focuses on time-critical monitoring applications. • In contrast to Naps, the proposed protocol requires the knowledge of the network density. To overcome this problem, DDEMA extends the approach presented in this section thru considering only neighboring information (section “Degree-Dependent Energy Management Algorithm (DDEMA)”).

Degree-Dependent Energy Management Algorithm (DDEMA) With prolonging the lifetime of WSNs and reducing overhead and redundancy, as targets, it is often desirable to keep some sensors inactive for some of the time (Shih et al. 2002; Ye et al. 2002; Lu et al. 2005) while maintaining a good level of network connectivity at all times. Mostly, energy management algorithms aim to safeguard full connectivity, k-connectivity, to ensure that any pair of sensors in the network is connected thru k paths. However, in large-scale WSNs exposed to severe natural hazards, enemy attack, and resource depletion, the full connectivity criterion may be overly restrictive and difficult to achieve. Moreover, the complexity and overhead involved in sleep scheduling algorithms that are required for maintaining full connectivity make their use in large-scale WSNs impractical. DDEMA views the connectivity of WSNs from a different perspective. Specifically, a simple measure of network functionality is the fraction of nodes in the largest connected component of the network; nodes in that component can communicate with an extensive portion of the network, while those in smaller components can communicate only with at most a few other nodes. For instance, the network is functional if the WSN is able to collect information from almost the entire coverage area even after a substantial number of sensor failures. On the other hand, if after many sensor failures, the WSN breaks down into isolated parts where even the largest component can only reach a few sensors, then the network is not considered to be functional. From this perspective, the characterization of network connectivity corresponds to the study of qualitative and quantitative properties of the largest component. A powerful technique for this study comes from the mathematical theory of percolation (Meester and Roy 1996; Grimmett 1999; Penrose 2003), which is a useful tool for the analysis of large-scale WSNs as shown in section “Uncoordinated Power Saving Mechanisms with Latency Considerations” and published in Gupta and Kumar (1999), Dousse et al. (2005), and Dousse et al. (2006). A percolation process resides in a random graph structure, where nodes or links are randomly designated as either “occupied” or “unoccupied.” When the graph structure resides in continuous space, the resulting model is described by continuum percolation (Meester and Roy 1996). A major focus of continuum percolation theory is the random geometric graph induced by a Poisson point process with constant density k. A fundamental result of continuum percolation regards a phase

4.1 Duty-Cycling Approach Taxonomy

131

transition effect, whereby the macroscopic behavior of the system is very different for densities below and above some critical value kc: • For k < kc (subcritical), the connected component containing the origin, or any other node, contains a finite number of points almost surely. • For k > kc (supercritical), the connected component containing the origin, or any other node, contains an infinite number of points with a positive probability. From a percolation-based connectivity perspective, energy management algorithm design can be modeled thru dynamic site percolation processes on random geometric graphs. An energy-saving mechanism based on this approach is presented in section “Uncoordinated Power Saving Mechanisms with Latency Considerations”; each sensor schedules its own active-sleeping process, independently of other nodes, and keeps its active time ratio strictly larger than kc =k. It was shown that the sensor network operating under this mechanism is percolated, in the supercritical phase, at all times. While the algorithm exhibited in section “Uncoordinated Power Saving Mechanisms with Latency Considerations” is simple and appealing, it has a serious shortcoming; each sensor is required to know the density or mean degree of the whole network. Such global information may be unavailable in a large-scale WSN, especially when changes of topology are expected. On the other hand, the node degree (the number of neighbors) of each sensor provides useful local information for designing more efficient distributed energy management algorithms. Typically, it is noted that a sensor with a higher degree needs to transmit to more neighbors and process more data, and thus may deplete its energy more quickly than nodes with a smaller degree. Moreover, a sensor with a large degree implies that its communication/sensing range is being densely covered by many sensors. Viewing these reasons, a sensor with large degree should be active for a smaller proportion of time in order to stay energy-efficient while achieving the overall sensing and communication task. DDEMA is a fully distributed energy management algorithm for large-scale WSNs; it allows each sensor to schedule its own activity based on its node degree, without knowledge of global network parameters. This mechanism is modeled using a degree-dependent dynamic site percolation process on random geometric graphs; a specification is reached for the conditions under which the resulting network is guaranteed to be percolated at all times. Useful findings are worth spotlighting: • Phase transitions in degree-dependent site percolation processes are specified. • Due to dynamic ON/OFF behaviors of the sensors, a delay is incurred for any transmission even when the propagation delay is ignored. The characteristics of this transmission delay for the proposed DDEMA are studied by modeling the problem as a degree-dependent first passage percolation process on random geometric graphs.

4 Energy Management Techniques for WSNs …

132

• Employing the subadditive ergodic theorem (Liggett 1985), it is shown that with negligible propagation delay: – The message delay scales linearly with Euclidean distance between the sender and the receiver when the resulting network is in the subcritical phase. – The delay scales sublinearly with the distance if the resulting network is in the supercritical phase. • When the propagation delay is on the same, or larger, order of the inactive periods, the message delay cannot be improved too much by transforming the network from the subcritical phase to the supercritical phase.

4.1.1.3

Appraisal of Topology Control Protocols

Location-driven topology control protocols require that sensor nodes can somewhat know their positions. Two schemes might achieve this objective: • Providing sensors with a GPS unit helps them to know their locations; but, since GPS is remarkably expensive and energy consuming, it is mostly impossible to install it on all nodes. In such a case, it would be tolerable to equip only a limited subset of nodes with a GPS and then derive the location of the others using other techniques (Langendoen and Reijers 2003). • Exploiting radio or sound waves might provide an alternative to recurring to GPS (Mao et al. 2007). However, the lack of the hardware suitable to acquire location information in commonly available sensor platforms makes connectivity-driven protocols generally preferable (Sect. 4.1.1.2), as they only require information attainable from local measurements. Noticeably, the energy efficiency of topology control protocols is tightly related to node density; moreover, the achievable gain in terms of network lifetime depends on the actual density. Topology control protocols can typically increase the network lifetime by a factor of 2–3 with respect to a network with nodes always ON (Mainwaring et al. 2002; Ganesan et al. 2004; Warrier et al. 2007). This factor may be not suitable for many practical applications. Consequently, topology control protocols should be coupled with other energy conservation techniques, such as those presented in this chapter and the following two chapters. However, the simultaneous use of multiple energy conservation schemes may lead to unforeseen consequences. In fact, although the combination of protocols should be transparent to the applications, the obtained results may though be very different from what one would expect (Melodia et al. 2005; Malesci and Madden 2006).

4.1 Duty-Cycling Approach Taxonomy

4.1.2

133

Power Management Protocols

As previously presented in Sect. 4.1, duty-cycling operated on active nodes is referred to as power management. Topology control and power management are complementary techniques that implement duty-cycling with a different granularity. Power management techniques can be further subdivided into two broad categories depending on the layer of the network architecture they are implemented at. Recalling Fig. 4.1, power management protocols can be implemented either as independent sleep/wakeup protocols running on top of a MAC protocol, typically at the network or application layer, or strictly integrated with the MAC protocol itself. Precisely defining (Anastasi et al. 2009): • Independent sleep/wakeup protocols permit a greater flexibility as they can be tailored to the application needs and, in principle, can be used with any MAC protocol without relying on topology or connectivity aspects. Independent sleep/ wakeup protocols can be further subdivided into three main categories: typically, on-demand, scheduled rendezvous, and asynchronous protocols. • MAC protocols with low duty-cycles permit to optimize medium access functions based on the specific sleep/wakeup pattern used for power management; they can be subclassified as TDMA-based, contention-based, and hybrid protocols. The subcategorization of power management protocols as independent sleep/ wakeup protocols and MAC protocols with low duty-cycle is illustrated in Fig. 4.8 and further detailed in Sect. 4.1.2.1 and Sect. 4.1.2.2, respectively.

Power management

Sleep/wakeup protocols

On-demand

Scheduled rendezvous

MAC with low duty-cycles

Asynchronous

TDMA

Contention-based

Fig. 4.8 Power management approach taxonomy (Anastasi et al. 2009)

Hybrid

4 Energy Management Techniques for WSNs …

134

4.1.2.1

Sleep/Wakeup Protocols

As previously discussed, sleep/wakeup schemes can be defined for a given component, i.e., the radio subsystem of the sensor node, without relying on topology or connectivity aspects. This section surveys the main sleep/wakeup schemes implemented as independent protocols on top of the MAC protocol, i.e., at the network or the application layer. Independent sleep/wakeup protocols can be further subdivided into three main categories (Anastasi et al. 2009): precisely, on-demand, scheduled rendezvous, and asynchronous schemes (Fig. 4.8). Before the detailed survey in sections “On-Demand Schemes,” “Scheduled Rendezvous Schemes,” and “Asynchronous Schemes,” each category deserves a brief description: • On-demand protocols take the most intuitive approach to power management. It is assumed that nodes can be signaled and awakened at any point of time and then a message is sent to the node. This is usually implemented by employing two wireless interfaces. The first radio is used for data communication and is triggered by the second ultra-low-power, or possibly passive, radio, which is used only for paging and signaling. STEM (section “Sparse Topology and Energy Management (STEM)”) and its variation (Miller and Vaidya 2004), and passive radio-triggered solutions (Gu and Stankovic 2005) are examples of this class of wakeup methods. Although these methods can be optimal in terms of both delay and energy, they are not yet practical. The main problem associated with on-demand schemes is how to inform the sleeping node that some other node is willing to communicate with it. Clearly, such schemes typically use multiple radios with different energy/performance tradeoffs, i.e., a low-rate and low-power radio for signaling, and a high-rate but more power-hungry radio for data communication. Downside issues prohibit the widespread use and design of such wakeup techniques: specifically, the limited available hardware options resulting in limited range and poor reliability, and the stringent system requirements. Consequently, there is a need for efficient scheduled wakeup schemes which are reliable and cost effective and can also guarantee the delay and lifetime constraints. • An alternative solution uses a scheduled rendezvous approach. The basic idea behind scheduled rendezvous schemes is that each node should wake up at the same time as its neighbors. Typically, nodes wake up according to a wakeup schedule and remain active for a short time interval to communicate with their neighbors. Then, they go to sleep until the next rendezvous time. In this category, the nodes follow deterministic, or possibly random, wakeup patterns (Ye et al. 2002b; Zheng et al. 2003; Van Dam and Langendoen 2003; Ye et al. 2004; Lu et al. 2005). Time synchronization among the nodes in the network is generally assumed. • An asynchronous sleep/wakeup protocol can be used. With asynchronous protocols, a node can wake up when it wants and still be able to communicate with its neighbors. This goal is achieved by properties implied in the sleep/wakeup scheme; thus, no explicit information exchange is needed among nodes (Zheng et al. 2003; Polastre et al. 2004). Although asynchronous methods are simpler to

4.1 Duty-Cycling Approach Taxonomy

135

implement, they are not as efficient as synchronous schemes, and in the worst-case their guaranteed delay can be very long.

On-Demand Schemes On-demand schemes are based on the idea that a node should be awakened just when it receives a packet from a neighboring node. This minimizes the energy consumption and, thus, makes on-demand schemes particularly suitable for sensor network applications with a very low duty-cycle, e.g., fire detection, surveillance of machine failures, and, more generally, all event-driven scenarios. In such scenarios, sensor nodes are in the monitoring state; i.e., they only sense the environment most of the time. As soon as an event is detected, nodes transit to the transfer state. On-demand sleep/wakeup schemes are aimed at reducing energy consumption in the monitoring state while ensuring a limited latency for transitioning in the transfer state. The implementation of such schemes typically requires two different channels, a data channel for normal data communication and a wakeup channel for awaking nodes when needed. Although it would be possible to use a single radio with two different channels, published proposals rely on two different radios. This allows immediate transmission of signal on the wakeup channel if a packet transmission is in progress on the other channel, thus reducing the wakeup latency. Typical on-demand schemes are the focus of sections “Sparse Topology and Energy Management (STEM)” and “Pipelined Tone Wakeup (PTW).” Sparse Topology and Energy Management (STEM) STEM is a topology management scheme designed to reduce energy consumption in the monitoring state to a bare minimum while ensuring satisfactory latency for transitioning to the transfer state (Schurgers et al. 2002). In fact, STEM allows to trade efficiently one design constraint (energy) for the other (latency). It may also be combined with techniques that leverage increased network density to obtain energy savings. In essence, STEM offers the designers full flexibility in trading latency, density, and energy versus each other. A mathematical model is also developed to govern these tradeoffs. For example, for a specific desired network lifetime and acceptable notification latency, the required network density can be calculated. When STEM is running in the network, this density will assure that both the latency and the network lifetime are as preferred. This mathematical model is therefore a tool for the network designer to choose the optimal parameter settings given the deployment requirements. At design time, the model allows the selection of the desired operating point in the latency-density-lifetime design space. In sensor networks, routing is directed toward two alternative approaches, namely flat multihop and clustering (Al-Karaki and Kamal 2004; Pantazis and Vergados 2007; Raghunandan and Lakshmi 2011; Liu 2012). Although STEM is

136

4 Energy Management Techniques for WSNs …

applicable to both of them, its main focus is on flat multihop routing. For clustered approaches, which are possibly hierarchical, STEM can be used to reduce the energy of the cluster heads. Topology management techniques, such as SPAN (section “Span”) and GAF (section “Geographical Adaptive Fidelity (GAF)”), have been proposed for flat multihop routing. They trade network density for energy savings while preserving the data forwarding capacity of the network; however, the absence of traffic in the monitoring state was not exploited. By integrating these schemes into STEM, it was possible to combine their benefits with those of STEM to achieve compounded energy savings. STEM is essentially a technique to quickly transition to the transfer state, while making the monitoring state as energy-efficient as possible. Other work has proposed to do this wakeup using a separate paging channel, by critically assuming that the listen mode of this paging radio is ultra-low power (Bachir et al. 2010). However, the difference in the transmission range between the data and the wakeup radio presents a major difficulty. STEM offers an alternative by trading energy for latency. Furthermore, if such a low-power radio is available, the energy savings are further improved by using it in a low duty-cycle. The work in McGlynn and Borbash (2001) describes an algorithm that also uses a low duty-cycle radio; it is designed for a different goal, namely, to discover the network topology some time after its deployment. It is less aggressive than STEM and, therefore, would result in much higher latencies to transition to the transfer state. The same principle of duty-cycling the radio is also adapted in the medium access control (MAC) protocol, called S-MAC, presented in Ye et al. (2002) and shown in section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC).” However, channel access and node wakeup are integrated together. In the monitoring state, where there are no data to forward, STEM is therefore more energy-efficient than S-MAC while assuring timely transitioning to the transfer state. In this transfer state, STEM allows for any MAC protocol, including S-MAC. The basic concept of STEM functioning involves three phases: • In the monitoring state, where there is no traffic to forward, only the node sensors and some preprocessing circuitry are ON. In this state, the node is asleep, i.e., in the sleep state. When a possible event is detected, the main processor is awakened to analyze the data in more detail. The radio is only turned ON if the processor decides that the information needs to be communicated to other nodes. The underlying logic can be understood from the radio mode power values listed in Table 4.2; these are for the TR1000 radio (Murata Electronics 2015b), where the transmit range is set to approximately 20 m (Raghunathan et al. 2002). This low-power radio has a data rate of 2.4 Kbps and uses ON-OFF keying (OOK) (Proakis and Salehi 2007) modulation. As can be observed from the table, the radio consumes considerable power except when completely turned OFF.

4.1 Duty-Cycling Approach Taxonomy Table 4.2 Characterization of radio power (Schurgers et al. 2002)

137

Radio mode

Power consumption (mW)

Transmit (Tx) Receive (Rx) Idle OFF

14.88 12.50 12.36 0.016

Fig. 4.9 Low-power listen mode (Schurgers et al. 2002)

Power

Listen

Listen T

TRx T Time

Low-power listen mode

• To forward traffic, nodes on the multihop path need to be awakened, i.e., transitioned from the monitoring to the transfer state. The concern is that these nodes have no way of knowing when to transition if they did not detect that event. There is a dilemma to resolve though. For energy reasons, the nodes should turn OFF their radio when in the monitoring state, but they still need to be told somehow if they should turn it back ON. As a solution, each node periodically turns ON its radio for a short time to listen if someone wants to communicate with it. In the monitoring state, instead of being asleep, a node goes into this low-power listen mode, as shown in Fig. 4.9. The period of the listen–sleep cycle is denoted as T. The node that wants to communicate, the initiator node, polls the node it is trying to wake up, the target node. As soon as the target node, which is in the low-power listen mode of Fig. 4.9, hears the poll, the link between the two nodes is activated. If the packet needs to be relayed further, the target node will become an initiator for the next hop and the process is repeated. • Once the link between nodes is activated, data are transferred using a MAC protocol. MAC protocols are designed to organize access to the shared medium, in addition to contacting nodes. STEM strategy is thus to decouple the transfer and wakeup functionalities; in the monitoring state, low-power listen mode, as little energy as possible, is consumed, and only in the transfer state the MAC protocol is started. After the previous introduction of the basic setup, and in light of practical convenience, the following alternative scenarios lead to the adoption of two separate radios, one for wakeup and the other for data: • Without special protocol provisions, nodes are not synchronized; consequently, they do not know the phase of each other wakeup–sleep cycles in the listen mode. To avoid missing the short time where the target node has its radio ON,

4 Energy Management Techniques for WSNs …

138 Fig. 4.10 Interference due to aggressive wakeup (Schurgers et al. 2002)

A

B Regular transmission Interference D

C

the initiator has to poll continuously. As the data arrivals are uncorrelated with the sleep cycles, it will take about half a cycle for the target to hear the poll. However, this aggressive polling causes problems, as shown in Fig. 4.10. A regular data transmission is going on between nodes A and B. When node C wants to wake up node D, its aggressive polls will collide with the ongoing data transmission, essentially acting as a jammer to node B. Despite possible recovery action from the MAC layer, the data communication between nodes A and B suffers from extra delays. • Since this aggressive scheme is needed to limit the wakeup latency, the solution is to completely separate data transfer from wakeup. A realistic choice is to use two radios operating in separate frequency bands. As shown in Fig. 4.11, the radio in band f1 is only turned ON in the transfer state, and the wakeup band f2 can be viewed as a separate paging channel. STEM is not limited by the availability of an ultra-low-power radio for this paging channel; instead, the most efficient radio available might be used and the energy consumption further reduced by putting it in the low-power listen mode. This allows trading energy savings versus latency, beyond the capabilities of the radio alone. In principle, one radio can be used and switched between frequencies. However, if a target node is already transferring data and also has to wake up another node, it has to interrupt its data transmission or postpone the wakeup. Both are undesirable, and therefore, two radios are used. The penalty of making the node more expensive is minimal, as the radio typically accounts for less than 15% of the cost of a sensor node, e.g., an extra TR1000 radio (Murata Electronics 2015b) on the MICA motes (Crossbow 2002a, b). The wakeup radio is not a low-power radio so as to avoid problems associated with different transmission ranges. Therefore, an asynchronous duty-cycle scheme is used on the wakeup radio as well. Each node periodically turns ON its wakeup radio for Tactive every duration T. In this STEM-B scheme, when a source node (initiator) has to communicate with a neighboring node (target), it sends a stream of periodic beacons on the wakeup channel. As soon as the target node receives a beacon, it sends back a wakeup acknowledgment and turns ON its data radio. If a collision occurs on the wakeup channel, any node that senses the collision activates its data radio “up” (no wakeup acknowledgment is sent in case of collision). The wakeup beacon transmission is repeated up to a maximum time unless a wakeup acknowledgment is received from the target node.

4.1 Duty-Cycling Approach Taxonomy

139

Power

Active f1

Sleep

Sleep

Time

f2

Listen

Listen

Listen

Power

Listen

Time Sleep

Fig. 4.11 Separate wakeup and data frequencies using two radios (Schurgers et al. 2002)

• An alternate method to separate data transmissions from wakeup is to assign a separate time slot to each, as shown in Fig. 4.12. In this case, the initiator only sends the poll in the slot to which the target is listening. Inappropriately, this option requires time synchronization as all nodes have to remain synchronized at all times, which brings about considerable overhead. The monitoring state would thus be much more energy hungry than in the two radio options of Fig. 4.11. The same energy savings as in the layout of Fig. 4.11 could be achieved by deploying more nodes, but the total network deployment cost would exceed that of using slightly more expensive nodes. That is why the solution with two radios is chosen (Zheng et al. 2012). In addition to STEM-B, the above described beacon-based approach, a variant, STEM-T, that uses a wakeup tone instead of a beacon is proposed (Schurgers et al. 2002). The main difference is that in STEM-T all nodes in the neighborhood of the initiator are awakened. In STEM-B, the interbeacon period is such that there is enough time to send the wakeup beacon and receive the related acknowledgment (Fig. 4.11). Let Twakeup denote the time required for transmitting a wakeup beacon and Twack denote the time to transmit the related acknowledgment. Since nodes are not synchronized, the receiver must listen to the wakeup radio for a time Tactive  2 Twakeup + Twack to ensure the correct reception of the beacon. Clearly, Tactive depends on the bit rate of network nodes. In low bit-rate networks, the time between successive active periods (T) must be very large to allow a low duty-cycle on the wakeup channel. This results in large wakeup latency, especially in multihop networks with a large hop count. Both STEM-B and STEM-T can be used in combination with topology control protocols. For example, in a practical case the combination of GAF (section

4 Energy Management Techniques for WSNs …

140

Active

Listen

Power

Listen

Time Sleep

Fig. 4.12 Separate data and wakeup along time (Schurgers et al. 2002)

“Geographical Adaptive Fidelity (GAF)”) and STEM can reduce the energy consumption to about 1% of that of a sensor network with neither topology control nor power management. This increases the network lifetime by a factor of 100 (Schurgers et al. 2002); yet, STEM trades energy saving for path setup latency. Pipelined Tone Wakeup (PTW) PTW is proposed to achieve a tradeoff between energy saving and wakeup latency (Yang and Vaidya 2004). Like STEM, PTW relies on two different channels for transmitting wakeup signals and data packets, and uses a wakeup tone to awake neighboring nodes. Hence, any node in the neighborhood of the source node will be awakened. Unlike STEM, in PTW the burden for tone detection is shifted from the receiver to the sender. This means that the duration of the wakeup tone is long enough to be detected by the receiver that turns ON its wakeup periodically. The rationale behind this solution is that the sender only sends a wakeup tone when an event is detected, while the receivers wake up periodically. Also, the wakeup procedure is pipelined with the packet transmission so as to reduce the wakeup latency and, hence, the overall message latency. The concept is illustrated in Fig. 4.13b with reference to the string topology network of Fig. 4.13a. Supposing that node A has to transmit a message to node D through nodes B and C, PTW functioning can be outlined in several steps: • At time t0, node A starts the procedure by sending a tone to the wakeup channel. This tone awakens all A’s neighbors. • At time t1, node A sends a notification packet to node B on the data channel to inform that the next data packet will be destined to node B. • Upon receiving the notification messages, all A’s neighbors but B learn that the following message is not intended for them. Therefore, they turn OFF their data radio. Instead, node B realizes it is the destination of the next data message, and replies with a wakeup acknowledgment on the data channel. • Then, node A starts transmitting the data packet on the data channel. At the same time, node B starts sending a tone on the wakeup channel to awake all its neighbors. As shown in Fig. 4.13, the packet transmission from node A to node

4.1 Duty-Cycling Approach Taxonomy A

141 B

C

D

(a) String topology network Node B wakes up its neighbors

Node A wakes up its neighbors

Wakeup channel

Node C wakes up its neighbors

… Time

A notifies B B acks A Data channel

B notifies C C acks B

t0

t1

t2

…

B sends a packet to C

A sends a packet to B t3

t4

t5

Time

(b) Pipelined wakeup procedure

Fig. 4.13 PTW functioning (Anastasi et al. 2009)

B on the data channel and the B’s tone transmission on the wakeup channel are done concurrently. Like STEM, the data transmission is regulated by the underlying MAC protocol. In Yang and Vaidya (2004), it is shown by simulation that, if the time spent by a sensor network in the monitoring state is greater than 10 min, PTW outperforms STEM significantly, in terms of both energy saving and message latency, especially when the bit rate of sensor nodes is low. As the energy consumption of the wakeup radio is generally not negligible, both STEM and PTW use an asynchronous sleep/wakeup scheme for enabling a duty-cycle on the wakeup radio as well. A different approach is thru using a low-power radio for the wakeup channel. The low-power radio is continuously in standby mode, and whenever it receives a signal it wakes up the data radio, minimizing thus the wakeup latency (Shih et al. 2002). This approach suffers some drawbacks though: • The transmission range of the wakeup radio is significantly smaller than that of the data radio, which might limit the applicability of such a technique, since a node may not be able to wake up a neighboring node even if it is within its data transmission range. Typically, the low-power radio operating at 915 MHz ISM band has a transmission range of approximately 332 ft in free space, while the IEEE 802.11 card operates at 2.4 GHz with a transmission range up to 1750 ft (Shih et al. 2002). • Using a second radio for the wakeup channel bears additional power consumption, which may not be negligible even when using a low-power radio. To overcome problems associated with the extra energy consumed by the wakeup radio, a radio-triggered power management scheme is investigated (Gu and

4 Energy Management Techniques for WSNs …

142

Node Antenna

Radio

ON/OFF CPU

Radio-triggered circuit

Interrupt

Fig. 4.14 Radio-triggered power management (Anastasi et al. 2009)

Stankovic 2005). The basic idea is to use the energy contained in wakeup messages, such as STEM-B beacon, or signals such as STEM-T and PTW tones, to trigger the activation of the sensor node. This approach is similar to the one used in active radio frequency identification (RFID) systems (Want 2006). The radio-triggered scheme, in its simplest form, is illustrated in Fig. 4.14. A special hardware component, a radio-triggered circuit, is used to capture the energy contained in the wakeup message (or signal) and uses such energy to trigger an interrupt for waking up the node. The radio-triggered approach is significantly different than using a standby radio to listen to possible wakeup messages from neighboring nodes. Basically, the standby radio consumes energy from the node while listening, while the radio-triggered circuit is powered by the wakeup message. The main downside of the radio-triggered approach is the limitation on the maximum distance from which the wakeup message can be sent. With the simple radio-triggered circuit proposed in Gu and Stankovic (2005), the maximum achievable distance is 3 m. This distance may be increased up to a few dozen meters at the cost of a more complex and expensive radio-triggered circuit and increased wakeup latency due to limitations on the electronic circuit. For instance, the radio triggered wakeup with addressing capability (RTWAC) solution can achieve a distance up to 7.5 m (Ansari et al. 2009).

Scheduled Rendezvous Schemes Scheduled rendezvous schemes require that all neighboring nodes wake up at the same time. Typically, nodes wake up periodically to check for potential communications; then, they return to sleep until the next rendezvous time. The major advantage of such schemes is that when a node is awake it is guaranteed that all its neighbors are awake as well. This allows sending broadcast messages to all neighbors. On the flip side, scheduled rendezvous schemes require nodes to be synchronized in order to wake up at the same time. Clock synchronization in WSNs is a significant research and project topic. Detailed surveys on time synchronization

4.1 Duty-Cycling Approach Taxonomy

143

techniques are available in Sivrikaya and Yener (2004) and Faizulkhakov (2007). In the schemes presented in sections “Wakeup Scheduling Patterns in WSNs” and “Optimal Wakeup Scheduling of Data Gathering Trees for WSNs,” it is assumed that nodes are synchronized by means of some synchronization protocol. Wakeup Scheduling Patterns in WSNs In this model, the goal is to minimize the worst-case end-to-end overall delay, which includes both transmission delay and detection delay (Keshavarzian et al. 2006). Unlike the model in Lu et al. (2005) where general traffic flows are assumed, a specific yet very common and practical traffic pattern is considered such that the basestation (a central node) is either the source of the messages (forward direction) or the sink of the messages (backward direction), as illustrated in Fig. 4.15. By focusing on this traffic pattern, energy-efficient methods were designed, and compared to other methods a significantly better delay performance is attained. For example, in a four-hop network when the nodes wake up on average once every two seconds, the proposed wakeup methods guarantee a worst-case overall delay of less than 3 s over four hops in both forward and backward directions while D-MAC (Lu et al. 2004), presented in Sect. “An Adaptive Energy-Efficient and Low-Latency MAC for Data Gathering in WSNs (D-MAC)” achieves a worst-case delay of 6 s (in both directions). Even for data collection or monitoring applications where the backward delay is important, the proposed wakeup scheduling patterns perform Level 1

Level 2 Backward direction

Basestation

Forward direction

Legend: Lk denotes the set of nodes in level k

Fig. 4.15 Network and traffic model (Keshavarzian et al. 2006)

Level 3

144

4 Energy Management Techniques for WSNs …

better than other schemes. For the same system, D-MAC guarantees a delay of 2.1 s for backward direction, while the proposed multi-parent technique achieves a delay of less than 1.15 s, which is almost half of the wakeup period of the nodes. The WSN considered in this model comprises several tens of energy-constrained sensor nodes that either notify an event to a basestation or receive commands/ queries from the basestation, possibly over multiple hops. The basestation is assumed to be less energy constrained; however, it does not necessarily provide greater radio bandwidth or range than other nodes. The suggested model involves several concepts that are to be itemized: • Traffic model. The proposed traffic model as depicted in Fig. 4.15 involves two kinds of communication paths in the network, namely: – Forward direction (downlink) where the basestation sends a message to one of the nodes in the network. – Backward direction (uplink) where a regular node sends a message to the basestation. In a WSN, some sensor nodes are equipped with passive event detection capabilities that allow a node to detect an event even while in sleep mode. Other nodes provide ultra-low-power, low-rate periodic sampling mechanisms for rare event detection. Upon detecting an event, the sensor node is awaken within microseconds and gets ready to transmit a notification message to the basestation. Similarly, the basestation is often required to transmit imperative commands or queries to sensor nodes. Messages in either directions, thus, originate at random times (asynchronously) implying that messages may potentially originate at an inopportune time when all other nodes in the network are in sleep mode and not ready to receive the message. While these messages occur infrequently, they reflect urgency; as such, their delivery requires non-negotiable worst-case delay bounds. Delay is defined as the time duration between generating a message at a node, basestation or a regular node, until its eventual delivery at the destination node. • Channel sniffing and wakeup. Nodes in the network wake up from time to time and sniff the channel for activity. This is performed thru listening to the channel for a very short period of time and measuring the received signal strength. If the signal strength exceeds a predetermined threshold, the node remains awake awaiting to receive a possible transmission; otherwise, it powers itself OFF. Wakeup for sniffing constitutes the most frequent operation in the network and consequently is the most energy-consuming activity. To highlight the significance of the wakeup power consumption, typical data are recalled. According to the datasheet of CC1100 radio (Texas Instruments 2005), if the nodes wake up once every second the average current consumption over one second is 15 lA, which gives a charge draw of 15 lA per wakeup. This current draw might seem negligible in comparison with average current draw of 15 mA for reception or transmission of packets at a data rate 250 Kbps. However, in a 24 h day of network operation, the energy consumed by a node due to wakeup will add up to

4.1 Duty-Cycling Approach Taxonomy

145

15 lA * 3 V * 86,400 s = 3.9 J. This energy can be used to transmit or receive almost 21 Mbits of data, surpassing thus the overall traffic that passes through a node in one day for many applications. The length of the sniffing period and the energy consumed while performing a wakeup critically determine the longevity of the network. In practice, the sniffing length is determined by several hardware limitations such as the warm-up time of the radio and the minimum time required to reliably detecting a signal in the channel. Sniffing periods are typically in the order of hundreds of microseconds to few milliseconds. • Time Synchronization. It is assumed that a network-wide time synchronization protocol maintains a consistent notion of time between various sensor nodes in the network. Time synchronization in WSNs is implemented in Elson et al. (2002). Ganeriwal et al. (2003) can achieve synchronization within few milliseconds. Per se, synchronizing nodes within an accuracy of few milliseconds as required by the wakeup schedules is a relatively easy task. Although the time synchronization protocol may create additional energy burden for the system, in most delay-sensitive applications this extra energy cost either is negligible in comparison with the energy consumed by the wakeup process and/or will be compensated by the energy saving that can be achieved by employing an efficient synchronous wakeup method. Therefore, for delay-sensitive applications, synchronous wakeup methods are preferred due to their overall energy efficiency. • Network Topology. Each node in the network is represented by a node in a graph, and a link between two nodes signifies their ability to communicate with each other. An initial connectivity graph is formed by the basestation during network initialization followed by occasional updates to account for temporal changes in the wireless channel (Vasudevan et al. 2005). While wireless link qualities are subjected to changes temporally, two static nodes that are connected via a reliable link (high signal-to-noise ratio) rarely experience a whole change in their connectivity over short periods of time. It is assumed that the sensor network deployment is dense enough such that every node has few neighbors with highly reliable links. In such a network, if only reliable links are used for communication, the connectivity graph itself is not subjected to frequent changes. As such, it is assumed that the connectivity graph formed by using the reliable links of the network is stationary. • Notation. There are N nodes in the network. Levels are assigned to various nodes in the network in a breadth-first order based on the connectivity graph. The basestation is assigned a level 0. The level of a node signifies the minimum number of hops from the node to the basestation. This is illustrated in Fig. 4.15. Lk denotes the set of nodes in level k. The maximum number of hops, maximum number of levels, in the network is denoted by h. In a scheduled wakeup scheme, each node must be able to decide the times to sniff the channel for possible receptions. The simplest scheme is to schedule a node to wake up periodically after a fixed time interval. In a more sophisticated

146

4 Energy Management Techniques for WSNs …

scheme, each node may follow a periodic wakeup pattern, i.e., a sequence of predetermined wakeup times that exhibit periodicity. The period of a wakeup pattern is denoted by T. During one period T, the node may wake up multiple times; therefore to compare consistently across various schemes, the effective wakeup period Teff is defined as: Teff ¼ lim s=Ns s!1

ð4:1Þ

where Ns represents the number of wakeups in a time duration s. So on average, the nodes wake up once every Teff seconds. Also, the effective wakeup rate is defined to be: Reff ¼ 1=Teff

ð4:2Þ

The power consumption due to wakeups is thus given by: Pwakeup ¼ E0 =Teff ¼ Reff  E0

ð4:3Þ

where E0 is the energy consumed for each wakeup. The value of E0 depends on the hardware and on the duration that a node stays awake in each wakeup. For example, for CC1100, E0 = 3 V * 15 lC = 45 lJ. In the proposed approach, different wakeup patterns are provided and compared numerically for delay and lifetime values under two scenarios: • Fixed-power case. In this case, different patterns are compared based on their worst-case guaranteed delay when they consume the same amount of power and therefore provide the same lifetime. It is assumed that the amount of power allocated for the wakeup process is fixed such that: Pwakeup ¼ 0:5  E0 ; thus Teff ¼ 2 s So, the nodes wake up on average once every two seconds. • Fixed-delay case. It is assumed that the worst-case delay required by the application is fixed such that the maximum delay should be less than one second: maxðD! ; D Þ
1 Then, the amount of power each pattern consumes to satisfy this delay requirement is calculated. To convert the power consumption to network lifetime, a fixed battery capacity is assumed to be (2.4 * 108) * E0. This value is obtained assuming 2/3 of the capacity of the two AA batteries (1000 mAh) used on a CC1100 radio with E0 = 45 J (15 lC) per wakeup. In both cases, it is assumed that the network has h = 4 hops.

4.1 Duty-Cycling Approach Taxonomy

147

Several wakeup patterns are provided in the proposed model, and among them the fully synchronized pattern, the shifted even and odd pattern, and the ladder pattern are displayed in this section: • For the fully synchronized wakeup pattern as shown in Fig. 4.16, all the nodes in the network wake up at the same time according to a simple periodic pattern with a fixed period T = Teff. This pattern is very similar to the S-MAC protocol (Ye et al. 2002, 2004). The delay of a message that arrives at the basestation and is forwarded to a node in level 3 is shown in the figure. The worst-case delay in the network is simply h * T, and due to the symmetry of the pattern, the distribution of delay in both forward and backward directions is the same: D! ; D  U ½ðh  1Þ  Teff ; h  Teff 

ð4:4Þ

EðD! Þ ¼ ðh  1=2Þ  Teff

ð4:5Þ

The notation X * U[a, b] denotes that X is a continuous random variable with uniform distribution over the range [a, b]. The delay and lifetime values under the two proposed scenarios are found to be: – For the fixed-power case with h = 4 and Teff = 2 s: D! ; D  U ½6; 8; thus maxðD! ; D Þ ¼ 8 s – For the fixed-delay case, the nodes should wake up every Teff = 250 ms which gives Pwakeup ¼ 4  E0 , and 23.1 months network lifetime. • A shifted even and odd wakeup pattern is derived from the fully synchronized wakeup pattern by shifting the wakeup pattern of the nodes in even levels by T/2 as illustrated in Fig. 4.17. The figure shows the worst-case delay scenario; a message arrives to a level 3 node immediately after the wakeup time of the parent of the node. In this case, the first hop requires T seconds and the following (h − 1) hops each take T/2 s. The worst-case delay is therefore ðh þ 1Þ  T=2, and the delay distribution is: D! ; D  U ½ðh  1Þ=2  Teff ; ðh þ 1Þ=2  Teff 

ð4:6Þ

EðD! Þ ¼ ðh=2Þ  Teff

ð4:7Þ

The delay and lifetime values under the two proposed scenarios are thus: – For the fixed-power case scenario: D! ; D  U ½3; 5; thus maxðD! ; D Þ ¼ 5 s

4 Energy Management Techniques for WSNs …

148

Wakeup

Wakeup

Wakeup

Wakeup

T

Basestation Message arrives at basestation

Sleep

Level 1

Level 2

Level 3 Forward delay

Fig. 4.16 Fully synchronized wakeup pattern (Keshavarzian et al. 2006)

– For the fixed-delay case to achieve one second delay Teff = 400 ms which gives Pwakeup ¼ 2:5  E0 , resulting in 37 months lifetime, which is clearly better than the 23.1 months of the fully synchronized wakeup pattern. Thru such a simple modification, the delay for shifted even and odd wakeup pattern is almost half of the delay for the fully synchronized pattern, and the lifetime is significantly increased. In fact, in Lu et al. (2005) it is proved that in a network with tree topology this pattern provides the best overall average delay among all simple one-wakeup-per-period patterns with different shifts. • In the ladder wakeup pattern, the nodes still follow the simple periodic pattern but the wakeup patterns of different levels are staggered. Figure 4.18 shows this pattern where the wakeup is staggered in the forward direction. This idea is similar to the common practice of synchronizing the traffic lights to turn green (wakeup) just in time for the arrival of vehicles (packets) from the previous intersections (hops). This pattern has been given different names such as staggered wakeup (D-MAC) (Lu et al. 2004), streamlined wakeup (Cao et al. 2005), and fast path algorithm (FPA) (Li et al. 2005). The time difference between the wakeup times of two nodes in adjacent levels is denoted by s. By decreasing this value, the forwarding time of the message can be minimized. However, an intermediate node should fully receive the message

4.1 Duty-Cycling Approach Taxonomy

149

Wakeup

Wakeup

Wakeup

Wakeup

T

Basestation

Sleep T

Level 1

Level 2

/2

Level 3 Message arrives at level 3 node

Backward delay

Fig. 4.17 Shifted even and odd wakeup pattern (Keshavarzian et al. 2006)

before it can forward it to the next level, so the value of s is limited by the size of the message and the time required to transmit it. Typically, s should be in the order of tens of milliseconds. This wakeup pattern is no longer symmetric, so the forward and backward delay distributions are different. In the forward direction, the first hop requires between zero and T seconds, and then the next (h − 1) hops each require only a short period of length s. So, noting that Teff = T: D!  U ½ðh  1Þ  s; Teff þ ðh  1Þ  s

ð4:8Þ

EðD! Þ ¼ Teff =2 þ ðh  1Þ  s

ð4:9Þ

For backward direction, the first hop again requires at most T seconds, and the next hops each take (T − s) seconds. Noticeably, the wakeup time of the basestation does not impact the forward delay. Hence, in order to reduce the backward delay, the basestation wakes up after (instead of before) the wakeup time of the L1 nodes (Fig. 4.18). The distribution of the backward delay is given by:

4 Energy Management Techniques for WSNs …

150

Wakeup

Wakeup

Wakeup

Wakeup

T

Basestation Message arrives at basestation

Sleep

Level 1

Level 2

Level 3

Forward delay Backward delay Fig. 4.18 Forward ladder wakeup pattern (Keshavarzian et al. 2006)

D  U ½ðh  2Þ  ðTeff  sÞ þ s; ðh  1Þ  Teff  ðh  3Þ  s EðD Þ ¼ ðh  3=2Þ  Teff  ðh  3Þ  s

ð4:10Þ ð4:11Þ

Assuming s ¼ 50 ms for both delay scenarios, the delay and lifetime values are obtained as: – For the fixed-power delay case with Teff = 2 s: D!  U ½0:15; 2:15; D  U ½3:95; 5:95 so the maximum delay is 5.95 s. – For the fixed-delay case to achieve one second delay in both directions, it is required that: Teff ¼ 350 m s;

thus Pwakeup ¼ 2:86  E0

4.1 Duty-Cycling Approach Taxonomy

151

which gives 32.4 months network lifetime. Notably, the delay in the forward direction is significantly reduced but the backward delay is almost the same as the fully synchronized wakeup pattern. So when the delays in both directions are considered, there is no major improvement in the worst-case delay or the lifetime of the network. The forward ladder wakeup pattern shown in Fig. 4.18 can be reversed to create the backward ladder pattern, which improves the backward direction and is essentially the same as the wakeup method used in D-MAC (Lu et al. 2004; Cao et al. 2005) and FPA (Li et al. 2005). Since nodes at different levels of the data gathering tree wake up at different times, the ladder wakeup pattern has several advantages with respect to the fully synchronized approach: • At a given time, only a small subset of nodes in the network will be active. Thus, the number of collisions is potentially lower as only a subset of nodes contends for channel access. • The active period of each node can be much shortened, thus resulting in energy saving. This scheme is also suitable for data aggregation. Parent nodes receive data from all their children before they forward such data to their own parent at the higher level. This allows parent nodes to filter data received from children or to aggregate them. The ladder wakeup pattern has though some drawbacks in common with the fully synchronized wakeup pattern: • Since nodes located at the same level in the data gathering tree wake up at the same time, collisions can potentially occur. • This wakeup pattern has limited flexibility due to the fixed duration of the active and wakeup periods.

Optimal Wakeup Scheduling of Data Gathering Trees for WSNs This scheme addresses the problem of scheduling wakeup intervals for all nodes in a WSN data gathering tree such that a preset delay deadline is met while utilizing the minimum amount of total power in the WSN (Ungjin et al. 2012). A multihop route in the data gathering tree, assumed to be a fixed topology, is used to transmit sensor data from a sensor node to the gateway node. However, since two adjacent nodes must be active at the same time in order for the two nodes to be able to communicate, a local time synchronization mechanism whereby two adjacent nodes know the clock values and wakeup periods are used by each other is assumed. In addition to its regular wakeup interval, if a node has data to send, it will become active during the next wakeup period of its parent node in the data gathering tree. For WSNs that have a delay limitation in data delivery from any node to the

152

4 Energy Management Techniques for WSNs …

gateway node, an optimal algorithm is devised to determine the wakeup frequency of each node. The power consumption model takes into account possible differences in data rates. For WSNs, a most essential requirement is low power usage by each sensor node since nodes are mostly powered thru small batteries that are hard to replace in the field. Therefore, modules that control a sensor node, such as the microcontroller, RF radio, and sensors, are designed for low power usage. Nevertheless, the most effective method for power conservation is the use of periodic sleep/wakeup cycles with long sleep periods followed by short wakeup periods during which necessary communication and sensing activities can take place. For example, the ATmega128 (Atmel 2011), one of the most popular microcontrollers used in WSNs, consumes several hundreds of times lower power during sleep mode than active mode, even when idle active. The approach to be presented in what follows offers particular features: • Assigning a different wakeup frequency for each node to minimize the total power consumption in the data gathering tree. Contrarily, several works related to wakeup scheduling adopted the same wakeup frequency for all nodes (Ye et al. 2004; Rajendran et al. 2006; Keshavarzian et al. 2006) presented in section “Wakeup Scheduling Patterns in WSNs” or probabilistic wakeup frequencies (Lai and Paschalidis 2006; Tang et al. 2011). • Ensuring delivery delays from any node to the gateway node. Although several other works considered the tradeoff between power consumption and delay, except for Cohen and Kapchits (2009), none of them guaranteed a delay bound. • Compared to the most similar scheme provided in Cohen and Kapchits (2009), the two main differentiating factors in the proposed approach are a more accurate power consumption model and an accounting for the differences in data rates of child and parent nodes in the data gathering tree. To model the system, a given WSN consisting of n nodes can be represented by an undirected graph G = (V, E). V is the set of sensor nodes in the WSN, and E is the set of edges between each pair of nodes that are able to communicate with each other. A data gathering tree, which is a spanning tree of G, is used to transfer data from each sensor node to the gateway node. The data gathering tree can be constructed by a spanning tree algorithm (Chen et al. 2002; Lee and Wong 2005). Although this appears to imply the use of a static data gathering tree, dynamic trees could also be considered by periodically changing the spanning tree being used. Given a static data gathering tree, the energy levels of nodes closest to the sink will tend to become the fastest to deplete because all paths to the sink must go through them. To prevent such a situation, it may be desirable to periodically reconfigure the data gathering tree being used, after moving or changing the sink node. If this is done using one of the spanning tree construction methods, the approach presented can also be effective with dynamic data gathering trees. Aiming at power conservation, every node, except the gateway node, which could have its own power supply source, goes to sleep and wakes up in a periodic

4.1 Duty-Cycling Approach Taxonomy

153

manner. When a node wakes up, it can receive data from its child nodes and can send data to its parent. A node that needs to forward data to its parent should transmit the data while the parent is awake. Therefore, it is assumed that each node in the WSN is globally or locally time synchronized with its neighboring nodes and is also aware of the times when its neighbors wake up. The energy conservation model used assumes a WSN with a very low sleep/wakeup duty-cycle (asleep most of the time). However, since only very loose time synchronization is required, methods with a long synchronization period could still be used (Elson et al. 2002; Ganeriwal et al. 2003). Data transfers occur when both the sender and the receiver nodes are awake. When a parent node wakes up, it stays awake until all of its child nodes complete sending their data. Also assumed, each node transmits at least one packet that contains its own sensed data whenever its parent wakes up. In the case of an intermediate node having at least one child node, it should transfer the packets it generates and those received from its child nodes. Received packets are temporarily stored until the parent node wakes up, and then they are forwarded to the parent. Under these assumptions and considerations, the power consumption of each node consists of three parts: • Power consumption for data reception. Let Dv be the amount of data received from the child nodes of node v when v wakes up. The received data are composed of the packets generated by the child nodes of v and the packets forwarded through them. Let dv be the size of the data packet generated by v. The set of child nodes of v will be denoted by c(v). Each node, except the gateway node, will sleep and wake up according to a preset schedule. Let fv denote the wakeup frequency of a node v. Then, the data delivery rate Dv can be represented by Eq. 4.12:

Dv ¼

8 > < > :

P

dci ðvÞ

if all child nodes of m are leaf nodes

ðdci ðvÞ þ fci ðvÞ  Dci ðvÞ =fv Þ

otherwise

Ci ðvÞ2cðvÞ

P

Ci ðvÞ2cðvÞ

ð4:12Þ For the upper component in Eq. 4.12, Dv is the sum of the data generated by the child nodes of v because leaf nodes only have the data they generate. For the lower component, the child nodes of v transfer data received from their child nodes as well as data they generated. A node ci(v) 2 c(v) receives data equal to fci ðvÞ  Dci ðvÞ per unit time from its child nodes. Therefore, the amount of data forwarded to v is this value divided by fv. Let erv be the average energy consumed by v to receive a packet from its child node and orv be the overhead energy used when receiving data, mainly due to synchronizing the listening channel while waiting for packets from child nodes. Since node v receives data from its child nodes at every wakeup interval, the power v for receiving data is represented by Eq. 4.13:

4 Energy Management Techniques for WSNs …

154

pv

r

  ¼ fv  erv  Dv þ orv

ð4:13Þ

• Power consumption for data transmission. A node sends data during every wakeup interval of its parent node. Therefore, the power consumption for data transmission is related to the wakeup frequency of the parent node. Let esv be the average energy consumed by v to send a packet to its parent node and osv be the overhead energy used to send data mainly consumed due to snooping the channel to avoid contention with other nodes. Then, the power consumed by v to send data is denoted by Eq. 4.14: pv

s

    ¼ fpðvÞ  esv  dv þ fv  Dv =fpðvÞ þ osv   ¼ fpðvÞ  esv  dv þ osv þ fv  esv  Dv

ð4:14Þ

  The expression dv þ fv  Dv =fpðvÞ represents the amount of data sent by node v. The scaling factor fv =fpðvÞ for Dv is applied for the same reason as in Eq. 4.12. • Power consumption for other activities. Sensor nodes execute additional activities, such as time synchronization, localization, or routing tree construction, as well as data gathering. Even though these activities are periodically executed, since they are independent of the wakeup frequency for data gathering, the power consumption for these activities can be considered as constants. Let pv_b be the power consumed for all such activities. From Eqs. 4.12 to 4.14, the power consumption of an arbitrary node v in the data gathering tree can be as Eq. 4.15 illustrates: pv ¼ p v ¼ fv 

þ pv s þ pv b  r     ev þ esv  Dv þ orv þ fpðvÞ  esv  dv þ osv þ pv

r

ð4:15Þ b

The average energy consumed, esv , of the gateway node is zero because the gateway node only receives data from its child nodes and does not forward data to any other node. Also, the wakeup frequency fv of every leaf node is zero because leaf nodes do not wake up to receive data. To formalize the problem, let T(vg) represent the data gathering tree and vg be the root of T(vg), and vg is also the gateway node of G. Let PATH(v) be the set of nodes on the data delivery path from v to vg in the data gathering tree. Then, the formalization can be as follows: • Problem MIN_WAKEUP. Given a data gathering tree T(vg), determine the wakeup frequency of each node in T(vg) that minimizes: X      ðfv  erv þ esv  Dv þ orv þ fpðvÞ  esv  dv þ osv Þ v2T ðvg Þ

4.1 Duty-Cycling Approach Taxonomy

subjected to the following constraints: – For every node except leaf nodes in T(vg),

155

P

1=fi
D.

i2PATHðvÞ

– For every node except leaf nodes in T(vg), fv > 0. Δ in the first constraint above denotes the maximum delay permitted when delivering data from any node in the data gathering tree to the gateway node. If a child node of v wishes to send data to v, it must wait until the next wakeup interval of v. In the worst case, this waiting time is equal to the sleep interval of v. Thus, since the sleep interval will typically be much longer than the wakeup interval in order to achieve effective power conservation, the data delivery delay of v is approximated by the inverse of fv. Delay variations caused by processing or channel contention are typically much smaller than the above waiting time and are thus omitted for simplicity. The overhead power component pv_b is also not included in the objective function in the MIN_WAKEUP problem because pv_b is independent of the wakeup frequency of each node. The optimal wakeup frequency assignment (OWFA) for the MIN_WAKEUP problem achieved several objectives: • Minimizing the total average power consumption of the tree while limiting the total delay for delivering data from any node in the tree to the gateway node. The power consumption model takes into account the differences in data rates of each node in the tree. • Analysis proves that OWFA assigns optimal wakeup frequencies to all nodes. • Simulation results under various conditions show that OWFA consumed about 8.6–24.3% less average power and thus resulted in a 7.4–26.0% longer network lifetime over the alternative wakeup frequency determination algorithm that does not consider differences in data rates (Cohen and Kapchits 2009). • OWFA can also be applied to a dynamic tree topology leading to a significant increase in the network lifetime when compared to a static tree topology.

Asynchronous Schemes Asynchronous wakeup was first studied in Tseng et al. (2003); an investigation was carried on several wakeup schedules, and a necessary modification is done on IEEE 802.11 power saving mechanism (PSM) to support asynchronous wakeup. As compared to the scheduled rendezvous wakeup mechanisms, asynchronous wakeup does not require clock synchronization. Each node follows its own wakeup schedule in idle states, as long as the wakeup intervals among neighbors overlap. To meet this requirement, nodes usually have to wakeup more frequently than in the scheduled rendezvous mechanisms. On the flip side, asynchronous wakeup is easier to implement and can ensure network connectivity even in highly

156

4 Energy Management Techniques for WSNs …

dynamic networks. In other words, asynchronous wakeup trades energy consumption for the robustness of network connectivity. A key challenge is to derive schedules that have minimum idle state energy consumption with bounded neighbor discovery latency. In sections “Asynchronous Wakeup Protocol (AWP) for Ad Hoc Networks” and “Random Asynchronous Wakeup (RAW) Protocol for Sensor Networks,” power management via asynchronous wakeup schemes is elucidated. Asynchronous Wakeup Protocol (AWP) for Ad Hoc Networks In this scheme, a systematic approach is adopted to design asynchronous wakeup protocol (AWP) in ad hoc networks (Zheng et al. 2003). Several fundamental questions related to asynchronous wakeup are raised and answered: • Given a desirable delay for neighbor discovery, what is the minimum percentage of time a node has to be awake? • Is there an optimal schedule that can achieve such minimum value? • How to design a wakeup protocol using the optimal schedule? • How can power management be performed with asynchronous wakeup? To answer the first two questions, there is a problem formulation of generating wakeup schedules as a block design problem in combinatorics. Then, a derivation of the theoretical limit of the wakeup schedule is conducted, and an optimal solution is given to achieve minimum idle state energy consumption with bounded neighbor discovery latency. To answer the last two questions, the theoretical base is laid out, and two protocol design issues are addressed: • Efficient implementation of the wakeup schedule. A neighbor discovery and neighbor schedule bookkeeping protocol are devised to operate without requiring slot boundaries to be aligned. The protocol is also resilient to packet collision and network dynamics. • Power management using asynchronous wakeup. Two design choices are considered to determine how a node transitions among different power management modes, specifically, slot-based power management and on-demand power management. In slot-based power management, power management modes are managed slot by slot based on the number of buffered packets for a particular neighbor, while in on-demand power management the transitions between power management states are triggered by the presence/lack of certain communication events. In both cases, a desirable communication schedule can be overlaid over the wakeup schedule. To state the problem, a multihop ad hoc network is represented by a directed graph G(V, E), where V is the set of network nodes (|V| = N) and E is the set of edges. If node vj is within the transmission range of node vi, then an edge (vi,vj) is in E. Bidirectional links are assumed; therefore, if (vi,vj) 2 E then (vj,vi) 2 E. The neighboring set of a node v is denoted by N(v). The major objective of asynchronous

4.1 Duty-Cycling Approach Taxonomy

157

wakeup mechanisms is to maintain network connectivity regardless of the power states of nodes. The term “connectivity” is used loosely, in the sense that a topologically connected network may not be connected at any time; instead, all the nodes are reachable from a node within a finite amount of time. In the absence of data transmission, i.e., when a node is in the idle state, a wakeup mechanism associates each node with a slot schedule of length T, termed as the wakeup schedule function (WSF). The WSF of a node v is represented by a P i polynomial of order T − 1 as fx ðvÞ ¼ T1 i¼0 ai  x , where T is the length of the schedule, a = 0 or 1 for 8i 2 ½0; T  1, and x is a placeholder. If ai = 1, the node should wake up in slot i. For two neighboring nodes to communicate, their wakeup schedules have to overlap regardless of the difference of their clocks. By definition, kv = fv(1) is the total number of slots in which node v is scheduled to be awake every T slots. If the schedules of two nodes u and v overlap, the amount of time it takes for node u to reach node v is bounded by T; i.e., T is the worst-case delay to discover a neighbor due to power saving. Given a fixed value of T, it is required to make kv =T as small as possible, subjected to the constraint that the schedules of any two neighboring nodes overlap by m slots. The rationale behind seeking a small value of kv is to enable nodes to stay in the low-power mode as much as possible. Given the wakeup schedule fu(x) for node u, it is easy to show that fuk ¼ fu ðxÞ  xk ðmodðxT  1ÞÞ represents the cyclic shift of the original schedule by k slots. Let K be the “and” operator between two polynomials and |∙| be the number of nonzero items of a polynomial. The following definition is made: • Definition 1: The degree of overlapping between two WSFs fu(x) and fv(x) denoted as C(u, v) is defined to be the minimum number of common items of fu ðxÞ  xl ðmodðxT  1ÞÞ and fv ðxÞ xk ðmodðxT  1ÞÞ for any integer l, k 2 [0, T − 1], i.e., C ðu; vÞ ¼ minl;k2½0;T1 ful ðxÞ ^ fvk ðxÞ. Definition 1 mandates that the number of overlapping slots should be shift-invariant. Therefore, the problem of designing optimal WSFs can be stated as follows: Optimal WSF design problem. Given a fixed value of T, minimize k such that c(u, v)  m, for all 8u 2 V; and 8v 2 NðuÞ; where k is the assembled average of the number of active slots in every T slots. How small the value of k can be in the optimal WSF design problem. Theorem 4.1 shows the constraints that ku, 8u 2 V, must satisfy: Theorem 4.1 Bound for the WSF design. Consider any two neighboring nodes u and v with WSFs of length T, fu(x) and fv(x), respectively. Given T and m, the necessary condition for c(u,v)  m is kv  ku  m  T, where ku = fu(1) and kv = fv(1) are, respectively, the number of slots in which node u and node v are pffiffiffiffiffiffiffiffiffiffiffiffi scheduled to be awake every T slots. It was proved that k  m  T .

4 Energy Management Techniques for WSNs …

158

It follows immediately from Theorem 4.1 that when kv = k, 8m 2 V, i.e., the duty-cycles of all nodes are the same (symmetric design). The following corollary gives the lower bound of the minimum duty-cycle in the symmetric WSF design problem: Corollary 4.1 Bound for symmetric design. Given T and m, and kv = k, 8m 2 V, pffiffiffiffiffiffiffiffiffiffiffiffi the feasible set of the optimal WSF design problem satisfies k  m  T . The problem of symmetric WSF design can then be mapped to the symmetric block design problem in combinatorics:

Schedule

• Definition 2: (v, b, r, k, k)-design is a family of b blocks of size k from a set V of size v, such that each element of V appears in r blocks and every two elements of V appear in k common blocks. In particular, if b = v, or equivalently r = k, a (v, b, r, k, k)-design is called a (v, k, k)-design or a symmetric design. It can be simply shown that under symmetric design every two blocks have common elements. As an example, consider the following design (124, 235, 346, 457, 561, 672, 713), where the number gives the positions of active slots. As shown in Fig. 4.19, this is a (7, 3, 1)-design; i.e., there are 7 blocks of size 3, and any two blocks have one common element. In the context of the WSF design problem, T = b, m = k, i.e., the length of each schedule (block) is T, and the number of overlapping (common elements) of any two schedules (blocks) is m. This scheme ensures that each node will be able to contact any of its neighbors in a finite amount of time. However, the packet latency introduced may be large, especially in multihop networks. In addition, it never happens that all neighbors are simultaneously active. Therefore, it is not possible to broadcast a message to all neighbors.

1

2

3

4

5

6

7

Slot number

Fig. 4.19 Slot assignment under the (7, 3, 1)-design (Zheng et al. 2003)

4.1 Duty-Cycling Approach Taxonomy

159

• In the case of asymmetric design, the duty-cycles of nodes are different. Theorem 4.1 gives a necessary condition for ensuring connectivity among neighboring nodes. In the case m = 1, this necessary condition can be trivially satisfied by assigning ku = T and kv = 1 for m 2 N(u). Asymmetric design is applicable in heterogeneous networks, where there exist powerful nodes with abundant power supply. These nodes can be assigned WSFs with high duty-cycle and serve as relay nodes for power-constrained nodes that use WSFs with low duty-cycles. After the clarified buildup of the theoretical foundations and requirements, a wakeup schedule is needed to maintain network connectivity when nodes are in power saving mode, and a power management policy is needed to further facilitate effective communication while saving as much as energy as possible. If nodes can only communicate during the active slots of the wakeup schedule, the capacity of the network will be greatly reduced and the delay experienced by packets may be prohibitively long, i.e., in the worst case, n*T for communication over an n-hop path. A power management policy overlays a desirable communication schedule over the wakeup schedule to decide when a node should go to sleep and wakeup. The relationship between the wakeup schedule and the communication schedule devised by a power management policy is illustrated in Fig. 4.20. To verify the effectiveness of the design, the proposed scheme was implemented in ns-2 (Fahmy 2016) using IEEE 802.11 MAC, without the power management component. Simulation studies indicated several motivating results: • The proposed wakeup schedule design guarantees that any two neighboring nodes can detect each other in finite time without global clock synchronization.

Power saving mode

Active mode

Wakeup schedule Awake

Sleep

Communication schedule Send /Receive

Idle

Power management policy

Fig. 4.20 Relationship between the wakeup schedule and the communication schedule (Zheng et al. 2003)

160

4 Energy Management Techniques for WSNs …

• In conjunction with the power management policies, the proposed wakeup protocol can achieve communication efficiency comparable to the case without power management while saving energy up to 70%. • The energy consumed under the asynchronous wakeup mechanism with the (7, 3, 1)-design is approximately half that without power management, while the energy consumed under the (73, 9, 1)-design is approximately 1/3–1/4. • As compared to the asynchronous wakeup mechanism with the (7, 3, 1)-design, the mechanism with the (73, 9, 1)-design can achieve much lower energy consumption per unit data delivery at the expense of slightly higher packet loss rate. • On-demand power management in conjunction with the asynchronous wakeup mechanism with the (73, 9, 1)-design can achieve a good balance between energy consumption and packet delivery ratio.

Random Asynchronous Wakeup (RAW) Protocol for Sensor Networks Random asynchronous wakeup (RAW) is a power saving technique for sensor networks that reduces energy consumption without significantly affecting the latency or connectivity of the network (Paruchuri et al. 2004). RAW builds on the observation that when a region of a shared-channel wireless network has a sufficient density of nodes, only a small number of them need to be active at any time to forward traffic for active connections. RAW is a distributed, randomized algorithm where nodes make local decisions on whether to sleep or to be active. Each node is awake for a randomly chosen fixed interval per time frame. High node density leads to the existence of several paths, between two given nodes, whose path length and delay characteristics are similar to the shortest path. Thus, a packet can be forwarded to any of several nodes in order to be delivered to the destination without much affecting the path length and delay experienced by the packet if compared to when forwarded through the shortest path. System lifetime due to RAW increases with the increase of idle-to-sleep energy consumption ratio and the density of the network. Knowing that energy conservation is of paramount importance in WSNs, several solutions address the problem of energy wastage due to idle listening. The main sources of energy wastage are collisions, idle listening, overhearing, and control packet overhead: • MAC protocols (Ye et al. 2002; Chen et al. 2002; Van Dam and Langendoen 2003) whether contention-based, like CSMA, or scheduled protocols, like TDMA, target avoiding collisions. • The second major energy wastage source is idle listening, which occurs when the receiver is listening to the channel to gather possible data. As noted throughout this chapter, the energy spent during idle listening is comparable to the energy spent during transmitting or receiving.

4.1 Duty-Cycling Approach Taxonomy

161

• Overhearing occurs when a node receives packets that are destined to other nodes. Overhearing unnecessary packets can be a significant factor in energy wastage when the network is highly loaded or when the node density is high. • Lastly, sending, receiving, and listening of control packets consume energy, which reduces the effective throughput. One approach to prevent energy wastage due to the above sources is to control the node receiver by setting it to sleep mode when no data are expected and to wakeup mode when communication is expected (Yang and Vaidya 2004). To reduce latency, RAW uses the concept of stateless non-deterministic geographic forwarding (SNGF) (He et al. 2003). Unlike geographic routing where a packet is forwarded to a node that is closest to the destination, a RAW packet can be forwarded to any node in a forwarding set, as detailed later. This design reduces the energy consumption due to idle listening and reduces latency because of the presence of multiple forwarding nodes. RAW mainly consists of two components, routing based on forwarding sets and random wakeup scheme: • Routing based on forwarding sets. The routing methodology in RAW is designed to take advantage of the fact that WSNs are densely deployed. A high node density results in the existence of several paths between two given nodes, whose path lengths are very close to the length of the shortest path. Thus, a packet can be forwarded to any of such several paths in order to be delivered to the destination without affecting the path length and delay when compared to the shortest path. RAW allows for a node to be active during a randomly chosen fixed interval in each time frame. This removes the necessity of time synchronization and makes the protocol implementation very simple. In the geographic routing protocol, a packet is forwarded to a neighboring node that is closest to the destination. However, in a WSN, in which not all nodes might be active at a given point of time, a packet can be forwarded to the active neighbor that is closest to the destination, or the packet can be queued until the closest neighbor becomes active, and the packet can then be forwarded to this neighbor. A modification of the geographic routing protocol is made such that a packet is sent to any of the active neighbors that meet a forwarding criterion. A neighboring set and a forwarding candidate set are defined as follows: – The neighbor set of node i is the set of nodes that are inside the radio range R of node i: NSi ¼ fnodejnodeðnode; node iÞ
Rg

ð4:16Þ

– The forwarding candidate set of node i for a given destination is the set of potential neighboring nodes to which node i can forward a packet. Two criteria are adopted for defining the forwarding candidate set: One is based

162

4 Energy Management Techniques for WSNs …

on path lengths, and the other is based on geographic distance to the destination: Hop-based forwarding candidate set (h-FCS). The forwarding criterion for a given source s and destination d states that a neighbor k of s is a node in FCS if: H ðk; d Þ\H ðs; d Þ þ D

ð4:17Þ

where H(i, j) is the hop length of the shortest path between nodes i and j. When Δ = 0, it is implied that a shortest path between s and d exists through node k. When Δ > 2, every neighbor of s belongs to h-FCS. This is because, for a given neighbor k, there always exists a path s ! k ! s … ! d whose length is H(s, d) + 2, thus satisfying the forwarding criterion. Also, it should be noted that unless Δ = 0, selecting a forwarding node based on this forwarding criterion does not guarantee that a packet reaches the destination, as the path length to the destination from any two neighbors in the path can be the same. Computing h-FCS requires each node to know the shortest path length to all other nodes in the network. Thus, this criterion of selecting FCS might not be very appealing owing to the computational overhead involved. To overcome this overhead, a selection of FCS is proposed based on the geographic distances between the nodes. Distance-based forwarding candidate set (d-FCS). The forwarding criterion for a given source s and destination d states that a neighbor k of s is a node in FCS if: Dðk; d Þ\Dðs; d Þ  Th

ð4:18Þ

where D(i, j) is the geographic distance between nodes i and j. Thus, if a neighbor k is closer to the destination by at least a threshold, Th, than node s itself, then k belongs to the forwarding candidate set (Fig. 4.21). The d-FCS selection criterion guarantees that there would be no loops in the path, since a node always forwards a packet to a node that is closer to the destination than itself. At the same time, this simple criterion cannot guarantee the delivery of a packet to the destination in the presence of holes. At high network densities, it can be safely assumed that holes would not exist. In case holes are present, criteria for selection are to be extended (Fang et al. 2006). Routing based on forwarding sets increases the path length. The Th value limits the maximum path length, since with each transmission a packet traverses at least a distance Th toward the destination. Intuitively, due to increased path lengths, it might seem that forwarding set-based routing adds additional overhead in terms of energy consumption.

4.1 Duty-Cycling Approach Taxonomy

163

L d L-Th

s

Fig. 4.21 Forwarding candidate set (Paruchuri et al. 2004)

However, when combined with the random wakeup scheme, the total energy consumed by a sensor in RAW is lower. • Random wakeup scheme. Each node wakes up once in every slot for a predetermined time and goes back to sleep. To elaborate, consider for each sensor node time slots of fixed interval T and active time Ta (Ta < T). Thus, if there are m neighbors in the forwarding set of node s to which a packet destined to d can be transmitted, then the probability that at least one of those nodes is awake, when s is awake, is given by: P ¼ 1  ð1  2  Ta =T Þm

ð4:19Þ

It was shown that even for Ta as low as 15%, and at low node density, a node could find an active neighbor to whom it can forward the packet with high probability (>82%). For higher densities, the probability is even higher. Thus, even if a node is active for a randomly selected duration Ta, there is a high probability that a packet can be forwarded to the destination. Random wakeup as the basis of RAW is composed of two modules, neighbor discovery and packet forwarding: – The neighbor discovery procedure. Whenever a node i wakes up, it broadcasts a beacon message piggybacking its own id, the start time of its wakeup period, and other information subjected to channel contention/resolution rule. To implement the protocol, each node keeps a two-hop neighbor list in which each entry has the fields shown in Fig. 4.22.

Node id Clock Schedule Lifespan Location Fig. 4.22 Entry fields in a neighbor list maintained by each node (Paruchuri et al. 2004)

4 Energy Management Techniques for WSNs …

164

A new entry is added whenever a new neighbor is discovered. Also, among the neighbors of i, node j that has been awake for the longest period sends a beacon message to i as an acknowledgment, and it also piggybacks its neighbor list. All nodes that receive the acknowledgment beacon update their neighbor lists according to the neighbor list of j. This ensures consistency in neighbor lists of all nodes. – Packet forwarding. A greedy geographical routing protocol forwards a packet to an active neighbor that is closest to the destination at each hop. Whenever a node i has a packet destined to node d, it selects a node k from its one-hop neighbor list, such that k is closer to d than any other active neighbor of i, and k is closer to the destination by at least Th. The threshold Th limits the length of a path to a maximum Dðs; d Þ  R=Th, where D(s, d) is the distance between the source and the destination, and R is the transmission range of the sensors. To evaluate the performance of RAW, simulation was carried on using OMNeT++, a discrete event simulation framework (Fahmy 2016). All simulations were based on a network of dimension 5R  5R, where R is the transmission range of a sensor node. Various node densities were considered. The model parameters and limits on transmission bit rates and energy ratings are set according to Crossbow MICA2 sensor nodes (Crossbow 2002). Power consumption in the model is based on the amount of current a Crossbow MICA2 sensor node radio transceiver draws, as listed in Table 4.3. A radio transmission rate of 76.8 Kbps is assumed. Several findings were grasped after extensive simulation runs: • Effect of different threshold values, Th, on the performance of RAW. It was spotted that as Th increases, the number of neighbors, which are closer to the destination by at least Th than the node itself, decreases and, hence, the size of the forwarding candidate set (FCS) decreases as well. Thus, the probability that a node in the FCS is active decreases, and hence the probability that a packet is buffered increases. This leads to latency increase at higher Th values. • Effect of Th on the delivery ratio. The delivery ratio is the ratio of packets received at the corresponding destinations to the number of packets generated for the network as a whole. At lower thresholds, though the latency is low, the path length can be very high, thus resulting in many transmissions and collisions. At the same time, for a higher Th, a packet would be buffered several times, thus increasing the latency and decreasing the capacity of the network. The best delivery ratios are obtained at Th = R/3 where a maximum delivery ratio around 98% is attained.

Table 4.3 Typical current draw values (Crossbow 2002)

Transmit

Receive

Idle

Sleep

15 mA

8 mA

7 mA

2 lA

4.1 Duty-Cycling Approach Taxonomy

165

• Effect of time frame length. When setting the time frame length, it should be observed that once a node wakes up, it has to be active for at least a period that allows the node to transmit the beacon message, receive a reply to the beacon, and transmit at least one data packet. The higher the active period of a node, the higher is the frame length, in order to maintain low duty-cycle, Ta/T. This results in higher latency, as a packet will be buffered for longer periods. At the same time, if the active period is short, not many packets can be forwarded and packets might be buffered several times, resulting in higher delay. • Performance of RAW. The performance metrics of interest are the amount of power consumed, the packet delivery ratio, and the latency experienced by the packets. Specifically, it was observed that: – The delivery ratio increases with active time Ta. RAW shows better performance in large networks, and it is scalable with respect to density; in fact, the performance improves with density. – The longer active time Ta for a node results in a lower latency, and higher probability that a node finds a neighbor to which it can forward the packet. – RAW consumes about 65% less energy than what the scheme without power management consumes. It is important to observe the tradeoffs between latency, delivery ratio, and energy consumed, against the amount of time a node is active. A higher node active time can achieve better latency and delivery ratio, but at the cost of increasing the amount of energy consumed. The appropriate choice of node active time depends on the type of application the WSN is deployed for, as well the amount of latency and delivery ratio the network can tolerate. • Compared to AWP (section “Asynchronous Wakeup Protocol (AWP) for Ad Hoc Networks”), the average latency in RAW is significantly lower at intermediate nodes, while the energy consumption of RAW is comparable to that of AWP. Contrary to AWP, RAW does not consider mobile networks. Appraisal of Sleep/Wakeup Protocols A recapitulation of the sleep/wakeup protocols presented in Sect. 4.1.2.1 is now due. On-demand protocols have several assets that make them maybe ideal when compared to scheduled rendezvous and asynchronous protocols. Explicitly specifying (Anastasi et al. 2009): • They maximize energy saving, since nodes remain active only for the minimum time required for communication. • There is only a significantly limited impact on latency, because the target node wakes up immediately as soon as it realizes that there is a pending message. • Unluckily, the adoption of a radio-triggered wakeup scheme is almost impractical, because it can be only applied when the distance between nodes is

4 Energy Management Techniques for WSNs …

166

extremely short (a few meters). Introducing an additional wakeup radio is a more promising direction, especially for event detection applications. Nevertheless, the wakeup radio is expensive, and generally it is not shipped with commonly used sensor platforms. So, when a second radio is not available or convenient, other techniques, such as the scheduled rendezvous and the asynchronous wakeup schemes, can be used. Both of them trade energy saving for an increased latency inflicted on messages traveling through many hops. The scheduled rendezvous scheme is convenient for its suitability to data aggregation and support of broadcast traffic. But, it requires nodes to be synchronized, which in some cases can be hard to achieve or expensive, in terms of additional protocol overhead for synchronization. On the other hand, asynchronous wakeup protocols have their own features: • They do not require a tight synchronization among network nodes. • They are commonly easier to implement and can ensure network connectivity even in highly dynamic scenarios where synchronous, i.e., scheduled rendezvous, schemes become inadequate. • This greater flexibility is compensated by lower energy efficiency; nodes need to wake up more frequently than in scheduled rendezvous protocols. Therefore, asynchronous protocols usually result in a higher duty-cycle for network nodes than their synchronous counterparts. Moreover, the support to broadcast traffic is problematic. Due to the wider applicability and the properties of scheduled rendezvous and asynchronous approaches, they might provide the most promising solutions in the class of sleep/wakeup protocols. However, there is still prospect for improvements over the techniques discussed in sections “Scheduled Rendezvous Schemes” and “Asynchronous Schemes”: • For instance, scheduled rendezvous protocols should relax the assumptions of clock synchronization among nodes, so that a coarse-grained time reference should be sufficient. Alternatively, they could embed a time synchronization solution as well, so that their timing requirements can be guaranteed without requiring a separate protocol. • While in the design of asynchronous protocols, exploiting cross-layer information seems to be an often forgotten factor.

4.1.2.2

MAC Protocols with Low Duty-Cycle

Several MAC protocols for WSNs have been proposed (Ye and Heidemann 2004; Demirkol et al. 2006). In the following sections, focus will be mainly on power management issues rather than on channel access methods. Most of these protocols adopt a low duty-cycle scheme for power management. MAC protocols are

4.1 Duty-Cycling Approach Taxonomy

167

classified according to the taxonomy illustrated in Fig. 4.8. Basically, they are TDMA-based, contention-based, and hybrid protocols: • Time-division multiple access (TDMA) protocols. They naturally enable a duty-cycle on sensor nodes as channel access is done on a slot-by-slot basis. As nodes need to turn ON their radio only during their own slots, the energy consumption is ideally reduced to the minimum required for transmitting/ receiving data. • Contention-based protocols. They are the most popular class of MAC protocols for WSNs. They achieve duty-cycling by tightly integrating channel access functionalities with a sleep/wakeup scheme similar to those described above. The sole difference is that in these protocols the sleep/wakeup algorithm is not a protocol independent of the MAC protocol, but is tightly coupled with it. • Hybrid protocols. They adapt the protocol behavior to the level of contention in the network. They behave as a contention-based protocol when the level of contention is low, and switch to a TDMA scheme when the level of contention is high. Most MAC protocols for WSNs have been based on conventional wireless protocols, especially the IEEE 802.11. These protocols typically provide a general-purpose mechanism that works reasonably well for a large set of traffic workloads. Several goals are usually considered in a MAC protocol for WSN applications, exclusively: • • • • • • •

Low-power operation. Effective collision avoidance. Simple implementation, small code, and RAM size. Efficient channel utilization at low and high data rates. Reconfigurable by network protocols. Tolerant to changing RF and networking conditions. Scalable to large numbers of nodes.

In sections “TDMA-Based MAC Protocols,” “Contention-Based MAC Protocols,” and “Hybrid MAC Protocols,” such MAC protocols are surveyed in proper depth. TDMA-Based MAC Protocols In TDMA-based MAC protocols (Heinzelman et al. 2000; Haartsen 2000; Arisha et al. 2002; Li and Lazarou 2004; Rajendran et al. 2006), time is divided into periodic frames where each frame consists of a certain number of time slots. Every node is assigned to one or more slots per frame, according to a certain scheduling algorithm, and uses such slots for transmitting/receiving packets to/from other nodes. In many cases, nodes are grouped to form clusters with a cluster head, which is in charge to assign slots to nodes in the cluster as in Bluetooth (Haartsen 2000), LEACH (Heinzelman et al. 2000), and energy-aware TDMA-based MAC (Arisha et al. 2002).

168

4 Energy Management Techniques for WSNs …

In sections “Traffic-Adaptive Medium Access Protocol (TRAMA),” “A Lightweight Medium Access Control (L-MAC) Protocol for WSNs,” and “Flow-Aware Medium Access (FLAMA),” TRAMA, L-MAC, and FLAMA are described and compared. Traffic-Adaptive Medium Access Protocol (TRAMA) TRAMA is introduced for energy-efficient collision-free channel access in WSNs (Rajendran et al. 2006)3. TRAMA reduces energy consumption by ensuring that unicast and broadcast transmissions incur no collisions, and by allowing nodes to assume a low-power, idle state, whenever they are not transmitting or receiving. TRAMA assumes that time is slotted and uses a distributed election scheme based on information about traffic at each node to determine which node can transmit at a particular time slot. Using traffic information, TRAMA avoids assigning time slots to nodes with no traffic to send and also allows nodes to determine when they can switch OFF to idle mode and not listen to the channel. TRAMA is revealed to be fair and correct, as no idle node is an intended receiver and no receiver suffers collisions. As for channel access in TRAMA, it is energy-efficient while maintaining good throughput, acceptable latencies, and fairness. Energy efficiency is attained by transmission schedules that avoid collisions of data packets at the receivers and by having nodes switch to low-power radio mode when there are no data packets intended for those nodes. Adequate throughput and fairness are achieved by means of an inherently fair transmitter-election algorithm that promotes channel reuse as a function of the competing traffic around any given source or receiver. TRAMA derives the collision-free transmission schedules based on more than a trait: • The identifiers of nodes that are one and two hops away. • The current time slot. • Traffic information that specifies which node intends to transmit to which other node. Hence, the “sleep schedule” of a node is a direct function of the traffic going through the node and its neighbors, and is synchronized automatically when nodes exchange information about their identifiers and their traffic. In contrast to prior MAC protocols proposed for WSNs, TRAMA provides support for unicast, broadcast, and multi-cast traffic to only a set of one-hop neighbors. TRAMA differs from sensor-MAC (S-MAC), shown in section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC),” which also provides explicit energy conservation mechanisms in two fundamental ways. First, TRAMA is inherently collision-free as its medium access control mechanism is schedule-based as opposed to the contention-based S-MAC.

3

First published in SenSys’03.

4.1 Duty-Cycling Approach Taxonomy

169

Signaling slots

Transmission slots

Signaling slots

Random-access

Scheduled-access

Random-access

Switching period

Fig. 4.23 TRAMA time organization (Rajendran et al. 2006)

Second, TRAMA uses an adaptive, dynamic approach based on current traffic patterns to switch nodes to low-power mode, while S-MAC scheme is static based on a predefined duty-cycle. A single, time-slotted, channel is assumed for both data and signaling transmissions. Figure 4.23 shows the overall time slot organization of the protocol. Time is organized as sections of random-access and scheduled-access periods. Random-access slots are signaling slots, and scheduled-access slots are transmission slots. Because the data rates of a sensor network are relatively low, the duration of time slots is much larger than typical clock drifts. For example, for a 115.2 Kbps radio, a transmission slot of approximately 46 ms is used to transmit 512-Byte application layer data units. Hence, clock drifts in the order of milliseconds can be tolerated; yet, typical clock drifts are in the order of microseconds or even less. This allows very simple time stamp mechanisms (Ganeriwal et al. 2003; Dai and Han 2004) to be used for node synchronization. When much smaller clock drifts must be assumed and more expensive nodes are to be used, nodes can be time synchronized using techniques such as GPS (SparkFun Electronics 2016). The length of a transmission slot is fixed based on the channel bandwidth and data size. Signaling packets are usually smaller than data packets, and thus transmission slots are typically set as a multiple of signaling slots to allow for easy synchronization. In TRAMA, transmission slots are seven times longer than signaling slots. Basically, TRAMA employs a traffic-adaptive distributed election scheme that selects receivers according to schedules announced by transmitters. Nodes using TRAMA exchange their two-hop neighborhood information, and the transmission schedules specifying in chronological order which nodes are the intended receivers of their traffic, and then select the nodes that should transmit and receive during each time slot. Accordingly, TRAMA is built upon three components: • The neighbor protocol (NP). NP propagates one-hop neighbor information among neighboring nodes during the random-access period using the signaling slots to obtain consistent two-hop topology information across all nodes. During the random-access period, nodes perform contention-based channel acquisition and thus signaling packets are prone to collisions. Transmission slots are used for collision-free data exchange and also for schedule propagation. • The schedule exchange protocol (SEP). SEP allows nodes to exchange two-hop neighbor information and their schedules. Essentially, schedules contain current information on traffic coming from a node, i.e., the set of receivers for the traffic

170

4 Energy Management Techniques for WSNs …

originating at the node. A node has to announce its schedule using SEP before starting actual transmissions. SEP maintains consistent schedule information across neighbors and updates the schedules periodically. • The adaptive election algorithm (AEA). AEA uses neighborhood and schedule information to select the transmitters and receivers for the current time slot, leaving all other nodes in liberty to switch to low-power mode. AEA selects transmitters and receivers to achieve collision-free transmission using the information obtained from NP and SEP. This is because electing both the transmitter and the receiver(s) for a particular time slot is a necessity to achieve energy efficiency in a collision-free transmission schedule. Random transmitter selection leads to collisions, and electing the transmitters and not the receivers for a given time slot leads to energy waste, since all the neighbors around a selected transmitter have to listen in the slot, even if they are not to receive any data. Furthermore, selecting a transmitter without regard to its traffic leads to low channel utilization, as the selected transmitter may not have any data to send to the selected receiver. Hence, to improve channel utilization, AEA uses traffic information, i.e., which sender has traffic to which receivers. Knowingly, WSNs are dynamic, nodes may fail if power drained, or new nodes may be added when additional sensors are deployed. To accommodate topology dynamics, TRAMA alternates between random-access and scheduled-access. TRAMA starts in random-access mode where each node transmits by selecting a slot randomly. Nodes can only join the network during random-access periods. The duty-cycle of random-access versus scheduled-access depends on the type of network: • In more dynamic networks, random-access periods should occur more often. • In more static scenarios, the interval between random-access periods could be larger, because topology changes need to be accommodated only occasionally. • In the case of WSNs, there is very little or no mobility, depending on the type of application. Hence, the main function of random-access periods is to permit node additions and deletions; time synchronization could also be done during this period. During random-access periods, all nodes must be in either transmit or receive state, so they can send out their neighborhood updates and receive updates from neighbors. Thus, the duration of the random-access period plays a significant role in energy consumption. During random-access periods, signaling packets may be lost due to collisions, which can lead to inconsistent neighborhood information across nodes. To guarantee consistent neighborhood information with some degree of confidence, the length of the random-access period and the number of retransmissions of signaling packets are set accordingly. In Rajendran et al. (2006), it is shown that, for a network with an average of N two-hop neighbors, the number of signaling packet retransmissions should be 7 and the retransmission interval should be 1.44  N to guarantee 99% packet delivery. Hence, the length of the random-access period will be 7  1.44  N.

4.1 Duty-Cycling Approach Taxonomy

171

Several details are to be enlisted to further clarify TRAMA functioning: • NP gathers neighborhood information by exchanging small signaling packets during the random-access period. Signaling packets carry incremental neighborhood updates, while if there are no updates, signaling packets are sent as “keep-alive” beacons. Each node sends incremental updates about its one-hop neighborhood as a set of added and deleted neighbors. These signaling packets are also used to maintain connectivity between the neighbors. A node times out a neighbor if it does not hear from that neighbor for a certain period of time. The updates are retransmitted such that a 0.99 probability of success is ensured. Because a node knows the one-hop neighbors via its one-hop neighbors, consistent two-hop neighborhood information eventually makes its way across the network. • SEP establishes and maintains traffic-based schedule information required for the transmitter and receiver selection, such as slot reuse for the transmitter, and sleep state switching for the receiver. A node schedule captures a window of traffic to be transmitted by the node. This information is periodically broadcasted to the node one-hop neighbors during scheduled-access. To generate a schedule, each node computes a SCHEDULE_INTERVAL based on the rate at which packets are produced by the typical application. The SCHEDULE_ INTERVAL of a node represents the number of slots where the node can announce the schedule to its neighbors according to the current state of its MAC layer queue. The node then precomputes the number of “winning slots” in the interval [t, t + SCHEDULE_INTERVAL] where it has the highest priority among its two-hop neighbors (contenders). Because during these slots, the node will be selected as the transmitter, it announces the intended receivers for such slots. Alternatively, if a node does not have enough packets to transmit, it announces that it gives up the corresponding slot(s). Other nodes that have data to transmit can make use of these “vacant” slots. A node’s last winning slot in this interval is reserved for broadcasting the node schedule for the next interval. • Nodes announce their schedule via schedule packets. Because nodes have two-hop topology information obtained through NP, there is no need to send receiver addresses in the schedule packet. Instead, nodes convey intended receiver information using a bitmap whose length is equal to the number of one-hop neighbors. A summary of a node schedule is also sent with every data packet. Schedule summaries help minimizing the effects of packet loss in schedule dissemination. Each schedule has an associated time-out, and nodes are not allowed to change the schedule until it expires; this is to ensure consistency across one-hop neighborhood schedules. • Nodes maintain schedule information for all their one-hop neighbors. The schedule information is consulted whenever a node has the highest two-hop priority that permits to decide if it will actually transmit the data it has and will thus use the slot, or it will give up the slot to another node in the neighborhood. An analytical model was developed, and its results were verified by simulation using the QualNet network simulator (Fahmy 2016). The underlying physical layer

172

4 Energy Management Techniques for WSNs …

model used for the experiments was based on the TR1000 (Murata Electronics 2015b), a typical radio used in WSNs. The TR1000 radio used by the UC Berkeley Motes (Crossbow 2002; Hill and Culler 2002) is a short-range, low data rate at a maximum of 115.2 Kbps with built-in support for low-power sleep state. The average power consumption in transmit, receive, and sleep modes is 24.75 mW, 13.5 mW, and 15 lW, respectively. The maximum transition time for switching is 20 ls. The modulation type used in the physical layer is amplitude shift keying (ASK), while the receiver threshold is −75 dBm. The topology considers 50 nodes uniformly distributed over a 500 m  500 m area. The transmission range of each node is 100 m, and the topology is such that the nodes have 6 one-hop neighbors on average. The average size of the two-hop neighborhood for this network is 17 nodes. Two different types of traffic load were considered. In a scenario, node traffic is statistically generated based on exponentially distributed interarrival time to stress test protocol performance for different arrival rates. TRAMA performance was also tested when driven by data gathering applications, which are considered typical of WSNs. The analytical model quantified the delay for scheduling-based MACs. It is shown that TRAMA is well suited for applications that are not delay sensitive while requiring high delivery guarantees and energy efficiency. Furthermore, through extensive simulations, TRAMA performance was compared against a number of contention-based and scheduled-based MAC protocols. From the simulation results, it was found that: • Significant energy savings can be evidently achieved by TRAMA depending on the offered load, since nodes can sleep for up to 87% of the time. • TRAMA also realizes a higher throughput when compared to contention-based protocols. Gains obtained are around 40% over S-MAC and CSMA, and around 20% over IEEE 802.11, because TRAMA avoids collisions due to hidden terminals. • In general, scheduled-based MACs exhibit higher delays than contention-based MACs. In the case of TRAMA, the delay is higher than random selection protocols, e.g., NAMA (Bao and Garcia-Luna-Aceves 2001), due to the scheduling overhead. Recognizing its energy efficiency benefits, the use of traffic information also makes TRAMA adaptive to the WSN application at hand. However, TRAMA adaptiveness comes at a price, namely the complexity of its election algorithm and scheduling overhead for announcing traffic information. A Lightweight Medium Access Control (L-MAC) Protocol for WSNs The L-MAC protocol is designed to account for the physical layer properties in WSN nodes (van Hoesel and Havinga 2004). The goal of the protocol is minimizing the number of transceiver switches, to make the sleep interval for sensor nodes adaptive to the amount of data traffic and to limit the complexity of

4.1 Duty-Cycling Approach Taxonomy

173

implementation. L-MAC is based on the EMACS protocol (Nieberg et al. 2003) designed in the European research project EYES (Havinga et al. 2003). In MAC protocols that are designed for WSNs, the properties of the underlying hardware are not to be bypassed. In general, the simple and inexpensive receivers envisioned in the sensor nodes that form the WSN need to be trained to the incoming RF signal before an acceptable bit error rate (BER) can be established. The training is often done by transmitting a known sequence to the receiver, allowing it to adjust its input sensitivity and adopt its timing synchronization to that of the transmitter. This preamble adds significantly to the energy costs of transmitting a message. From a technical standpoint, transceivers suffer from startup effects. Transceivers that use crystals to derive their mixer frequencies typically switch OFF their oscillators when they are put in low-power state. To re-enable transmit or receive functions, the crystal oscillator has to be restarted, which takes time and consumes a usually undefined amount of energy. Frequent transceiver state switches can drastically reduce the lifetime of the network. The L-MAC inspiring, TDMA-based EMACS protocol divides time into slots that nodes can use to transfer data without having to contend for the medium or to deal with energy-wasting transmission collisions. A node can assign only one slot to itself and controls it. After the frame length, which consists of several time slots, the node again has a reserved period of time. A time slot is further divided into three sections: • Communication request (CR). In the CR section, other nodes can send requests to the node that is controlling the current time slot. Nodes that have a request will pick a random start time in the short CR section to make their request. These messages are comparable to RTS messages in S-MAC. Communication in this section is not guaranteed to be collision-free. Nodes that do not have a request for the current slot owner will keep their transceiver in a low-power state during the entire CR section. • Traffic control (TC). During a time slot, its controller will always transmit a TC message. The time slot controller indicates in its TC message what communication will take place in the data section. If a node is not addressed in the TC section nor its request was approved, then the node will resume in standby state during the entire data section. The TC message can also indicate that the controlling node is about to send an omnicast message. During a time slot that is not controlled by any node, all nodes will remain in sleep state. • Data section. After the TC section, the actual data transfer takes place. Energy efficiency is a main challenge when designing a MAC protocol for WSNs. Sensor nodes need saving on every bit that is transmitted, to ensure an acceptable network lifetime, while limiting latency and throughput degradation. Sensors equipped with transceiver, processor, and memory will be deployed by thousands or their multiples; hence, the costs of a single smart sensor must be at a minimum. This does not translate to just limiting resources, but also to lessening complexity of the hardware. Multi-channel transceivers, for instance, are higher priced than

4 Energy Management Techniques for WSNs …

174 Table 4.4 Transceiver TR1001 data (van Hoesel and Havinga 2004)

Parameter

Value

Energy consumption for Tx Energy consumption for Rx Energy consumption for sleep Switch time sleep/Tx Switch time sleep/Rx

21 mW 14.4 mW 15 lW 16 ls 518 ls

single-channel versions. In L-MAC, a single-channel transceiver with three operational states is chosen, namely transmit, receive, and standby. As depicted in Table 4.4 for the TR1001 transceiver (Murata Electronics 2014), transmitting consumes more power than receiving, while standby lies beneath the power consumption of receiving by a factor of 1000 or beyond. These parameters are used in the physical layer model of the simulator to study network lifetime. Based on time-division multiple access (TDMA), L-MAC considers several basics: • Time is divided into time slots that nodes can obtain to transfer data without having to contend for the medium, or to deal with energy-wasting transmission collisions. • Only one time slot is assigned to each node. • Unlike traditional TDMA-based systems, the time slots in L-MAC are not divided among the networking nodes by a central manager; instead, a distributed algorithm is used (Nieberg et al. 2003). • During its time slot, a node will always transmit a message consisting of two parts: control message and a data unit. Because a time slot can only be controlled by a single node, this node can communicate collision-free. • The control message, listed in Table 4.5, has a fixed size and is used for several purposes: – It carries the ID of the time slot controller. – It indicates the distance of the node to the gateway in hops for simple routing to a network gateway. – It addresses the intended receiver and reports the length of the data unit.

Table 4.5 Control message contents (van Hoesel and Havinga 2004)

Description

Size (Bytes)

Identification Current slot number Occupied slots Distance to gateway Collision in slot Destination ID Data size Total

2 1 4 1 1 2 1 12

4.1 Duty-Cycling Approach Taxonomy

• • • •

175

The control data will also be used to maintain synchronization between the nodes; therefore, the nodes also transmit the sequence number of their time slot in the frame. The transmission of the control data is carefully timed by the nodes, although it is not assumed that nodes have clocks with high accuracy, and the clock drift is negligible in a single frame, even for clocks with low accuracy. When a receiving neighboring node is not addressed in that control message, or the message is not addressed as an omnicast message, the node will switch OFF its power-consuming transceiver only to wake up at the next time slot. If a node is addressed, it will listen to the data unit, which might not fill the entire remainder of the time slot. Both transmitter and receiver(s) turn OFF their transceivers after the message transfer has completed. A short timeout interval ensures that nodes do not waste energy for idle listening in time slots that are not controlled. It is only possible for a node to transmit a single message per frame. The maximum size of the data unit is 256 Bytes, but this value can easily be adapted to the expected WSN traffic. A node may glue messages together to the same destination to prevent high latency. To set up the network, several steps are to be followed:

• When the nodes are powered ON, they are all unsynchronized. In order to get synchronized, the gateway will take the initiative to start controlling a time slot. The control messages of the gateway are to be received by its one-hop neighbors. These neighbors will synchronize their clocks to the gateway. After one frame, the one-hop neighbors are aware of all time slots that are owned by possible gateways in their reception range. • Next, the recently synchronized nodes pick a random time slot, that is not occupied, to control. • The time slot occupancy is efficiently encoded by a number of bits equal to the number of time slots in a frame. Nodes can start controlling a time slot when the slot is considered to be free by all its neighbors and therefore all nodes are required to maintain a table of their neighborhood. This method ensures that a time slot is only reused after at least three hops and that no collisions of messages will occur. In practice, the CTS message in sensor-MAC (S-MAC) takes care of a similar distance between two transmissions at the same time. • Because there is a chance at network setup that nodes will pick the same time slot to control, nodes inform their neighborhood when a collision occurs between control messages. The nodes that did transmit the colliding control messages will give up their time slot and will choose again a random, not yet controlled, time slot after a backoff time depending on the identification number (ID) of the node. Clearly, such ID-dependent backoff time ensures that all nodes are capable of controlling a time slot, and that the nodes in the network can communicate with each other collision-free.

176

4 Energy Management Techniques for WSNs …

• Nodes will maintain their time slots until their battery runs out or if they are informed that their time slot is colliding with another one. • The number of time slots must be larger than the maximum connectivity in the network. This ensures that every node in the network can find an empty slot in finite time. In L-MAC design, there are 32 time slots in a frame. To route through the gateway, each node keeps track of its hop distance to a designated gateway node and broadcasts this information in its control message. When a message arrives, either generated by the node itself or received from another node, the node looks up in its neighbor table for a neighboring node that is closer to the gateway than itself, and picks this node as destination for the message. In case of multiple neighbors closer to the gateway, the node randomly picks one from the candidates. Eventually, the message will arrive at the gateway. Simulation study was done thru the discrete event simulator OMNeT++ (Fahmy 2016) with implementations of EMACS, S-MAC, and L-MAC, together with a framework for WSNs (Dulman and Havinga 2003). In the simulator, a physical layer with energy model is implemented to record the sending, receiving, and standby energy consumptions of the nodes. Additionally, care is accorded to the energy consumed in switching between sending and receiving for the TR1001 transceiver (Table 4.4). Compared to EMACS and S-MAC, L-MAC was able to extend the network lifetime by factors of 2.4 and 3.8, respectively. Flow-Aware Medium Access (FLAMA) FLAMA is an energy-efficient medium access control (MAC) protocol designed for WSNs (Rajendran et al. 2005). FLAMA achieves energy efficiency by preventing idle listening, data collisions, and transmissions to a node that is not ready to receive packets; it adapts medium access schedules to the traffic flows exhibited by the application. FLAMA is simple enough so that it can be run by nodes with limited processing, memory, communication, and power capabilities. The performance of FLAMA is evaluated through simulations and testbed experimentation. By then, most existing energy-efficient MAC protocols for WSNs have employed a contention-based approach; the sensor-MAC (S-MAC) protocol is a typical instance. The main drawback of contention-based MAC protocols is that the probability of collisions increases with the offered load, which degrades channel utilization and wastes energy. This motivated research into distributed schedule-based medium access methods. FLAMA is a schedule-based MAC protocol that leverages traffic predictability in WSN applications. Traffic information can be determined by having the application explicitly specify its traffic characteristics or by using traffic prediction techniques at each node. Depending on the application at hand, traffic prediction can be relatively simple. Explicitly, in environmental monitoring, periodic data gathering generates data streams over a collection tree rooted at the information sink and spanning all relevant nodes. When sending data, each node transmits

4.1 Duty-Cycling Approach Taxonomy

177

upstream to the next hop toward the sink; this information could be used to determine the next-hop node for a node transmission. FLAMA uses the concept of flows to characterize application traffic patterns. Flows represent one-hop traffic information and specify the transmitter, the receiver (s), and the rate at which packets are sent. Flow-based traffic information helps determining transmission schedules, as well as when nodes should be in receive mode or can switch to low-power sleep state. Unlike TRAMA (section “Traffic-Adaptive Medium Access (TRAMA) Protocol”) that attempts to realize adaptive scheduling in WSNs, FLAMA does not require explicit schedule announcements during scheduled-access periods. Alternatively, application-specific traffic information is exchanged among nodes during random-access in order to reflect the driving application-specific traffic patterns or flows. This allows FLAMA to still adapt to changes in traffic behavior and topology, e.g., node failure. FLAMA is characterized by main features that are: • The distributed maintenance of energy-efficient, collision-free transmission schedules based on two-hop neighborhood information and implicit traffic information. • Low transmission delays with limited processing and storage requirements. • Robust operation that accommodates topology changes. FLAMA uses a simple traffic-adaptive distributed election scheme for energy-efficient channel access. It requires two-hop neighborhood and flow information in the neighborhood to perform the election. Using only two-hop neighborhood information makes FLAMA scalable. Time is organized in periods of random-access and scheduled-access intervals as shown in Fig. 4.24. A single channel is assumed for data and signaling; however, FLAMA can be easily extended to handle multiple channels. Channel access is contention-based during random-access and time-slotted during scheduled-access periods. During random-access, neighbor discovery, time synchronization, and implicit traffic information exchange are performed. Data transmission happens during scheduled-access. Using periodic random-access periods allows FLAMA to adapt to topology and traffic changes in the network. FLAMA design considerations and implementation issues are enlisted in the coming notions:

Signaling transmission

Data transmission slots

Signaling transmission

Random-access period (contention-based channel access)

Scheduled-access period (time-slotted channel access)

Random-access period (contention-based channel access)

Fig. 4.24 FLAMA time organization (Rajendran et al. 2005)

178

4 Energy Management Techniques for WSNs …

• The implementation of FLAMA is customized for data gathering applications, an important class of WSN applications. In data gathering applications, it is assumed that the sink initially sends out a query requesting data from sensing nodes. As the replies from the sensors are forwarded back, a tree rooted at the sink spanning all relevant nodes is established. Then, sensor nodes sample readings periodically and send them to the sink over the collection tree. On its way to the sink, data might be aggregated (Intanagonwiwat et al. 2002) to minimize energy consumption. The sink is used as the synchronization point for the other nodes; e.g., the sink may be connected to a backbone network and is synchronized to it. For this type of data gathering scenarios, traffic is predictable and exhibits regular patterns, which can be exploited when designing MAC layer protocols tailored for these applications. Since data are sent back to the sink along the forwarding tree, nodes can easily determine incoming and outgoing flows. More specifically, a node has incoming flows from all its children in the tree and it has only one outgoing flow to its parent. The sink does not have any outgoing flows. If R is the rate at which sensed data are generated at sensor nodes, all the nodes in the network except for the sink have an originating flow with data rate R that exists for a period specified by the sink. A node has to either forward or aggregate flows that are incoming from its children. If no data aggregation is employed, the outgoing flow rate is the sum of the incoming flow rates from the children and the originating flow rate R. If data aggregation is used, then the outgoing flow rate remains constant at R. FLAMA assigns node weights based on the resulting flow rates and performs traffic-adaptive scheduling. How FLAMA acquires this information during the random-access period is described in the coming bullet. • Random-access is used for time synchronization, exchanging neighbor information, and establishing flow information. In the case of data gathering applications, establishing flow information is essentially constructing the data forwarding tree. The data gathering node, or sink, initiates tree formation and time synchronization. Every node in the tree synchronizes with its parent using a pairwise time synchronization algorithm based on time stamps. Hence, during the random-access period the following tasks that are necessary for FLAMA operation are performed: – Network-wide time synchronization. – Data forwarding tree formation. – Traffic flow information exchange and weight computation for traffic-adaptive election. – Two-hop neighborhood information and corresponding node weight exchange. Nodes running FLAMA start in random-access mode, and the radio is in either transmit or receive state. Control frames are exchanged every SYNC_INTERVAL. Two types of control frames, SYNC and SYNC_REQ, are

4.1 Duty-Cycling Approach Taxonomy

179

exchanged during random-access, while channel access is based on carrier sensing. Other than the source and destination information, the control frame also includes the node outgoing flow weight, the node parent, time stamp, and a neighbor update list. The neighbor update list contains node identifiers for one-hop neighbors, their announced weights, and receive time stamps. In the case of data gathering applications, each node has only one outgoing flow toward the parent; hence, it suffices to announce a single weight for the node. Other applications might need to announce multiple node weights based on the number of outgoing flows. FLAMA requires time synchronization between two-hop neighbors (Ganeriwal et al. 2003; Dai and Han 2004). The basic idea is time-stamping the packet at the lowest possible level and using these time stamps to calculate clock drifts. A sender-initiated time synchronization mechanism is used such that a node can send a SYNC frame only after synchronizing the clock with its parent. Otherwise, nodes send SYNC_REQ frames to discover parents. The sender, or the parent, initiates time synchronization by sending a SYNC frame with its local time stamp T1. The receiver acquires the frame at its local time T2. Clearly, T2 = T1 + d + s, where d is the clock drift and s is the propagation delay. The receiver replies with SYNC_REQ to the parent with its local time stamp T3, and the sender receives the packet at its local time T4. Then, T4 = T3 − d + s. Using T1, T2, T3, and T4, both d and s can be calculated. As the receiver has to adjust its clock based on the sender, the sender sends back a SYNC frame announcing the time stamp T4 to the receiver. The receiver computes the clock drift d thru Eq. 4.20 and adjusts its clock: d ¼ ðT2  T1 þ T3  T4Þ=2

ð4:20Þ

Once a node becomes synchronized with its parent, it can start sending SYNC frames and synchronize downstream nodes. This process eventually synchronizes the entire network. Time stamps are generated at the physical layer to improve the accuracy. A node updates its child information whenever it receives SYNC frames with its node identifier as the parent. The length of the random-access period is fixed based on the time required to complete the synchronization and tree formation processes. During random-access periods, signaling packets may be lost due to collisions. Hence, the interval should be long enough to accommodate signaling retransmissions. In FLAMA implementation, exchanged state information is minimized and kept by nodes. Each node maintains the following neighborhood information: – Parent identifier (2 Bytes). – Clock drift information (T1, T2, T3, T4, and offset, for a total of 20 Bytes). – One-hop neighbor table where each entry has 8 Bytes of information, namely node identifier, isChild flag, receive time stamp, node weight.

4 Energy Management Techniques for WSNs …

180

– Two-hop neighbor table, namely node identifier and node weight, with each entry having 3 Bytes of information. FLAMA uses node weights to adjust transmission schedules based on how much traffic individual nodes generate. Node weight calculation is illustrated using the example shown in Fig. 4.25, where arrows represent traffic flows with rates. For example, node B has three incoming flows, from nodes C, D, and E with rates FC, FD, and FE, respectively, and a single outgoing flow to node A with a rate FB. The outgoing flow rate FB is a function of incoming flow rates and is given by: FB ¼ Forigin þ c  FC þ d  FD þ e  FE

ð4:21Þ

where c, d, and e denote the fraction of the flow that is forwarded. If the flows are “terminal flows,” then c, d, and e are 0. Forigin denotes the rate of the originating flow (if any) from node B. Node weights are directly proportional to the outgoing flow rate. Hence, node B’s weight is decided based on FB and is announced during random-access. • During scheduled-access, channel access is time-slotted. The slot interval is fixed based on a maximum physical layer frame size. The packet size is 128 Bytes, which is the maximum physical layer packet size for TinyOS CC1000 physical radio module (Texas Instruments 2007). A guard interval is added to the time slot duration to account for synchronization errors and radio mode switching, and is set to a multiple of the maximum possible clock drift. The number of slots in the scheduled-access period is decided based on the duty-cycle for scheduled-access. The distributed election algorithm is used to decide the state of each node at every slot. The distributed election algorithm schedules collision-free transmissions; its design is driven by the fact that sensing nodes are typically limited in processing and memory resources. Essentially, for each node, the election algorithm decides which radio mode to use in the current slot; the choices are transmit,

Fig. 4.25 Traffic flows (Rajendran et al. 2005)

C D

FC FD B FB A

FE

E

4.1 Duty-Cycling Approach Taxonomy

181

receive, or sleep. FLAMA ensures that there is only one transmitter in the two-hop neighborhood and thus avoids hidden terminal collisions. The election algorithm requires that each node maintains a list of one-hop and two-hop neighbors and their corresponding weights, and parent information. A node can transmit if it has the highest two-hop priority for the given time slot and it has data to send. A node should be in receive mode if it is not the highest two-hop priority node and its highest one-hop priority node is a child. Otherwise, a node can go to sleep. While in receive mode waiting for data, the node can switch to sleep mode if it does not start receiving data for PREAMBLE_INTERVAL. Node weights computed during the random-access period are incorporated into the election algorithm to provide more channel access for nodes with higher traffic rate. This makes FLAMA traffic-adaptive while maintaining the simplicity of the election algorithm. Collision freedom is obtained by allowing only one transmitter in the two-hop neighborhood. Due to limited neighborhood information and the distributed nature of the algorithm, care should be taken to prevent a node from sleeping when a neighbor is transmitting a data packet destined to this node. FLAMA performance is evaluated by both simulation and testbed experimentation. The main goal of the simulation experiments is to highlight the importance of application awareness in channel access scheduling. To be reminded, TRAMA (section “Traffic-Adaptive Medium Access (TRAMA) Protocol”) is designed for general applications and hence has to propagate traffic information explicitly and periodically. FLAMA, on the other hand, establishes flows based on traffic patterns exhibited by WSN applications and does not need to propagate traffic information explicitly. S-MAC is also designed for general applications and does not account for application-specific traffic patterns. Testbed experiments target establishing the feasibility of implementing a TDMA-based, application-aware MAC protocols on sensor nodes and also to check the advantages of FLAMA over traditional contention-based channel access protocols. Several metrics are used to assess performance comparisons of the protocols: • Average packet delivery ratio. It is the ratio of the number of packets received at the sink to the number of packets sent by all sensor nodes. For broadcast traffic, a packet is counted to be received only if it is received by all one-hop neighbors. • Percentage sleep time. It is the ratio of the time spent in low-power sleep mode to the total experiment run time. • Latency is computed as the average per-hop latency for the network. • Average queue drops provide the average number of packets dropped at the MAC layer queue. Simulation using QualNet (Fahmy 2016) exhibits interesting results: • FLAMA achieves better delivery ratio than TRAMA and S-MAC. This is due to the fact that FLAMA performs traffic-adaptive scheduling without incurring much overhead. Nodes near the sink have a larger outgoing flow rate and are

182

• •

• •

•

4 Energy Management Techniques for WSNs …

preferred in the election process. Whereas, TRAMA needs to propagate traffic information periodically which results in a significant overhead during scheduled-access period. Hence, for the given simulation duration, FLAMA was able to service more packets than TRAMA. The synchronized listen and sleep cycles of S-MAC affect neighbor discovery and data throughput in multihop forwarding. This is because S-MAC restricts transmitting or receiving packets to a specific small time window. In S-MAC, depending on the contention window size for transmitting data and synchronization packets, collisions occur due to hidden terminals. This affects neighbor discovery significantly as the synchronization packets are sent by unreliable broadcasts. Hence, the average delivery ratio at the sink is significantly less for S-MAC when compared with scheduling-based protocols. The queueing delay for FLAMA is significantly 75 times less than that of TRAMA. S-MAC achieves less delay than FLAMA in the studied topology. This is due to the delay involved in the election algorithm, which depends on the two-hop neighborhood size. However, FLAMA achieves much higher reliability than S-MAC. Hence, the end-to-end application-perceived delay is much higher for S-MAC due to retransmissions. FLAMA achieves significant energy savings when compared to TRAMA and S-MAC. This is because FLAMA exchanges less information than TRAMA during scheduled-access periods. As expected, for both FLAMA and TRAMA the energy savings are proportional to the offered load. For S-MAC, energy savings depend on the fixed duty-cycle.

In testbed experimentation, FLAMA was implemented on TinyOS for MICA2 motes and its performance was compared with that of S-MAC. Similar to S-MAC, FLAMA is implemented on top of the radio communication stack developed at USC/ISI and UCLA for the MICA2 platform (Ye et al. 2002a). For both S-MAC and FLAMA, there is no MAC layer buffer to queue up frames; hence, a frame from the application is dropped if the send buffer is full. From experimentation, it was found that: • FLAMA significantly outperforms S-MAC in terms of delivery ratio, drop rate, and energy efficiency. • For the adopted topology, the average service time for S-MAC is in the order of 700 ms, while for FLAMA the service time is around 100 ms. Hence, the number of packets dropped for S-MAC is significantly higher than that of FLAMA. This affects the end-to-end reliability measured at the sink. • FLAMA delay is dependent on the number of two-hop nodes. • FLAMA achieves energy savings comparable to S-MAC, and this is in spite of the fact that FLAMA has the radio ON during the entire random-access period. This clearly demonstrates the importance of using an adaptive scheduling approach for channel access in sensor networks.

4.1 Duty-Cycling Approach Taxonomy

183

For low data rates, it was realized that: • FLAMA achieves perfect reliability while S-MAC reliability is 75%; this is because FLAMA avoids collision and transmissions to sleeping node. Also, it does not exchange any control packets during the scheduled-access period and hence the channel contention level is lower. • On the other hand, even though S-MAC uses RTS/CTS handshakes to avoid hidden terminal collisions, it loses data packets due to increased service time and also due to RTS/CTS handshake failures. • Low offered loads tend to benefit contention-based protocols. Contention-Based MAC Protocols In all shared-medium networks, medium access control (MAC) is an important technique that enables the successful operation of the network. One fundamental task of a MAC protocol is to avoid collisions from interfering nodes. There are many MAC protocols that have been developed for wireless voice and data communication networks. Typical examples include the time-division multiple access (TDMA), code-division multiple access (CDMA), and contention-based protocols like IEEE 802.11. When considering MAC protocols, the major sources of energy waste have been identified to be: • Collision. When a transmitted packet is corrupted, it has to be discarded; hence, follow-on retransmissions increase energy consumption. Collision increases latency as well. • Overhearing, meaning that a node picks up packets that are destined to other nodes. • Control packet overhead. Understandably, sending and receiving control packets consume energy. • Idle listening, i.e., listening to receive possible traffic that is not sent. This is especially true in many WSN applications. If nothing is sensed, nodes are in idle mode for most of the time. However, in many MAC protocols such as IEEE 802.11 ad hoc mode or CDMA, nodes have to listen to the channel to receive possible traffic. Measurements have shown that idle listening consumes 50– 100% of the energy required for receiving (Ye et al. 2004). Most sensor networks are designed to operate for long time, and nodes will be in idle state for a long time. Thus, idle listening is a dominant factor of energy waste in such cases. Sections “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC),” “An Adaptive Energy-Efficient MAC Protocol for WSNs (T-MAC),” “An Adaptive Energy-Efficient and Low-Latency MAC for Data Gathering in WSNs (D-MAC),” and “Versatile Low-Power Media Access for Sensor Networks (B-MAC),” respectively, analyze and compare S-MAC, T-MAC, D-MAC, and B-MAC.

184

4 Energy Management Techniques for WSNs …

Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC) Sensor-MAC (S-MAC) is a MAC protocol explicitly designed for WSNs (Ye et al. 2004). Reducing energy consumption is the primary goal of S-MAC; it also achieves good scalability and collision avoidance by utilizing a combined scheduling and contention scheme. To achieve the primary goal of energy efficiency, the main sources that cause inefficient use of energy were identified, as well as the tradeoffs that can be made to reduce energy consumption. The effectiveness and performance of S-MAC were tested on a testbed built on Atmel ATmega128L 8-bit microcontroller (Atmel 2011). The mote runs on TinyOS, the very small event-driven operating system (TinyOS 2012). S-MAC targets to reduce energy waste from all causes mentioned in section “Contention-Based MAC Protocols”: specifically, idle listening, collision, overhearing, and control overhead. In exchange, it tolerates some performance reduction in both per-hop fairnessand latency. S-MAC establishes low duty-cycle operation on nodes in a multihop network; it reduces idle listening by periodically putting nodes into sleep state. While in sleep state, the radio is completely turned OFF. In protocols for traditional data networks like the IEEE 802.11, bandwidth utilization is a big concern, and nodes normally operate in fully active mode (Xiao and Rosdahl 2002; Cooklev 2004). Switching to low duty-cycle mode (called power saving mode) is an option at each IEEE 802.11 node, and it normally happens when a node has been idle for long time. In S-MAC, however, the low duty-cycle mode is the default operation of all nodes; nodes only become more active when there is traffic in the network. To reduce control overhead and latency, S-MAC introduces coordinated sleeping among neighboring nodes. Also, S-MAC recognizes that how is latency meaningful depends on what application is running. In applications such as surveillance or monitoring, nodes will be vigilant for long time, but largely inactive until something is detected. These applications can often tolerate some additional messaging latency, because the network speed is typically orders of magnitude faster than the speed of a physical object. During a period where there is no sensing event, there is normally very little data flowing in the network. In a fraction of seconds, latency is not critical and can be traded off for energy savings. S-MAC, therefore, lets nodes periodically sleep if otherwise they will be idle. This design reduces energy consumption, but increases latency, since a sender must wait for the receiver to wake up before it can send out data. S-MAC introduces adaptive listen to meaningfully reduce such latency. Message passing is one of the S-MAC concerns. In traditional wireless voice or data networks, each user desires equal opportunity and time for accessing the medium to send or receive packets for his applications; per-hop MAC-level fairness is thus critical. However, by WSN nature, all nodes cooperate for a single common task; at any particular time, one node may have dramatically more data to send than some other nodes. Hence, fairness is not important as long as application-level performance is not degraded. S-MAC reinstates the concept of message passing to efficiently transmit long messages. A long message is divided into small fragments

4.1 Duty-Cycling Approach Taxonomy

185

that are transmitted in a burst. Accordingly, a node that has more data to send gets more time to access the medium. From a per-hop MAC-level perspective, this is unfair for nodes that have some short packets to send. However, message passing saves energy by reducing control overhead and avoiding overhearing. In-network data processing characterizes WSNs; it greatly reduces energy consumption compared to transmitting all the raw data to the end node (Intanagonwiwat et al. 2002). Techniques such as data aggregation can reduce traffic, while collaborative signal processing can reduce traffic and improve sensing quality. In-network data processing requires store-and-forward processing of messages. A message is a unit of data that a node can process (average, filter, etc.), and it may be long and composed of many small fragments. Thus, MAC protocols that promote fragment-level fairness actually increase message-level latency. In contrast, message passing reduces message-level latency by trading off the fragment-level fairness. From the above discussion, S-MAC targets reducing energy consumption from all sources of energy waste, i.e., idle listening, collision, overhearing, and control overhead. Thus, two basic concepts are satisfied, periodic listen and sleep, and collision avoidance: • Periodic listen and sleep. In many WSN applications, nodes are idle for long time if no sensing event happens. Given the fact that the data rate is very low during this period, it is not necessary to keep nodes listening all the time. S-MAC reduces the listen time by putting nodes into periodic sleep state. The scheme shown in Fig. 4.26 holds several essentials: – Each node sleeps for some time and then wakes up and listens to see if any other node wants to talk to it. During sleeping, the node turns OFF its radio and sets a timer to be awakened later. – A complete cycle of listen and sleep is a frame. The listen interval is normally fixed according to physical layer and MAC layer parameters, e.g., the radio bandwidth and the contention window size. The sleep interval can be changed according to different application requirements, which actually changes the duty-cycle. For simplicity, these values are the same for all nodes. – The duty-cycle is defined as the ratio of the listen interval to the frame length. – All nodes are free to choose their own listen/sleep schedules. However, to reduce control overhead, neighboring nodes synchronize together. That is, they listen at the same time and go to sleep at the same time. It should be noticed that not all neighboring nodes could synchronize together in a multihop network. Two neighboring nodes A and B may have different schedules if they must synchronize with different nodes, C and D, respectively, as shown in Fig. 4.27.

4 Energy Management Techniques for WSNs …

186 …

Listen

Sleep

Listen

Sleep

… Time

Fig. 4.26 Periodic listen and sleep (Ye et al. 2004)

Fig. 4.27 Node synchronization (Ye et al. 2004)

A

B

C

D

– Nodes exchange their schedules by periodically broadcasting a SYNC packet to their immediate neighbors. A node talks to its neighbors at their scheduled listen time, thus ensuring that all neighboring nodes can communicate even if they have different schedules. In Fig. 4.27, for example, if node A wants to talk to node B, it waits until B is listening. The period for a node to send a SYNC packet is called the synchronization period. A characteristic of S-MAC is that it forms nodes into a flat, peer-to-peer topology. Unlike clustering protocols, S-MAC does not require coordination through cluster heads. One advantage of this loose coordination is that it can be more robust to topology change than cluster-based approaches. The downside of the scheme is the increased latency due to the periodic sleeping. Furthermore, the delay can accumulate on each hop. To reduce such latency, as shown below, coordinated sleeping is performed through choosing and maintaining schedules, maintaining synchronization, and adaptive listening. • Collision avoidance. If multiple neighbors want to talk to a node at the same time, they will try sending when the node starts listening; hence, they need to contend for the medium. Among the contention protocols, the IEEE 802.11 performs well on collision avoidance. S-MAC follows similar procedures, including virtual and physical carrier sense, and the RTS/CTS exchange accounting for the hidden terminal problem (Bharghavan et al. 1994). To attain collision avoidance, many necessary steps are followed: – There is a duration field in each transmitted packet that indicates how long the remaining transmission will be. If a node receives a packet destined to another node, it knows how long to keep silent from this field. The node records this value in a variable called the network allocation vector (NAV) and sets a timer for it. – When the timer fires, the node decrements its NAV until it reaches zero. – Before initiating a transmission, a node first looks at its NAV. If its value is not zero, the node determines that the medium is busy. This is called virtual carrier sense. – Physical carrier sense is performed at the physical layer by listening to the channel for possible transmissions.

4.1 Duty-Cycling Approach Taxonomy

187

– Carrier sense time is randomized within a contention window to avoid collisions and starvations4 (Raynal 2013). The medium is determined as free if both virtual and physical carrier senses indicate that it is free. All senders perform carrier sense before initiating a transmission. If a node fails to get the medium, it goes to sleep and wakes up when the receiver is free and listening again. Broadcast packets are sent without using RTS/CTS. Unicast packets follow the sequence of RTS/CTS/DATA/ACK between the sender and the receiver. After the successful exchange of RTS and CTS, the nodes will use their normal sleep time for data packet transmission; they do not follow their sleep schedules until they are done transmitting. With the low duty-cycle operation and the contention mechanism during each listen interval, S-MAC effectively addresses the energy waste due to idle listening and collisions. Periodic sleeping, as introduced, effectively reduces energy waste on idle listening. S-MAC nodes coordinate their sleep schedules rather than randomly sleeping on their own. This coordinated sleeping is performed through choosing and maintaining schedules, maintaining synchronization, and adaptive listening as detailed below: • Choosing and maintaining schedules. Each node chooses a schedule and exchanges it with its neighbors, before starting its periodic listen and sleep. All nodes maintain a schedule table that stores the schedules of all their known neighbors. To choose its schedule and establish its schedule table, a node goes thru multiple steps: – It first listens for a fixed amount of time, which is at least the synchronization period. If it does not hear a schedule from another node, it immediately chooses its own schedule and starts to follow it. Meanwhile, the node tries to announce the schedule by broadcasting a SYNC packet. Broadcasting a SYNC packet follows the normal contention procedure. The randomized carrier sense time reduces the chance of collisions on SYNC packets. – If the node receives a schedule from a neighbor before choosing or announcing its schedule, it sets its own schedule to be the same as the one received. Then, the node will announce its schedule at its next scheduled listen time. – There are two cases if a node receives a different schedule after it chooses and announces its own schedule:

4

Starvation is a problem encountered in concurrent computing where a process is perpetually denied necessary resources to process its work. Starvation may be caused by errors in a scheduling or mutual exclusion algorithm, but can also be caused by resource leaks and can be intentionally caused via a denial-of-service attack. In computer networks, especially wireless networks, scheduling algorithms may suffer from scheduling starvation; an example is maximum throughput scheduling.

188

4 Energy Management Techniques for WSNs …

If the node has no other neighbors, it will discard its current schedule and follow the new one. If the node already follows a schedule with one or more neighbors, it adopts both schedules by waking up at the listen intervals of the two schedules. • Maintaining synchronization. Since neighboring nodes coordinate their sleep schedules, the clock drift on each node can cause synchronization errors. Two techniques are used to make it robust to such errors: – All exchanged time stamps are relative rather than absolute. – The listen period is significantly longer than clock drift rates. For instance, the listen time of 0.5 s is more than 10 times longer than typical clock drift rates. Compared to TDMA schemes with very short time slots, S-MAC requires much looser time synchronization. Although the long listen time can tolerate fairly large clock drift, neighboring nodes still need to periodically update each other with their schedules to prevent long-term clock drift. The synchronization period can be quite long. The testbed measurements show that the clock drift between two nodes does not exceed 0.2 ms/s. Schedule updating is accomplished by sending a SYNC packet. The SYNC packet is very short and includes the address of the sender and the time of its next sleep. In order for a node to receive both SYNC packets and data packets, its listen interval is divided into two parts. The first one is for SYNC packets, and the second one is for data packets, as shown in Fig. 4.28. Each part has a contention window with many time slots for senders to perform carrier sense. For example, if a sender wants to send a SYNC packet, it starts carrier sense when the receiver begins listening; a time slot is randomly selected to finish its carrier sense. If it has not detected any transmission by the end of that time slot, it wins the contention and starts sending its SYNC packet. The same procedure is followed when sending data packets. • Adaptive listening. The scheme of periodic listen and sleep considerably reduces the time spent on idle listening when traffic load is light. However, when a sensing event happens, it is desirable that the sensing data can be passed through the network without too much delay. When each node strictly follows its sleep schedule, there is a potential delay on each hop, whose average value is proportional to the length of the frame. Therefore, a mechanism is introduced to switch the nodes from the low duty-cycle mode to a more active mode. The node that overhears its neighbor transmissions, ideally only RTS or CTS, wakes up for a short period of time at the end of the transmission. Thus, if the node is the next-hop node, its neighbor will be able to immediately pass the data to it instead of waiting for its scheduled listen time. If the node does not receive anything during the adaptive listening, it will go back to sleep until its next scheduled listen time. As shown in Fig. 4.28, if the next-hop node is a neighbor of the sender, it will receive the RTS packet. If it is only a neighbor of the receiver, it will receive the

4.1 Duty-Cycling Approach Taxonomy

189

Listen

Receiver For SYNC

For RTS

For CTS

Sleep Time

Tx SYNC Sender 1

CS Sleep Time Tx RTS

Sender 2

Got CTS

CS Send data Time Tx SYNC

Sender 3

CS

Tx RTS

Got CTS

CS Send data Time

Legend: CS stands for carrier sense. Fig. 4.28 Timing relationship between a receiver and multiple senders (Ye et al. 2004)

CTS packet from the receiver. Thus, from the duration field in the RTS and CTS packets, both the neighbors of the sender and receiver will learn how long the transmission is. They will thus be able to adaptively wake up when the transmission is over. Also shown, the interval of the adaptive listening does not include the time for the SYNC packet as in the normal listen interval. SYNC packets are only sent at scheduled listen time to ensure that all neighbors can receive it. To give priority to the SYNC packet, adaptive listen and transmission are not performed if the duration from the time the previous transmission is finished to the normally scheduled listen time is shorter than the adaptive listen interval. Noticeably, not all next-hop nodes can overhear a packet from the previous transmission, especially when the previous transmission starts adaptively, i.e., not at the scheduled listen time. So if a sender starts a transmission by sending out an RTS packet during the adaptive listening, it might not get a CTS reply. In this case, it just goes back to sleep and will try again at the next normal listen time.

4 Energy Management Techniques for WSNs …

190

Now, as S-MAC concepts were introduced, the analytical analysis of latency provides a precise description of the multihop latency and quantifies the delay introduced by periodic sleeping in S-MAC. Several delays are inherent to a multihop network using contention-based MAC protocols; they are available as well in both S-MAC and IEEE 802.11-like protocols. Specifically, these delays are: • Carrier sense delay. It is introduced when the sender performs carrier sense; its value is determined by the contention window size. • Backoff delay. It happens when carrier sense fails, either because the node detects another transmission or because collision occurs. • Transmission delay. It is determined by channel bandwidth, packet length, and the coding scheme adopted. • Propagation delay. It goes with the distance between the sending and the receiving nodes. In WSNs, node distance is normally very small, and the propagation delay can be ignored. • Processing delay. It occurs because the receiver needs to process the packet before forwarding it to the next hop. This delay is based on the computing power of the node and the efficiency of in-network data processing algorithms. • Queueing delay. It depends on the traffic load; in heavy traffic, it becomes a dominant factor. Sleep delay is an extra delay in S-MAC. It is caused by the periodic sleeping of each node. When a sender gets a packet to transmit, it must wait until the receiver wakes up. It is called sleep delay since it is instigated by the sleep of the receiver. The latency of different MAC protocols is analyzed in the simple case that the traffic load is very light, e.g., only one packet is moving through the network, so that there is no queueing delay and backoff delay. Also assumed, the propagation delay and the processing delay can be ignored. In this case, only carrier sense delay, transmission delay, and sleep delay are taken into account. Supposing that there are N hops from the source to the sink. The carrier sense delay, tcs,n, at hop n is random at each hop, and its mean value, tcs, is determined by the contention window size. The transmission delay, ttx, is fixed if the packet length is fixed. Several scenarios are considered: • For a MAC protocol without sleeping. When a node receives a packet, it immediately starts carrier sense and tries to forward it to the next hop. The average delay at hop n is tcs,n + ttx. The entire latency over N hops is:

DðNÞ ¼

N  X n¼1

tcs;n þ ttx



ð4:22Þ

4.1 Duty-Cycling Approach Taxonomy

191

Thus, the average latency over N hops in the MAC without sleeping is: E ½DðNÞ ¼ N ðtcs þ ttx Þ

ð4:23Þ

Equation 4.23 shows that, in the MAC protocol without sleeping, the multihop latency linearly increases with the number of hops. The slope of the line is the average carrier sense time plus the packet transmission time. • For S-MAC, which introduces a sleep delay at each hop, denoted by ts,n for the nth hop, and assuming that all nodes along the path follow the same sleep schedule, a frame is a complete cycle of listen and sleep, with length denoted by Tf. Noticeably, the listen interval is fixed, and the frame length can be changed by adjusting the sleep interval. To reflect a very low duty-cycle
10%, it is assumed that Tf has a value much larger than ttx. The delay at hop n is: Dn ¼ ts;n þ tcs;n þ ttx

ð4:24Þ

• In S-MAC without adaptive listening, contention (carrier sense) only starts at the beginning of each frame, i.e., the time each node starts listening. After a node receives a packet in a frame, it has to wait until the next-hop node to wake up, which is the beginning of the next frame. This indicates that: Tf ¼ tcs;n1 þ ttx þ ts;n

ð4:25Þ

So, the sleep delay at hop n is:   ts;n ¼ Tf  tcs;n1 þ ttx

ð4:26Þ

Substituting Eqs. 4.24 and 4.26 becomes: Dn ¼ Tf þ tcs;n  tcs;n1

ð4:27Þ

There is an exception on the first hop, because a packet can be generated on the source node at any time within a frame. So, the sleep delay on the first hop, ts,1, is a random variable whose value lies in (0, Tf). Suppose ts,1 is uniformly distributed in (0, Tf), its mean value is Tf /2. Combining it with Eq. 4.27, the overall delay of a packet over N hops is: DðNÞ ¼ D1 þ

N X n¼2

Dn ¼ ts;1 þ ðN  1Þ  Tf þ tcs;N þ ttx

ð4:28Þ

4 Energy Management Techniques for WSNs …

192

n i

ts,n

n+1 j

tcs,n + ttx Listen

n+2 k

tcs,n+1 + ttx

ts,n+2

Sleep

l

tcs,n+2 + ttx Listen

Sleep Time

Tf Fig. 4.29 Adaptive listen reduces sleep latency by at least half (Ye et al. 2004)

Thus, the average latency of S-MAC without adaptive listen over N hops is: E ½DðNÞ ¼ N  Tf  Tf =2 þ tcs þ ttx

ð4:29Þ

Equation 4.29 shows that the multihop latency also linearly increases with the number of hops in S-MAC when each node strictly follows its sleep schedules. The slope of the line is the frame length Tf. Compared with Eq. 4.23, Tf is normally much larger than (tcs + ttx) due to the very low duty-cycles. Therefore, periodic sleeping introduces an additional delay at each hop. • For S-MAC with adaptive listening. Figure 4.29 shows part of a multihop network, where the three hops are denoted as n to (n + 2). All nodes are assumed to follow the same sleep schedule. Suppose node i first waits for node j to wake up at its normally scheduled listen time and starts carrier sense for sending data from that moment. The delay at hop n is still expressed as Eq. 4.24. During the RTS/CTS exchange between nodes i and j, the next-hop node k is also listening and overhears j’s CTS packet. So, node k knows when the transmission from i to j will end. The adaptive listen mechanism will wake up node k immediately after the previous transmission is over; it also lets node j start carrier sense for sending to k at that time. Thus, the delay at hop (n + 1) is: Dn ¼ tcs;n þ 1 þ ttx

ð4:30Þ

There is no sleep delay when compared with the delay at the previous hop. If the frame length Tf is larger than ðtcs;n þ tcs;n þ 1 þ 2  ttx Þ, the packet will travel over two hops in just one frame. This condition holds in the following analysis, since it is assumed that Tf is much larger than ttx. On the other hand, node l is two hops away from node j; it may not be able to overhear j’s CTS packet as k does. In this case, l cannot wake up when the transmission from j to k is done. When j starts sending to k during the normal sleep time, node l is not aware of it, since it is in sleep state. Therefore, node l will not be able to wake up when the transmission from j to k is over. To start

4.1 Duty-Cycling Approach Taxonomy

193

its transmission, node k has to wait until l’s normal listen time. The delay on hop (n + 2) is again expressed by Eq. 4.24. Hence, the sleep delay occurs at every other hop in S-MAC with adaptive listen. The latency over N hops is: DðNÞ ¼ ts;1 þ tcs;1 þ ttx þ tcs;2 þ ttx þ ts;3 þ þ tcs;N1 þ ttx þ tcs;N þ ttx ð4:31Þ From Fig. 4.29: Tf ¼ tcs;n þ ttx þ tcs;n þ 1 þ ttx þ ts;n þ 2

ð4:32Þ

Equation 4.31 can be simplified to be: DðNÞ ¼ ts;1 þ ðN=2  1Þ  Tf þ tcs;N1 þ tcs;N þ 2  ttx

ð4:33Þ

The average latency over N hops in S-MAC with adaptive listen is: E ½DðN Þ ¼ N  Tf =2 þ 2  tcs þ 2  ttx  Tf =2

ð4:34Þ

The average latency in S-MAC with adaptive listen still linearly increases with the number of hops. The slope of the line is Tf /2; compared with Eq. 4.29 of no adaptive listen, it is reduced by half. Equation 4.34 is obtained under the assumption that only one-hop neighbors can hear each other, but two-hop neighbors cannot hear each other. In real world, this is not generally true. The latest implementation of S-MAC was on MICA motes based on Atmel ATmega128L microcontroller with 128 KByte Flash and 4 KByte data memory (Atmel 2011). The MICA motes are equipped with the RFM TR3000 radio transceiver (Murata Electronics 2015a) and a matched whip antenna. The modulation scheme is the amplitude shift keying (ASK). The power consumptions of the radio in receiving, transmitting, and sleep modes are 14.4 mW, 36 mW, and 15 W, respectively. S-MAC implementation was not based on the standard communication stack in the TinyOS release; instead, a stack was implemented with new features that are critical to S-MAC (Ye et al. 2002a). The goal of the experimentation was to reveal the fundamental tradeoffs of energy, latency, and throughput in S-MAC: • Measurement of energy consumption: The energy consumption on the radio is obtained by measuring the amount of time the radio on each node has spent in sleep, idle, receiving, or transmitting modes. The energy consumption in each mode is then calculated by multiplying the time with the required power to operate the radio in that mode. The energy consumption is compared for IEEE 802.11-like protocol without sleep, S-MAC without periodic sleep, and S-MAC with periodic sleep. Several test scenarios were devised:

4 Energy Management Techniques for WSNs …

194

– Tests on a two-hop network: The topology is a two-hop network with two sources and two sinks. The traffic load is changed by varying the interarrival period of messages. In this scenario, the message interarrival period varies from the highest 1 s rate to the lowest at 10 s. If the message interarrival period is 5 s, a message is generated every 5 s by each source node. Three MAC operating modes are implemented and compared: typically, IEEE 802.11-like protocol without sleep, S-MAC without periodic sleep, and S-MAC with periodic sleep at 50% duty-cycle. The average energy consumption is measured on the source nodes. It was realized from experimentation that: For a heavy traffic, when the message interarrival time is less than 4 s, the IEEE 802.11 MAC uses more than twice the energy used by S-MAC. Since idle listening rarely happens, energy savings from periodic sleeping are very limited. S-MAC achieves energy savings mainly by avoiding overhearing and efficiently transmitting long messages. When the message interarrival period is larger than 4 s, traffic load becomes light. In this case, S-MAC has the best energy performance and far outperforms IEEE 802.11 MAC. Message passing with overhearing avoidance also performs better than IEEE 802.11 MAC. However, when idle listening dominates the total energy consumption, the periodic sleep plays a key role for energy savings. Compared with IEEE 802.11 MAC, message passing with overhearing avoidance saves almost the same amount of energy under all traffic conditions. This is due to overhearing avoidance among neighboring nodes. – Tests on a multihop network: The adopted topology is a linear network with 11 nodes, as shown in Fig. 4.30. The nodes are configured to send at the minimum transmission power and are placed along a 1-m line. The first node is the source, and the last node is the sink. The traffic load is varied by changing the packet interarrival time on the source node. In this setup, the packet interarrival time changes from 0 s to 10 s, where at 0 s all the packets are generated and queued at the same time on the source node. There is no fragmentation on all messages. Three different operation modes of S-MAC are compared. The first one is 10% duty-cycle without adaptive listen, the second one is 10% duty-cycle with adaptive listen, and the last one is a fully active mode where periodic sleep is completely disabled. Interestingly, the obtained results conform with those of the two-hop network: Source 1

Sink

… 2

3

Fig. 4.30 Ten-hop linear topology (Ye et al. 2004)

10

11

4.1 Duty-Cycling Approach Taxonomy

195

S-MAC with periodic sleep achieves substantial energy savings over the MAC without periodic sleep, especially when traffic load is light. Comparing the two MAC modules running at the 10% low duty-cycle, the one with adaptive listen achieves better energy efficiency than the one without adaptive listen, especially when traffic load is heavy. The main reason is that the adaptive listen largely reduces the overall time needed to pass the fixed amount of data through the network. • Measurement of end-to-end latency. As S-MAC trades off latency for energy savings, it can have longer latency in a multihop network due to the periodic sleep on each node. Adaptive listen, as earlier described, is designed to minimize such additional latency. The ten-hop network topology of Fig. 4.30 is used to measure the end-to-end latency of S-MAC. The measurements are also on the same S-MAC modes used for measuring the energy consumption. No fragmentation is assumed on the messages. Two extreme traffic conditions are assumed, the lowest traffic load and highest traffic load: – Under the lowest traffic load, the second message is generated on the source node after the first one is received by the sink. To do this, a coordinating node is placed near the sink; when it hears that the sink receives the message, it signals the source directly by sending at the highest power. Under this traffic load, there is no queueing delay on each node. Compared with the MAC without sleep, the extra delay is only caused by the periodic sleep on each node. – Under the highest traffic load, all messages are generated and queued on the source node at the same time. So, there is a maximum queueing delay on each node including the source node. For both traffic scenarios, the latency of each message is measured from the time it is generated on the source node. Interesting results are obtained as outlined in what follows for the modes 10% duty-cycle without adaptive listen, 10% duty-cycle with adaptive listen, and the fully active mode where periodic sleep is completely disabled: – In the lowest traffic load: For all three S-MAC modes, the measured mean message latency on each hop increases linearly with the number of hops. However, S-MAC operating at 10% duty-cycle without adaptive listen has much higher latency. The reason is that each message has to wait for one sleep cycle on each hop. The latency of S-MAC with adaptive listen, by comparison, is very close to that of the MAC without any periodic sleep, because adaptive listening often allows S-MAC to immediately send a message to the next hop. However, this mode does not reach the lowest latency in the MAC of fully active mode. To be reminded, adaptive listen cannot guarantee the immediate transmission at each hop. If a node sends an RTS but fails to get a CTS from the intended receiver, it has to wait for its next cycle.

4 Energy Management Techniques for WSNs …

196

S-MAC with adaptive listen has about twice the average latency of the MAC in fully active mode, except for the first one or two hops. For either low duty-cycle modes, the variance in latency is much larger than that in the fully active mode, and it increases with the number of hops. The large variance is due to the fact that some messages may miss the sleep cycles of certain nodes. – In the highest traffic load: The 10% low duty-cycle mode without adaptive listen has the highest mean message latency on each hop, about 3–5 times that of the 10% low duty-cycle with adaptive listen mode. For 10% duty-cycle with adaptive listen mode, the latency is twice that in the fully no-sleep cycle active mode. • Measurement of end-to-end throughput. Throughput is the maximum possible number of bytes of data delivered in a time unit, excluding any control packets. Contention happens at each hop, which can significantly reduce throughput. Throughput is evaluated on the ten-hop network of Fig. 4.30 for different traffic loads. The measured throughput on node n represents the (n−1) hops across the network. Out of experimentation, several findings are reached: – For the highest traffic load, periodic sleeping reduces throughput at each hop. – Compared with the fully active mode without sleep cycles, the 10% low duty-cycle modes with and without adaptive listen only achieve about 1/2 and 1/8 of the throughputs at 10 hops, respectively. Throughput is understandably lower because latency is higher, since sometimes sending is delayed. – Similar to reducing latency, the 10% low duty-cycle mode with adaptive listen significantly improves the end-to-end throughput. – The results certify that, for the three MAC modes, throughput drops as the number of hops increases, due to the RTS/CTS contention in the multihop network. – For different traffic loads, i.e., different message interarrival times on the source, the throughput from the source to the sink is measured. It is shown that both throughputs of fully active mode without sleep cycles and that of the 10% duty-cycle with adaptive listen mode decrease as traffic load decreases (higher message interarrival time). When the traffic load is very low, they all approach that of the non-adaptive mode (no-sleep cycles), because the three MAC modes spend about the same time to finish transmitting the same number of messages. During this long time between two messages, it is worthless to spend more energy trying to increase throughput, since there is no enough traffic. It can be observed from the illustrated measurements of energy consumption, end-to-end latency, and end-to-end throughput that when traffic load is very heavy (interarrival time less than 1.5 s), the 10% duty-cycle with adaptive listening and the no-sleep modes both show statistically equivalent performance that is

4.1 Duty-Cycling Approach Taxonomy

197

significantly better than the 10% duty-cycle without adaptive listen. Performance is evaluated in energy–time cost per byte (J * s/Byte). For this traffic, both adaptive listen and no-sleep are almost always active, while the added delay of non-adaptive sleep requires extra transmission time and lower overall energy–time efficiency. At lower traffic load (interarrival time longer than 4 s), the energy–time cost without sleeping quickly exceeds the cost of sleep modes. The cost of no-sleep mode grows linearly in the limit. The 10% duty-cycle adaptive and non-adaptive sleeping modes become statistically equivalent at lower traffic load (interarrival time at 9 s or more). This result indicates that the overhead for adaptive listening is minimal. The benefits of adaptive listen occur at moderate to high traffic loads. Hence, periodic sleeping provides significant energy performance at light traffic load, while at heavy load, adaptive listening is able to adjust to traffic and provides energy performance as good as no-sleep. This indicates that S-MAC with adaptive listening is convenient for WSNs where traffic is intermittent. Although S-MAC achieves low-power operation, it is not simple to implement and lacks both scalability and tolerance to changing network conditions. As the size of the network increases, S-MAC must maintain an increasing number of schedules of surrounding nodes, which incurs additional overhead through repeated rounds of resynchronization. An Adaptive Energy-Efficient MAC Protocol for WSNs (T-MAC) The timeout-MAC (T-MAC) protocol is based on reducing idle listening by transmitting all messages in bursts of variable length and sleeping between bursts (Van Dam and Langendoen 2003); it is an enhanced version of S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”) specifically designed for variable traffic load. The length of the active time is dynamically determined to obtain an optimal active time under variable load. The active time is ended intuitively thru timing out on hearing nothing. T-MAC protocol design and evaluation were based on OMNeT++ discrete event simulation and on a realistic hardware model built for the EYES wireless sensor nodes (Havinga et al. 2003). Throughout section “4.1.2.2,” it was iterated that energy usage is the main design criterion for MAC protocols. Energy waste has been identified due to idle listening, collisions, protocol overhead, and overhearing. Idle listening is the main cause of energy waste, especially when messages are infrequent. Although reducing the idle listening time is a solution in fixed duty-cycle protocols, like S-MAC, it is not optimal. It is essential to remember that while latency requirements and buffer space are generally fixed, the message rate will usually vary. If important messages are not to be missed and unimportant messages should not have been sent, nodes must be deployed with an active time that can handle the highest expected load. Whenever the load is lower than expected, the active time is not optimally used and energy will be wasted on idle listening.

4 Energy Management Techniques for WSNs …

198 A

Contend

RTS CTS

Data

ACK

B

C

Contend

Contend

TA Legend: Node C overhears the CTS from node B and will not disturb the communication between node A and node B. TA must be long enough for node C to hear the start of CTS.

Fig. 4.31 T-MAC basic data exchange (Van Dam and Langendoen 2003)

Figure 4.31 displays the basic scheme of the T-MAC protocol. Every node periodically wakes up to communicate with its neighbors and then goes to sleep again until the next frame; meanwhile, new messages are queued. Nodes communicate with each other using RTS, CTS, data, ACK, which traditionally provides both collision avoidance and reliable transmission. T-MAC involves the following design fundamentals: • Clustering and synchronization. Frame synchronization in T-MAC is inspired by virtual clustering, described in S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”), and itemized in what follows: – When a node comes to life, it starts by waiting and listening: If it hears nothing for a certain amount of time, it chooses a frame schedule and transmits a SYNC packet that contains the time until the next frame starts. If the node, during startup, hears a SYNC packet from another node, it follows the schedule in that SYNC packet and transmits its own SYNC accordingly. – Nodes retransmit their SYNC once in a while. They must also listen for a complete frame sporadically, so they can detect the existence of different schedules. This allows new and mobile nodes to adapt to an existing group. – If a node has a schedule and hears SYNC with a different schedule from another node, it must adopt both schedules. It must also transmit SYNC with its own schedule to the other node, to let it know about the presence of another schedule. Adopting both schedules means that the node will have an activation event at the start of both frames. – Nodes must start a data transmission only at the start of their own active time. At that time, both neighbors with the same schedule, and neighbors that have adopted the schedule as extra, are awake. If a node would start transmission at the start of a neighbor frame, it might be transmitting to another, sleeping neighbor. Notably, broadcasts need only to be transmitted once.

4.1 Duty-Cycling Approach Taxonomy

199

A T-MAC node keeps listening and potentially transmitting while in an active period. An active period ends when no activation event has occurred for a time TA. An activation event is: – – – – –

The firing of a periodic frame timer. The reception of any data on the radio. The sensing of communication on the radio, e.g., during a collision. The end of transmission of a node own data packet or acknowledgment. The knowledge, acquired through overhearing prior RTS and CTS packets, that a data exchange of a neighbor has ended.

• Calculating the active time, TA. A node sleeps if it is idle and not in an active period. Note that TA is an upper bound on the idle listening time per frame at the end of the active time. The described timeout scheme moves all communication to a burst at the beginning of the frame. Since messages between active times must be buffered, the buffer capacity determines an upper bound on the maximum frame time. A node should not go to sleep while its neighbors are still communicating, since it may be the receiver of a subsequent message. Receiving the start of the RTS or CTS packet from a neighboring node is enough to trigger a renewed interval TA. Since an out-of-range node may not hear the RTS that starts a communication with its neighbor, the interval TA must be long enough to receive at least the start of the CTS packet (Fig. 4.31). This observation gives a lower limit on the length of the interval TA: TA [ C þ R þ T

ð4:35Þ

where, C is the length of the RTS contention interval, R is the length of an RTS packet, T is the turnaround time (the short time between the end of the RTS packet and the beginning of the CTS packet). During experimentation, TA was set to 1:5  ðC þ R þ T Þ, which proved to be satisfactory. A larger TA increases the energy used. • Fixed contention interval. In T-MAC, every node transmits its queued messages in a burst at the start of the frame. During this burst, the medium is saturated as messages are transmitted at the maximum rate. A node contends for winning the medium every time it sends an RTS. An increasing contention interval is not useful, since the load is mostly high and does not change. Therefore, RTS transmission in T-MAC starts by waiting and listening for a random time within a fixed contention interval tuned for maximum load. The contention time is always used, even if no collision has occurred yet. • RTS retries. When a node sends an RTS, but does not receive a CTS back, one of three scenarios might happen:

200

4 Energy Management Techniques for WSNs …

– The receiving node has not heard the RTS due to collision. – The receiving node is prohibited from replying due to an overheard RTS or CTS. – The receiving node is asleep. When the sending node receives no answer within the interval TA, it might go to sleep. However, for scenarios 1 and 2 this would not hold since there is a situation where the sending node goes to sleep, while the receiving node is still awake. Since such a situation might occur even at the first message of the frame, the throughput would dramatically decrease. Therefore, a node should retry resending the RTS if it receives no answer; if there still is no reply after two retries, it should give up and go to sleep. • Overhearing avoidance. The S-MAC protocol introduced the idea of sleeping after overhearing an RTS or CTS destined for another node. Since a node is prohibited from sending during that time, it cannot take part in any communication and may as well turn OFF its radio to save energy. Overhearing avoidance is an option in T-MAC. However, by experimentation, it was found that as a side effect, collision overhead becomes higher; a node may miss other RTS and CTS packets while sleeping and may cause some communication disturbance when it wakes up. Consequently, the maximum throughput decreases, for short packets by as much as 25%. So, although overhearing avoidance saves energy, it must not be used if maximum throughput is a requirement. • Asymmetric communication. The early sleeping problem occurs when a node goes to sleep while a neighbor still has messages for it. In the nodes-to-sink communication pattern, the early sleeping problem reduced the total possible throughput of T-MAC to less than half of the maximum throughput of traditional protocols as S-MAC. This problem might occur in any asymmetric communication pattern. Two T-MAC solutions are thus devised: – Future request to send. This scheme lets another node know that there is a message for it, but the requesting-to-send node is prohibited from using the medium. Hence, if a node overhears a CTS packet destined for another node, it may immediately send a future-request-to-send (FRTS) packet, like node C in Fig. 4.32. The FRTS packet contains the length of the blocking data communication (this information was in the CTS packet). A node must not send an FRTS packet if it senses communication just after the CTS or if it is prohibited from sending due to a prior RTS or CTS. From the timing information in the FRTS packet, a node that receives this packet knows it will be the future target of an RTS packet and must be awake by that time. As the FRTS packet would otherwise disturb the data packet that follows the CTS, the data packet must be postponed for the duration of the FRTS packet. To prevent any other node from taking the channel during this time, the node that sent the initial RTS (node A in Fig. 4.32) transmits a small data-send (DS) packet. After the DS packet, it must immediately send the normal data packet.

4.1 Duty-Cycling Approach Taxonomy

A

Contend

RTS CTS DS

201 Data

ACK

B

C

Contend

Contend

Active D

TA

Active FRTS

RTS

Fig. 4.32 FRTS packet exchange (Van Dam and Langendoen 2003)

Since the FRTS packet has the same size as a DS packet, it will collide with the DS packet, but not with the following data packet. The loss of the DS packet is not serious, as it contains no useful information. For the FRTS solution to work, TA must be increased with the length of a control packet (CTS), as shown in Fig. 4.32. Implementing the FRTS feature increased the maximum throughput in unidirectional communication patterns by 75%. However, due to the higher overhead of DS and FRTS packets, energy usage also increased slightly. One may want to use FRTS packets only if a reasonably high load in a more or less unidirectional communication pattern is expected. However, when the load is low, the number of exchanged packets, and therefore the increased overhead, is also low. – Taking priority on full buffers. In the second scheme, when a node transmit/ routing buffers are almost full, it may favor sending over receiving. This means that when a node receives an RTS packet destined for it, it immediately sends its own RTS packet to another node, instead of replying with a CTS, as usual. The full-buffer priority scheme has two effects. First, the node has a higher chance of transmitting its own message, since it effectively wins the medium upon hearing a competing RTS; in Fig. 4.33, node C may transmit to node D after losing contention to node B. Thus, there is a lower probability for the early sleeping problem to occur. Second, a limited form of flow control is introduced into the network, which is advantageous in a nodes-to-sink communication pattern; in Fig. 4.33, node B is prevented from sending until node C has enough buffer space. Care is to be taken with the full-buffer priority scheme, since it is risky in a high-load situation where communication is not unidirectional. When all

4 Energy Management Techniques for WSNs …

202

A

Contend

B

Contend

C

Contend

Active D

TA

RTS CTS

Data

ACK

Fig. 4.33 Taking priority on full buffers (Van Dam and Langendoen 2003)

nodes in an omnidirectional communication pattern start taking priority, chances of collisions increase rapidly. Therefore, T-MAC uses a threshold; a node may only use this scheme when it has lost contention at least twice. Experimentally, this threshold kept the performance in an omnidirectional communication pattern, while still increasing the maximum throughput in a unidirectional pattern. T-MAC protocol design and evaluation were based on the OMNeT++ discrete event simulation package (Fahmy 2016); also, a realistic hardware model was built for the EYES wireless sensor nodes (Havinga et al. 2003). Energy usage in the model is based on the amount of energy the real nodes use: specifically, 20 lA while sleeping, 4 mA while receiving, and 10 mA while transmitting a DC-balanced signal (RFM 2004, Texas Instruments 2004, and Murata Electronics 2014). Using these modeled nodes, a network of 100 nodes in a 10 by 10 grid was built. The radio range was such that non-edge nodes all have eight neighbors (Fig. 4.34). Several simulation scenarios were devised and tested as bulleted in what follows: • Nodes-to-sink communication. In this experiment, nodes send messages to a single sink node at the corner of the network. Messages are routed from node to node with a slightly randomized shortest path algorithm. No data aggregation is used. For the T-MAC protocol, overhearing avoidance, the full-buffer priority, and the FRTS mechanisms were used. Several results occurred as revealed in Fig. 4.35:

4.1 Duty-Cycling Approach Taxonomy

203

Fig. 4.35 Nodes-to-sink at a 100-Bytes message length (Van Dam and Langendoen 2003)

Average energy used (mA/node)

Fig. 4.34 Part of the simulation network (Van Dam and Langendoen 2003)

Load (Message/node/second)

– T-MAC uses less energy than S-MAC; this is because in nodes-to-sink communication the load varies with the location of nodes, and there is more traffic in the neighborhood of the sink node. – The maximum throughput of the T-MAC protocol is less than that of the S-MAC protocol. This is mainly due to variations on the early sleeping problem as described when presenting the asymmetric communication. – The relative throughput of T-MAC, as compared to S-MAC, decreases when the messages are larger. In that case, the penalty of the early sleeping problem is higher. • Early sleeping problem. This run tests the effectiveness of the measures that address the early sleeping problem in the previously explained asymmetric communication. From the acquired results illustrated in Fig. 4.36, it was found that:

4 Energy Management Techniques for WSNs …

Fig. 4.36 Nodes-to-sink options at a 20-Bytes message length (Van Dam and Langendoen 2003)

Average energy used (mA/node)

204

Load (Message/node/second)

– The FRTS mechanism increases maximum throughput by approximately 75% (0.08 vs. 0.14 messages per second), at an energy cost. – Adding the full-buffer priority mechanism adds approximately 30% (0.14 vs. 0.18 messages per second) without the cost of extra energy. • Event-based local unicast. In this more realistic run, events occur in the network with a frequency of one event per 10 s. Events have an average duration of 5 s and affect an area of approximately 9 nodes. These nodes then send local unicast messages to their neighbors for the duration of the event. A neighbor that receives one of these messages replies with a probability of 20%. For T-MAC, overhearing avoidance is adopted without FRTS and full-buffer priority. The obtained findings are itemized below and shown in Fig. 4.37: – T-MAC uses much less energy than either S-MAC or CSMA, especially when the message frequency increases during events. – The maximum frequency that T-MAC can handle is lower than that of S-MAC, similarly for the nodes-to-sink communication pattern. – T-MAC suffers from the early sleeping problem, because there relatively are many edge nodes. Regarding T-MAC features and from the above presentation, it can be concluded that: • Timeouts present a simple but effective way to address the idle listening problem in an environment with variable network traffic load. Implementation with T-MAC has shown the advantages, both in simulation and on real hardware. During a high load, nodes will communicate without sleeping. However, during a very low load, nodes will use their radios for as little as 2.5% of the time, saving as much as 96% of the energy compared to a traditional protocol. • The T-MAC protocol can easily be implemented; its final implementation requires no more than 42 Bytes of state. • The timeout scheme, as implemented by T-MAC, uses significantly less energy than a scheme with a fixed duty-cycle, such as S-MAC, when applied to an

Fig. 4.37 Event-based unicast pattern at a 20-Bytes message length (Van Dam and Langendoen 2003)

205

Average energy used (mA/node)

4.1 Duty-Cycling Approach Taxonomy

Load (Message/node/second)

•

• • • • •

environment with varying message rates. The advantage depends mainly on the amount of variation in the message rate. Even in a homogeneous, non-varying environment T-MAC can be advantageous, since it requires no tuning to the specific message rate, while S-MAC must be tuned precisely for optimal energy savings. Deployment of T-MAC can therefore be faster. T-MAC improves on S-MAC energy usage by using a very short listening window at the beginning of each active period. After the SYNC section of the active period, there is a short window to send or receive RTS and CTS packets. If no activity occurs in that period, the node returns to sleep. By changing the protocol to have an adaptive duty-cycle, T-MAC saves power at a cost of reduced throughput and additional latency. T-MAC, in variable workloads, uses one-fifth of the power of S-MAC. In homogeneous workloads, T-MAC and S-MAC perform equally well. T-MAC suffers from the same complexity and scaling problems of S-MAC. Shortening the active window in T-MAC reduces the ability to snoop on surrounding traffic and adapt to changing network conditions. Simulations of T-MAC have shown the early sleeping problem, which decreases throughput dramatically in an asymmetric communication pattern. The FRTS and the full-buffer priority mechanisms were proposed to increase the throughput by more than 100%, which was still less than optimal. This is a tradeoff to the adaptiveness of the protocol. However, very high message rates are not expected in WSNs.

An Adaptive Energy-Efficient and Low-Latency MAC for Data Gathering in WSNs (D-MAC) In many WSN applications, the major traffic pattern consists of data collected from several source nodes to a sink through a unidirectional data gathering tree. Data gathering MAC (D-MAC) is an energy-efficient and low-latency MAC protocol designed and optimized for such data gathering trees (Lu et al. 2004). Prior to

206

4 Energy Management Techniques for WSNs …

D-MAC, proposed protocols based on activation/sleep duty-cycles suffered from a data forwarding interruption problem, whereby nodes on a multihop path to the sink are not all notified of ongoing data delivery, thus resulting in significant sleep delay. There is also a problem when each single source has low traffic rate, but the aggregated rate at an intermediate node is larger than what the basic duty-cycle can accommodate. The interference between nodes with different parents could instigate interruption of a traffic flow because the nodes on the multihop path are not notified of the data transmission requirements. Concerned with such issues, D-MAC is designed to attain several objectives: • Solving the interruption problem by giving the active/sleep schedule of a node an offset that depends upon its depth on the tree. This scheme allows continuous packet forwarding because all nodes on the multihop path can be notified of the data delivery in progress. • Adjusting node duty-cycles adaptively according to the traffic load in the network by varying the number of active slots scheduled in an interval. • Proposing a data prediction mechanism and the use of more-to-send (MTS) packets in order to alleviate problems pertaining to channel contention and collisions. Simulation using ns-2 (Fahmy 2016) has shown that by exploiting the application-specific structure of data gathering trees in WSNs, D-MAC provides significant energy savings and latency reduction while ensuring high data reliability. There are several works on reducing sleep delay and adjusting duty-cycle to the traffic load. Mechanisms are either implicit such as S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”) and T-MAC (section “An Adaptive Energy-Efficient MAC Protocol for WSNs (T-MAC)”) where nodes remain active on overhearing of ongoing transmission or explicit as in AWP (section “Asynchronous Wakeup Protocol (AWP) for Ad Hoc Networks”), in which there are direct duty-cycle adjusting messages. S-MAC proposed adaptive listening to reduce the sleep delay. In adaptive listening, a node that overhears its neighbor transmission wakes up for a short period of time at the end of the transmission, so that if it is the next hop of its neighbor, it can receive the message without waiting for its scheduled active time. In T-MAC, a node keeps listening and potentially transmitting as long as it is in an active period. An active period ends when no activation event has occurred for a certain time. The activation time events include reception of any data, the sensing of communication on the radio, and the end of transmission of a node’s own data packet or acknowledgment. FRTS is employed to solve the early sleep problem. The AWP offered a slot-based power management mechanism. If the number of buffered packets for an intended receiver exceeds a threshold L, the sender signals the receiver to remain ON for the next slot. A node intended to stay awake sends an

4.1 Duty-Cycling Approach Taxonomy

207

acknowledgment to the sender, indicating its willingness to remain awake in the next slot. The sender can then send a packet to the receiver in the following slot. The request is renewed on a slot-by-slot basis. However, in the mechanisms proposed prior to D-MAC, whether explicit or implicit, not all nodes beyond one hop away from the receiver can overhear the data communication; therefore, packet forwarding will stop after a few hops. This data forwarding interruption problem causes sleep latency for packet delivery. D-MAC employs a staggered active/sleep schedule to solve this problem and enable continuous data forwarding on the multihop path. Data prediction is used to enable active slot request when multiple children of a node have packets to send in a same sending slot, while MTS packet is used when nodes, with different parents, on the same level of the data gathering tree compete for channel access. To further elaborate on the data forwarding interruption problem, it is necessary to emphasize that: • This problem exists in implicit adaptive duty-cycle techniques because the overhearing range is limited by radio sensitivity to signals on air. Nodes that are out of the hearing range of both the sender and the receiver are unaware of ongoing data transmissions and therefore go to sleep until the next interval. The data forwarding process will then stop at the node whose next hop toward the sink is out of the overhearing range because it is in sleep mode. Packets will then have to be queued until the next active period, which increases latency. • For the explicit mechanism, the duty-cycle adjusting messages can only be forwarded to limited hops in an active period. Thus, out-of-range nodes go to sleep after their basic duty-cycle, leading to interrupted data forwarding. An active period is the portion of time in each interval when a node is active, unless there is more data to be sent/received. Assuming the active period only long enough to transmit one packet each hop, the performance of S-MAC and T-MAC protocols is recalled: • In S-MAC, only the next hop of the receiver can overhear the data transmission and remains active for a long period. Other nodes on the multihop path do not overhear the data transmission and thus go to sleep after the basic active period, resulting in the interruption of packet forwarding to the sink till the next duty-cycle. The delay with adaptive listening still increases linearly with the number of hops at a slope that is half the interval length. Therefore, compared with the case of no adaptive listening, the delay is only reduced by half. Meanwhile, nodes other than those next hops in the neighborhood of the sender and the receiver also overhear a data transmission and hence may remain active unnecessarily. • Similarly, in T-MAC, a node remains active if it senses any communication on the air. Typically, a radio interference range is more than twice its transmission range. Any neighbor nodes in the interference range of either the sender or the receiver will remain active. Many of the nodes do not participate in the data delivery but remain active for an unnecessarily long period, which wastes

208

4 Energy Management Techniques for WSNs …

energy. Meanwhile, only nodes in the interference range hear the communication, while the nodes out of the interference range and on the multihop path still go to sleep after their basic active period. Therefore, packets still suffer from the data forwarding interruption problem. The proposed FRTS can increase the number of packets delivered in one frame; however, as a side effect, it can help forward a packet one hop further. • The same problem happens in AWP where the request for a next active slot can be only received by the next hop. The nodes beyond that will still go to sleep after their basic active period. The hearing/interference range causes a tradeoff between latency and energy. If the hearing range is long, latency is reduced since more nodes on the path can overhear the communication and remain active. Meanwhile, nodes out of the path also overhear the communication and waste energy in idle listening on the increased active periods. D-MAC can tell all nodes on the path to stay active and/or increase their duty-cycles and all other nearby nodes to sleep in order to enable continuous data forwarding without incurring energy waste of unrelated nodes. As detailed below, D-MAC is built upon a staggered wakeup schedule, data delivery and duty-cycle adaptation in multihop chain, data prediction, and MTS packet: • Staggered wakeup schedule. D-MAC is proposed to deliver data along the data gathering tree, aiming at both energy efficiency and low latency. As a prelude for D-MAC, three communication patterns relevant to WSN applications are to be highlighted: – Local data exchange and aggregation purely among nearby nodes; these can be handled by clustering or simple medium access mechanisms. – Dispatching of control packets from the sink to sensor nodes. Such sink-originated traffic is small in number and may not be latency sensitive. A separate active slot can be reserved periodically with a larger interval length for such control packets. – Data gathering from sensor nodes to sink; it is the most significant traffic pattern in WSNs. For a WSN application with multiple sources and one sink, the data delivery paths from sources to sink follow a data gathering tree structure (Krishnamachari et al. 2002). Several assumptions are worth consideration: Routes may change during data delivery, while sensor nodes are fixed without mobility. A route to the sink is fairly durable so that a data gathering tree remains stable for a reasonable length of time. Flows in the data gathering tree are unidirectional from sensor nodes to sink. The sink is the only destination. All nodes except the sink will forward any packets they receive to the next hop, excluding local processing packets, which are handled in cluster.

4.1 Duty-Cycling Approach Taxonomy

209 Nodes in the chain

0

Source

Sink

Time span

2 3 4 5 6 7 Send

Receive

Sleep

Legend: Data transmission is between the nodes spread along the horizontal axis for the time span prolonging over the vertical axis. Fig. 4.38 D-MAC chain transmission (Lu et al. 2004)

In D-MAC, the activity schedule of nodes is staggered on the multihop path to wake them up sequentially like a chain reaction (Fig. 4.38). A communication interval is divided into receiving, sending, and sleep periods. In receiving state, a node is expected to receive a packet and send an ACK packet back to the sender. In the sending state, a node will try to send a packet to its next hop and in return will receive an ACK packet. In sleep state, nodes turn OFF their radio to save energy. The receiving and sending periods have the same length l, which is enough for one packet transmission and reception. Depending on its depth d in the data gathering tree, a node skews its wakeup scheme d*l ahead of the schedule of the sink. In this structure, data delivery can only be done in one direction, toward the root. Intermediate nodes have a sending slot immediately after the receiving slot. A staggered wakeup schedule has four advantages: – Since nodes on the path wake up sequentially to forward a packet to the next hop, sleep delay is eliminated if there is no packet loss due to channel error or collision. – A request for longer active period can be propagated all the way down to the sink, so that all nodes on the multihop path can increase their duty-cycle promptly to avoid data from being stuck in intermediate nodes. – Since the active periods are now separated, contention is reduced. – Only nodes on the multihop path need to increase their duty-cycle, while other nodes can still operate on the basic low duty-cycle to save energy. Typical for a multihop wireless network, contention-based MACs suffer from the hidden node problem. In media access protocol for wireless LAN

4 Energy Management Techniques for WSNs …

210

(MACAW) (Bharghavan et al. 1994), virtual and physical carrier sense and RTS/CTS exchange are utilized to reduce the hidden node problem. For large packet sizes, these small control packets are efficient to save possible high cost of a packet loss. However, for WSNs where packet size is usually small, the overhead of RTS/CTS could be significantly high compared to the actual data transmission cost. Therefore, the use of RTS/CTS is not favored in D-MAC. D-MAC, however, employs link layer ARQ through ACK control packet and data retransmission. The hidden node problem is mitigated, to some extent, thru scheduling active slots such that nodes on the same path do not cause hidden node collisions. Although ACK packets consume energy and bandwidth, they are essential for achieving reliability by recovering packets lost due to low-quality wireless channel and contention. If a sending node does not receive an ACK packet from a receiving node, it will queue the packet until the next sending slot. After three retransmissions, the packet will be dropped. Under D-MAC, nodes with the same depth will have the same offset and thus a synchronous schedule. During the sending period, nodes will compete for the channel. To reduce collision during this period, every node backs off for a backoff period (BP) plus a random time within a contention window at the beginning of a sending slot. Since the length of a sending slot is only enough for one packet transmission, there is no need for an increased exponential contention window; therefore, a fixed contention window is used. When a node receives a packet, it waits for a short period (SP) and then transmits the acknowledgment packet back to the sender. BP and SP are two interframe spaces with BP > SP in order to assure collision-free reception of the acknowledgment packet. Based on the above choices, the sending and receiving slot length l is set to: l ¼ BP þ CW þ DATA þ SP þ ACK

ð4:36Þ

where, CW is the fixed contention window size, DATA is the packet transmission time, and all packets are assumed to have the same length, ACK is the acknowledgment packet transmission time. Synchronization is needed in D-MAC. However, local synchronization is enough since a node only needs to be aware of its neighbor schedule. There exist techniques like the reference broadcast synchronization scheme (RBS) (Elson et al. 2002) that can achieve time synchronization precision of 3.68 ± 2.57 ls after four hops. Slot lengths are in the order of 10 ms.

4.1 Duty-Cycling Approach Taxonomy

211

• Data delivery and duty-cycle adaptation in multihop chain proceed thru several steps (Fig. 4.38): – Every node periodically switches between send, receive, and sleep states. – When there is no collision, a packet will be forwarded sequentially along the path to the sink, without sleep latency. However when a node has multiple packets to send at a sending slot, it increases its own duty-cycle and requests other nodes on the multihop path to increase their duty-cycles. – A slot-by-slot renewal mechanism is adopted. With limited overhead, a data flag is piggybacked in the MAC header, to indicate the request for an additional active period. – Before a node in its sending state transmits a packet, it sets the packet MORE-DATA flag if either its buffer is not empty, or it received a packet from the previous hop with MORE-DATA flag set. – The receiver checks the MORE-DATA flag of the packet it received, and if the flag is set, it also sets the MORE-DATA flag of its ACK packet to the sender. With the slot-by-slot mechanism and the policy to set MORE-DATA flag when the buffer is not empty, D-MAC can react quickly to traffic rate variation both to be energy-efficient and to maintain low data delivery latency. A node will decide to hold an additional active period if: – It sends a packet with the MORE-DATA flag set and receives an ACK packet with the MORE-DATA flag set. – It receives a packet with the MORE-DATA flag set. Even if a node decides to hold an additional active period, it does not remain active for the next slot but schedules a 3l sleep and then goes to the receiving state (Fig. 4.38). The reason for a 3l sleep is that the node knows that the following nodes on the multihop path will forward through the path in the next 3 slots. The maximum utilization of a chain of ad hoc nodes is 1/4 if the radio interference range is twice the transmission range (Li et al. 2001). So, the maximum sending rate for a node is one packet per four slots. To accommodate the possibility of short range between two neighbor nodes, a node will only send one packet every 5l in order to avoid collision. This might reduce the maximum network capacity by about 20%, but if the traffic load is more than 80% of the maximum channel capacity, duty-cycled mechanisms would not function efficiently, making such a concern pointless. A worthy consequence of the staggered wakeup schedule is that the MORE-DATA flag can be propagated to all the nodes on the multihop path. Measurements showed that the cost for switching the radio between active and sleep is small compared to energy savings in a 3l sleep period of about 30 ms (Raghunathan et al. 2002). • Data prediction. In a data gathering tree, there is a chance that each source rate is small enough for the basic duty-cycle, while the aggregated rate at an

4 Energy Management Techniques for WSNs …

212

intermediate node exceeds the capacity of the basic duty-cycle. For example, suppose a node C has two children A and B, both children have only one packet to send every interval. At the sending slot of an interval, only one child can win the channel and send a packet to the node. Assume A wins the channel and sends a packet to C. Since A’s buffer is empty, the MORE-DATA flag is not set in A packet. C then goes to sleep after its sending slot without a new active period. B’s packet would thus have to be queued until next interval. This results in sleep delay for packets from B. A scheme called data prediction is proposed to deal with this scenario. If a node in receiving state receives a packet, it predicts that its children still have packets waiting for transmission. It then sleeps only 3l after its sending slot and switches back to receiving state. All the following nodes on the path also receive this packet and schedule an additional receiving slot. In this additional receiving slot for data prediction, if no packet is received, the node will go to sleep directly without a sending slot. If a packet is received during this receiving slot, the node will wake up again 3l later after this sending slot. For a node in sending state, during its backoff period it overhears the ACK packet from its parent in the data gathering tree. Hence, it knows that this sending slot is already taken by its brother, but its parent will hold an additional receiving slot 3l, so it will also wake up 3l later after its sending slot. In this additional sending slot, the node then can transmit a packet to its parent. Figure 4.39 displays how the data prediction scheme works. There is an overhead brought by the data prediction scheme. After receiving the last packets from its children, a node will remain idle for a receiving slot, wasting thus energy in idle listening. Compared to the significant latency reduction by the data prediction, this additional overhead would be tolerable. Nodes in the chain 0

Sources

Sink

…

Time span

… 3 4 5 6

Legend: Data transmission is between the nodes spread along the horizontal axis for the time span prolonging over the vertical axis. Fig. 4.39 Data prediction to reduce sleep delay (Lu et al. 2004)

4.1 Duty-Cycling Approach Taxonomy

0 Legend:

213

T

T+

Time

Non successful transmission

Fig. 4.40 Sleep delay due to interference between two sending nodes (Lu et al. 2004)

• More to send (MTS). Although a node sleeps 3l before an additional active period to avoid collision, there is still a chance of interference between nodes on different branches of the tree. In Fig. 4.40, two nodes A and B are in interference range of each other with different parents in the data gathering tree. In the sending slot of one interval, A wins the channel and transmits a packet to its parent. Neither B nor its parent C holds additional active slots in this interval. Thus, B can only send its packet in the sending slot of the next interval, resulting in T sleep latency. Since C does not receive any packet in its receiving slot and B does not overhear ACK packet from C in its sending slot, data prediction scheme will not work. An explicit control packet, more to send (MTS), is proposed to adjust the duty-cycle under interference. The MTS packet is very short, just containing destination local ID and a flag. The MTS packet with flag set to 1 is a request MTS, while the MTS packet with flag set to 0 is a clear MTS. A node sends a request MTS to its parent if one of two conditions is true: – It cannot send a packet because the channel is busy. After the node backoff timer fires, it has no enough time to send a packet, and it does not overhear its parent ACK packet. The node then assumes losing the channel because of interference from other nodes. – It received a request MTS from its children. This targets propagating the request MTS to all nodes on the path. A request MTS is sent only once, before a clear MTS packet is sent. A node sends clear MTS to its parent if the following three conditions are true: – Its buffer is empty. – All request MTSs received from children are cleared. – It sends a request MTS to its parent and has not sent a clear MTS. A node, which sent or received a request MTS, will keep waking up periodically every 3l. It switches back to the basic duty-cycle only after sending a clear MTS to its parent or when all previous received request MTSs from its children were cleared. Similar to the slot-by-slot renewal and data prediction schemes, the higher duty-cycle requests by MTS packets are forwarded through the staggered schedule to all nodes on the multihop path. However, to reduce the overhead of

4 Energy Management Techniques for WSNs …

214

MTS packets, instead of sending MTS packets to renew the active period slot by slot, only two MTS packets are sent for MTS request/clear period. Performance evaluation of D-MAC is implemented thru the ns-2 network simulator with the Carnegie Mellon University (CMU) wireless extension (Fahmy 2016). For comparison, a simple version of S-MAC with adaptive listening is implemented, but without its synchronization and message passing scheme. Also, a comparison is performed with a full active CSMA/CA MAC without periodical sleep schedule. Three metrics are chosen to evaluate the performance of D-MAC: • Energy cost, it is the total energy cost to deliver a certain number of packets from sources to a sink; this metric shows the energy efficiency of the MAC protocols. • Latency, the end-to-end delay of a packet. • Throughput or delivery ratio, the ratio of the number of packets arriving at the sink to the number of packets sent by sources.

Width (meters)

According to the ns-2 simulator setup values, the energy cost ratio of the Tx/Rx/ idle radio modes is about 1.67/1/0.88. The sleeping power consumption is set to 0. The MTS packet is 3 Bytes long. According to the parameters of the radio and for 100 Bytes packet length, the receiving and sending slot l is set to 10 ms for D-MAC and 11 ms for D-MAC/MTS. The active period is set to 20 ms for S-MAC with adaptive listening. All schemes have a 10% basic duty-cycle, and this means a sleep period of 180 ms for D-MAC and S-MAC, and 198 ms for D-MAC/MTS. For a random data gathering tree topology, 50 nodes are distributed randomly in a 1000 m  500 m area as shown in Fig. 4.41. The sink node is at the right bottom

Length (meters)

Legend: The five selected data gathering trees are green colored. Fig. 4.41 A random 1000 m  500 m topology with 50 nodes (Lu et al. 2004)

4.1 Duty-Cycling Approach Taxonomy

215

corner. A data gathering tree is constructed by each node choosing from its neighbor the node closest to the sink as its next hop. In order to show the different packet latencies, a source should be at least three hops away from the sink. Five nodes at the margin are chosen as sources to testify the mechanism of data prediction and MTS. All sources generate reports at the same rate. Comparisons between full active CSMA/CA, S-MAC, D-MAC, and D-MAC/ MTS disclosed meaningful results regarding latency, energy consumption, throughput, and scalability: • Latency analysis yielded satisfactory outcomes: – Full active CSMA/CA has small delay for all traffic loads. – S-MAC, D-MAC, and D-MAC/MTS latencies increase significantly when the traffic load is larger than a certain threshold. – D-MAC/MTS can handle the highest traffic load with small delay compared to S-MAC and D-MAC. – Compared to the multihop chain under the same heavy traffic load, the latency in a data gathering tree is much higher. This is due to the interference between nodes in the same depth of the tree. The interference could cause data loss, as well schedule inconsistency and MTS packet loss, thus increasing the sleep latency. • Energy consumption study has shown that: – Energy costs of all the 50 nodes in the network are collected because potentially a MAC could cause unrelated nodes to maintain a higher duty-cycle. – D-MAC and D-MAC/MTS are the two most energy-efficient MAC protocols. D-MAC/MTS, however, consumes higher energy than D-MAC because of the overhead of MTS packets and more active period requests by MTS packets. • Throughput determination disclosed that: – In terms of end-to-end throughput, D-MAC/MTS has a good delivery ratio while S-MAC and D-MAC delivery ratios decrease when traffic load is heavy. • Scalability. The scalability of D-MAC is evaluated for a dense network where 100 nodes are randomly placed in a 100 m  500 m area. A data gathering tree rooted at the sink is constructed. All sources generate traffic at a rate of one message per 3 s. The obtained results are interesting: – As the number of sources increases, interference increases resulting in increased latency for S-MAC and D-MAC without MTS (Fig. 4.42). – D-MAC/MTS, however, can still maintain low latency at very small energy overhead compared to D-MAC without MTS (Fig. 4.43). – D-MAC/MTS also has the second delivery ratio next to full active CSMA (Fig. 4.44). This clearly shows the effectiveness of D-MAC/MTS.

4 Energy Management Techniques for WSNs …

216 Fig. 4.42 Mean packet latency for a data gathering tree at a different number of sources (Lu et al. 2004)

Latency (sec)

Full active S-MAC D-MAC D-MAC/MTS

Number of sources

Energy (Joules)

Fig. 4.43 Energy consumption for a data gathering tree at a different number of sources (Lu et al. 2004)

Number of sources

• Tradeoff between energy, throughput, and latency. Figure 4.45 shows the number of packets that can be sent per unit resource measured in terms of energy  latency as a function of the traffic load. From the figure, it is depicted that: – Because S-MAC achieves energy efficiency at the sacrifice of latency, it sends the least number of packets per (Joule  second). Hence, for applications that can tolerate message latency, S-MAC is favorable. But for applications that require real-time data delivery, S-MAC is not feasible due to the data forwarding interruption problem. – D-MAC and D-MAC/MTS can achieve both energy efficiency and low message latency.

4.1 Duty-Cycling Approach Taxonomy

217

Delivery ratio

Fig. 4.44 Data delivery ratio for a data gathering tree at a different number of sources (Lu et al. 2004)

Number of packets per (Joule

Fig. 4.45 Tradeoff between energy, latency, and throughput for a data gathering tree under different traffic loads (Lu et al. 2004)

second)

Number of sources

Source report interval (sec)

– D-MAC/MTS can operate with even smaller basic duty-cycle to save more energy when traffic is light and can still adapt to traffic bursts with high throughput, low latency, and small energy consumption. – When traffic load exceeds a certain threshold, a full active MAC is most suitable when taking both energy and delay into account. Since D-MAC can adjust duty-cycle to traffic load with small latency, the basic duty-cycle can be set small. But a lower duty-cycle could have longer initial sleep delay at the source node when a sensing reading occurs while the source radio is OFF. So, there should be a limit on the lowest basic duty-cycle for D-MAC operation. Importantly, with the same required application latency bound, D-MAC can operate on a lower basic duty-cycle than S-MAC or T-MAC, while being more energy-efficient.

218

4 Energy Management Techniques for WSNs …

Lastly, the comparison between D-MAC and S-MAC is only applicable under the specific data gathering tree scenario for unidirectional communication flow from multiple sources to a single sink. S-MAC is in fact a general-purpose energy-efficient MAC that can handle simultaneous data transmissions and flows between arbitrary sources and destinations. For applications that require data exchange between arbitrary sensor nodes, D-MAC cannot be used, and S-MAC will be the suitable choice. Versatile Low-Power Media Access for Sensor Networks (B-MAC) Berkeley-MAC (B-MAC) is a carrier sense media access (CSMA) protocol for WSNs that provides a flexible interface to obtain ultra-low-power operation, effective collision avoidance, and high channel utilization (Polastre et al. 2004). B-MAC is a configurable MAC protocol for WSNs; it is simple in both design and implementation as it has a small core and factors out higher-layer functionality. This minimalist B-MAC model is in contrast to the classic monolithic MAC protocols optimized for a general set of workloads. To achieve low-power operation, B-MAC employs an adaptive preamble sampling scheme that reduces duty-cycle and minimizes idle listening. B-MAC supports on-the-fly reconfiguration and provides bidirectional interfaces for system services to optimize performance, whether it is throughput, latency, or power conservation. B-MAC presents the design of a MAC protocol motivated by monitoring applications. An analytical model is built for a WSN monitoring application to demonstrate the effect of changing B-MAC parameters and to predict the application behavior. Also, thru experimentally comparing B-MAC to conventional IEEE 802.11inspired protocols, specifically S-MAC, it is shown that B-MAC flexibility results in better packet delivery rates, throughput, latency, and energy consumption. Before getting into B-MAC design concepts and functioning details, a differentiation from S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”) is required for proper evaluation of both protocols: • S-MAC is an example of WSN protocol classically designed; it provides an RTS/CTS mechanism for channel arbitration and hidden terminal avoidance, synchronization with neighbors to realize low-power operation, and message fragmentation for efficiently transferring bulk data. S-MAC is not only a link protocol, but also a network and organization protocol. Applications and services must rely on S-MAC internal policies to adjust their operation while node and network conditions change; such changes are opaque to the applications. • In contrast, the B-MAC protocol contains a small core of media access functionality. B-MAC uses clear channel assessment (CCA)5 (Engstrom and Gray 2008) and packet backoffs for channel arbitration, link layer acknowledgments 5

CCA is a technique used by MAC protocols to effectively avoid collisions, and it accurately determines if the channel is clear.

4.1 Duty-Cycling Approach Taxonomy

219

for reliability, and low-power listening (LPL) for low-power communication. B-MAC is just a link protocol, with network services like organization, synchronization, and routing built above its implementation. Although B-MAC neither provides multi-packet mechanisms like hidden terminal support or message fragmentation, nor enforces a particular low-power policy, it has a set of TinyOS interfaces that allow services to tune their operation in addition to the standard message interfaces. These TinyOS interfaces allow network services to adjust B-MAC mechanisms, including CCA, acknowledgments, backoffs, and LPL. By exposing a set of configurable mechanisms, protocols built on B-MAC can make local policy decisions to optimize power consumption, latency, throughput, fairness, and reliability. An array of principles forms the foundations of B-MAC design: • Since the ambient noise varies according to the environment, B-MAC employs software automatic gain control for estimating the noise floor. Signal strength samples are taken at times when the channel is assumed to be free, such as immediately after transmitting a packet or when the data path of the radio stack is not receiving valid data. A common method used in a variety of protocols, including IEEE 802.15.4, takes a single sample and compares it to the noise floor. This thresholding method produces results with a large number of false negatives that lower the effective channel bandwidth. Since noise has significant variance in channel energy whereas packet reception has fairly constant channel energy (Fig. 4.46), B-MAC searches for outliers6 (NIST/SEMATECH 2013) in the received signal such that the channel energy is significantly below the noise floor. If an outlier exists during the channel sampling period, B-MAC declares the channel clear since a valid packet could never have an outlier significantly below the noise floor. If five samples are taken and no outlier is found, the channel is busy. The effectiveness of outlier detection as compared to thresholding on a trace from a CC1000 transceiver (Texas Instruments 2007) is portrayed in Fig. 4.46. • B-MAC provides optional link layer acknowledgment support; if acknowledgments are enabled, it immediately transfers an acknowledgment code after receiving a unicast packet. If the transmitting node receives the acknowledgment, an acknowledge bit is set in the sender transmission message buffer. • B-MAC duty-cycles the radio through periodic channel sampling called low-power listening (LPL). Each time a node wakes up, it turns ON the radio and checks for activity; the period between consecutive wakeups is called check interval. If an activity is detected, the node powers up and stays awake for the time required to receive the incoming packet. After reception, the node returns to sleep. If no packet is received (false positive), a timeout forces the node back to sleep. Clear channel assessment (CCA) is critical to achieving low-power operation with this method. The noise floor estimation of B-MAC is used not

A value that “lies outside” (is much smaller or larger than) most of the other values in a set of data.

6

4 Energy Management Techniques for WSNs …

Outlier

Threshold

Signal (dBm)

220

Time (msec) Legend: The top graph is a trace of the RSSI indicator from a CC1000 transceiver. A packet arrives in the interval 22-54 msec. The middle graph shows the output of a thresholding CCA algorithm; a 1 indicates the channel is clear, a 0 indicates it is busy. The bottom graph shows the output of an outlier detection algorithm.

Fig. 4.46 CCA effectiveness for a typical wireless channel (Polastre et al. 2004)

only for finding a clear channel for transmission, but also for determining if the channel is active during LPL. False positives in the CCA algorithm, such as those caused by thresholding, severely affect the duty-cycle of LPL due to increased idle listening. • To reliably receive data, the preamble length is matched to the interval that the channel is checked for activity. If the channel is checked every 100 ms, the preamble must be at least 100 ms long for a node to wake up, detect activity on the channel, receive the preamble, and then receive the message. Idle listening occurs when the node wakes up to sample the channel and there is no activity. The interval between LPL samples is maximized so that the time spent sampling the channel is minimized. Transmit mode corresponds to the preamble length, and the listening mode corresponds to the check interval. A selection of eight different modes is provided (corresponding to 10, 20, 50, 100, 200, 400, 800, and 1600 ms for the check interval). Protocols may also set their own preamble length and check interval through the interface. • In WSNs, each node typically runs a single application. Since the RAM and ROM available on sensor nodes are extremely limited, keeping the size of the MAC implementation small is important. B-MAC is implemented in TinyOS (TinyOS 2012). Since it does not have the RTS/CTS mechanism or synchronization requirements of S-MAC, the implementation is both simple and small.

4.1 Duty-Cycling Approach Taxonomy

221

Above B-MAC, an RTS/CTS scheme is implemented, as well as a message fragmentation service using B-MAC control interfaces that have equivalent functionality to S-MAC RTS/CTS and fragmentation services. The operation of a WSN monitoring application is investigated, in the paragraphs to follow, thru an analytical model. The model allows calculating and setting B-MAC parameters to optimize the application overall power consumption. The model illustrates the effect of different application variables including duty-cycle, network density, and sampling rate. B-MAC TinyOS interfaces may be used by network services to adapt to current consumption. The node lifetime is determined by its overall energy consumption. For the lifetime to be maximized, the energy consumption must be minimized. The sum of the energies, E, is defined in units of millijoules per second (milliwatts). Calculating the total energy usage can be done by multiplying E by the node lifetime tl. For WSN applications, the energy used by a node consists of the energy consumed by receiving, transmitting, listening for messages on the radio channel, sampling data, and sleeping: E ¼ Erx þ Etx þ Elisten þ Ed þ Esleep

ð4:37Þ

Current (mA)

Sampling sensors is often expensive and affects the lifetime of the node as illustrated through the monitoring application built on the MICA2 mote and the CC1000 transceiver (Mainwaring et al. 2002). Application operations are portrayed in Fig. 4.47, and the sampling parameters are listed in Table 4.6. Each node needs

Time (msec) Legend: The node first starts in sleep state (a). Then wakes up on a timer interrupt (b). The node initializes the radio configuration and commences the radio startup phase. The startup phase (c) waits for the radio crystal oscillator to stabilize. Upon stabilization, the radio enters receive mode (d). After the receive mode switch time, the radio enters receive mode (e) and a sample of the received signal energy may begin. After the ADC starts acquisition, the radio is turned OFF and the ADC value is analyzed (f). With LPL, if there is no activity on the channel, the node returns to sleep (g).

Fig. 4.47 Application operations performed when the radio is turned ON (Polastre et al. 2004)

4 Energy Management Techniques for WSNs …

222 Table 4.6 Time and current consumption to satisfy primitive operations of the monitoring application (Polastre et al. 2004)

Operation

Time

Initialize radio (b)

350 ls

trinit

Current (mA) 6 crinit

Turn ON radio (c) Switch to RX/TX (d)

1.5 ms 250 ls

tron trx=tx

1 15

cron crx=tx

Time to sample radio (e) Evaluate radio sample (f) Receive 1 Byte Transmit 1 Byte Sample sensors

350 ls 100 ls 416 ls 416 ls 1.1 s

tsr tev trxb ttxb tdata

15 6 15 20 20

csr cev crxb ctxb cdata

1100 ms to start its sensors, sample, and collect data. Data are sampled every five minutes, that is every r ¼ 1=ð5  60Þ seconds. The energy associated with sampling data, Ed, is given by: td ¼ tdata  r Ed ¼ td  cdata  V

ð4:38Þ

where, td tdata cdata r V

is is is is is

the sampling time, the data length in seconds, the current consumed for sampling, the sampling rate in packets/sec, a 3 V MICA2 mote voltage (Crossbow 2002).

The energy consumed for transmitting, Etx, is given by:   ttx ¼ r  Lpacket þ Lpreamble  ttxb Etx ¼ ttx  ctxb  V

ð4:39Þ

where, ttx Lpacket Lpreamble ttxb ctxb

is is is is is

the the the the the

transmit time, packet length in bytes, length of the preamble in bytes, time to transmit one byte, current consumed in transmitting one byte.

For a periodic application with a uniform sampling rate, the node will detect and receive data when each of its n neighbors transmits a packet, regardless of the packet destination. Although receiving data from neighbors shortens a node lifetime, it allows services to snoop on the channel and make decisions based on channel activity.

4.1 Duty-Cycling Approach Taxonomy

223

The total time the node will spend on receiving can be constrained, and the upper bound on the energy consumed for receiving, Erx, may be calculated:   trx
n  r  Lpacket þ Lpreamble  trxb Erx ¼ trx  crxb  V

ð4:40Þ

where, trx Lpacket Lpreamble trxb crxb

is is is is is

the the the the the

receive time, packet length in bytes, length of the preamble in bytes, time to receive one byte, current consumed in receiving one byte.

In order to reliably receive packets, the LPL check interval, ti, must be less than the time of the preamble. Therefore, the below constraint holds:   Lpreamble  ti =trxb

ð4:41Þ

where trxb is the time to receive one byte. Given a check interval and associated preamble length, the time spent sampling the channel can be calculated. From Fig. 4.47, the energy consumption of a single LPL radio sample is Esample = 17.3 lJ. The total energy, Elisten , spent listening to the channel can be calculated as follows:   tlisten ¼ trinit þ tron þ trx=tx þ tsr  1=ti Elisten
Esample  1=ti

ð4:42Þ

where, trinit tron trx=tx tsr Esample 1=ti

is is is is is is

the the the the the the

time to initialize the radio, time to turn the radio ON, switch time to RX/TX, time to sample the radio, energy of a single-channel sample, channel sampling frequency.

The node must sleep for the remainder of the time, hence, the sleep time, tsleep, is the time remaining, in each second, that is not consumed by other operations: tsleep ¼ 1  td  ttx  trx  tlisten Esleep ¼ tsleep  csleep  V

ð4:43Þ

4 Energy Management Techniques for WSNs …

224

Table 4.7 Parameters for a monitoring application running B-MAC (Polastre et al. 2004) MICA2 mote parameters

B-MAC parameters

Application parameters

Notation

Parameter

Value

csleep cbatt V Lpreamble Lpacket ti n r

Sleep current Battery capacity Voltage Preamble length Packet length Radio sampling interval Neighborhood size Sampling rate

30 mA 2500 mAh 3V 271 Bytes 36 Bytes 100 ms 10 nodes 1/300 samples/s

where, td ttx trx tlisten csleep V

is is is is is is

the sampling time, the transmit time, the receive time, the time spent listening to the channel, the current consumed in sleeping, a 3 V MICA2 mote voltage (Crossbow 2002).

The lifetime of the node, tl, depends on the total energy consumed, E, and the battery capacity, Cbatt. Lifetime is bounded by the available capacity of the battery: tl ¼ ðCbatt  V=E Þ  60  60

ð4:44Þ

By solving the system of Eqs. 4.37–4.44 and entering the parameter values listed in Table 4.7, the minimum energy for a given network configuration can be calculated. Lifetime may be estimated at compile time or computed at runtime for a discrete set of values that provides reconfiguration feedback to network services. Experimentation and simulation were carried on to illustrate the effectiveness of a lightweight, configurable B-MAC protocol as compared to S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”) and T-MAC (section “An Adaptive Energy-Efficient MAC Protocol for WSNs (T-MAC)”): • Experimenting on B-MAC and S-MAC by micro-benchmarking was implemented using a set of simple workloads that represent an empirical characterization of protocol performance. The purpose was to show how the protocols react to typical working conditions: specifically, high contention, low to high throughput, low to high latency, and their correlation with power consumption. The use of RTS/CTS and message fragmentation services implemented using B-MAC interfaces are compared to those of the S-MAC protocol. Both B-MAC and S-MAC were included in TinyOS run on a MICA2 node (Crossbow 2002). Tests were performed in an unobstructed area with line of sight to every other node, and nodes were 1 m spaced and 15 cm elevated to

4.1 Duty-Cycling Approach Taxonomy

225

reduce near-field effects. To enable multihop networking, the RF output power of the node was trimmed to its minimum value. • In addition to the experimentation on a real WSN, T-MAC was simulated in MATLAB (Fahmy 2016). The time of each operation is calculated and multiplied by the current consumptions listed in Table 4.6 to get the overall expected power consumption of the T-MAC protocol. The aforementioned micro-benchmarks display B-MAC functionality and illustrate the effect of a wide array of network conditions on the energy consumption of S-MAC, T-MAC, and B-MAC. The metrics considered for functionality evaluation are channel utilization, throughput, fragmentation, and latency. The acquired results may be detailed in what follows: • Channel utilization. High channel utilization is critical for delivering a large number of packets in a short amount of time. In WSNs, quickly transferring bulk data is typically reflected in network reprogramming or extracting logged sensor data. By minimizing the time to send packets, network contention might be reduced, while in network reprogramming the network is up and reprogrammed as quickly as possible. To find the channel utilization under congestion, n nodes were placed equidistant from a receiver. Each node was transmitting as quickly as possible with the MAC protocol providing collision avoidance. The offered load was increased by adding transmitters. There is no node or radio duty-cycling in this test. The throughput achieved by B-MAC and S-MAC is shown in Fig. 4.48, and the findings are as bulleted:

Throughput (bps)

Percentage of channel capacity

– Better throughput is reached with fewer nodes trying to saturate the channel. With one transmitter, B-MAC outperforms S-MAC broadcast mode (RTS/

Number of nodes Legend: The throughput of each protocol is affected by the amount of nodes contending for the channel and the protocol overhead.

Fig. 4.48 Measured throughput of each protocol with no duty-cycle under a contended channel (Polastre et al. 2004)

4 Energy Management Techniques for WSNs …

226

– – –

–

CTS disabled) by 2.5 times and S-MAC unicast mode (RTS/CTS enabled) by 4.5 times. B-MAC outperforms S-MAC for broadcast traffic due to more sophisticated CCA and lower preamble overhead. For unicast traffic, S-MAC suffers from the overhead of RTS/CTS exchanges. Instead of using control messages like RTS/CTS for hidden terminal support, B-MAC relies on higher-layer services to send data in accordance with their traffic pattern. These services implement the appropriate hidden terminal support for their workloads. B-MAC exceeds the performance of S-MAC, but does not trade off fairness; each node achieves no more than 15% additional bandwidth than any other node. To yield higher B-MAC throughput, fairness can be bypassed as a requirement.

• Energy versus throughput. B-MAC was designed to run at both low and high data rates configured by dependent services. Low duty-cycle applications have extremely low network throughput; however, some application services, such as bulk data transfer, stress the high throughput functionality of the MAC protocol. For B-MAC, the optimal check interval ti is calculated for the traffic pattern; the test is run, and the energy consumption is calculated. For S-MAC, the optimal duty-cycle is calculated for the traffic pattern such that the data arrive within a set 10 s latency bound. The results for a ten-node network are displayed in Fig. 4.49:

Fig. 4.49 Measured power consumption for maintaining a given throughput in a ten-node network (Polastre et al. 2004)

Power consumed (mW or mJ/sec)

– At low data rates, S-MAC can use an extremely low duty-cycle to transmit and receive data. As the amount of data increases, S-MAC duty-cycle increases. As the duty-cycle increases, there are more active periods each with a dedicated SYNC period. Due to the overhead of the SYNC period at the beginning of each wakeup, the S-MAC energy consumption increases linearly.

Throughput (bps)

227

Energy per byte (mJ/Byte)

Energy per byte (mJ/Byte)

4.1 Duty-Cycling Approach Taxonomy

Fragment size (Bytes)

Fragment size (Bytes)

(a) 10 second message generation rate

(b) 100 second message generation rate

Fig. 4.50 Effective energy consumption per byte (Polastre et al. 2004)

– In B-MAC, at low throughput, long preambles are sent with a long check interval ti. Due to the tradeoff between idle listening and packet length, the overhead dominates the energy consumption. The overhead of LPL is mitigated by a more frequent check interval when the throughput exceeds 45 bps. Noteworthy, B-MAC power consumption below 45 bps is within 25% of S-MAC power consumption; however, B-MAC has significantly less states and no synchronization requirements. – Services using B-MAC may easily reconfigure the link protocol to change the check interval based on the network bandwidth, whereas services using S-MAC oblige creating a new schedule and resynchronizing. • Fragmentation. Small periodic data packets are the characteristic workload in WSNs, but certain cases arise where larger transfers are needed. S-MAC supports large message fragmentation using an RTS/CTS exchange for channel reservation (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”). To compare B-MAC directly to S-MAC design goal of efficient message fragmentation, similar network topology and operation are embraced. The tests shown in Fig. 4.50 disclosed interesting results: – B-MAC without fragmentation control is simply the B-MAC protocol with the default parameters. Each fragment is an independent packet with a long preamble. However, the middleware service could adjust B-MAC to minimize the energy consumed during bulk transfer. – In the measurements with fragmentation control, a message fragmentation service is built using TinyOS B-MAC interfaces (Polastre et al. 2004). The first fragment of the message is sent with LPL enabled and with extra bytes to inform the receiver of the number of fragments. The remaining fragments are sent with LPL disabled. After the last fragment, the sender and receiver re-enable LPL. Such flexibility helps realizing significant power savings and

4 Energy Management Techniques for WSNs …

228

Latency (msec)

Fig. 4.51 End-to-end latency (Polastre et al. 2004)

Number of hops

–

– – –

efficiency identical to S-MAC, without the additional overhead including RAM and ROM usage. When the message transmission period is large (Fig. 4.50b), the overhead of S-MAC is apparent; the energy consumption per byte of both S-MAC and T-MAC is higher at all fragment sizes than B-MAC with fragmentation support. The simple B-MAC approach with long preambles for each fragment yields the same power consumption as S-MAC without the additional complexity. When there is no activity on the channel, T-MAC removes the overhead incurred by S-MAC thru using adaptive active periods to return to sleep much quicker. The energy cost of breaking up a short message into even shorter fragments is significantly high in all protocols, which is not a worthy option in WSNs.

• Latency. When S-MAC is permitted to increase latency, the node duty-cycle might be reduced, thus preserving energy. The latency tests on the ten-hop network as provided in section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)” were reproduced for comparison with B-MAC (Fig. 4.30). In each test, the source node sends 20 messages with a payload of 100 Bytes. There is no fragmentation on any message. Figure 4.51 displays the test outcomes: – The latencies of B-MAC and S-MAC increase linearly with the number of hops. – The slope of the latency increases with the overhead of the MAC protocols. – When duty-cycling is disabled, the effect of RTS/CTS exchanges in S-MAC results in a much steeper slope than B-MAC. – For low-power communications (100 ms check), B-MAC has a slope almost identical to S-MAC with adaptive listening; however, the Y-intercept is much lower.

4.1 Duty-Cycling Approach Taxonomy

229

Energy (mJ)

Fig. 4.52 Effect of latency on power consumption (Polastre et al. 2004)

Latency (sec)

– Since the first S-MAC packet cannot be sent until an active period, it is delayed by at most 1150 ms. Through adaptive listening, S-MAC does not incur an expected 1150 ms additional delay at each hop. – Link layer acknowledgments, a protocol feature of B-MAC, increase latency by an insignificant amount. To better evaluate the effect of increasing the latency on reducing power consumption, the throughput was set to a fixed one 100 Bytes packet per 10 s interval. For S-MAC, the end-to-end latency of the ten-hop network was measured while the duty-cycle was varied. For B-MAC, the optimal ti was selected given the latency and throughput. The acquired results are shown in Fig. 4.52: – For latencies under 6 s, B-MAC performs significantly better than S-MAC. As S-MAC approaches the 10-second latency limit, its power gets lower than that of B-MAC. When the latency exceeds 3 s, B-MAC power consumption is bounded by the cost of idle listening. – In contrast, the best-case performance of S-MAC relies on synchronizing the entire ten-hop network and using adaptive listening to transmit the data through the network in one active period. – The tests emphasize the importance of reconfiguration in WSNs. If the latency required for the application is relaxed, S-MAC could achieve lower energy consumption than B-MAC. But, S-MAC is not reconfigurable as it operates at a single setting, which limits saving energy. To conclude, since B-MAC is lightweight and configurable, several WSN protocols such as S-MAC and T-MAC may be implemented efficiently as services that use its primitives. S-MAC and T-MAC are more than just link protocols; they perform synchronization, organization, fragmentation, and hidden terminal support and could benefit from B-MAC flexibility.

230

4 Energy Management Techniques for WSNs …

Hybrid MAC Protocols The basic idea behind hybrid MAC protocols revolves around switching the protocol behavior between carrier sense multiple access (CSMA) and time-division multiple access (TDMA), depending on the level of contention. Depending on the WSN to be designed and deployed, the appropriate hybrid protocol is suggested. CSMA is popular in WSNs due to its simplicity, flexibility, and robustness. It does not require much infrastructure support; specifically, no clock synchronization and global topology information are required, and dynamic node joining and leaving are handled gracefully without extra operations. These advantages, however, come at the cost of trial and error; a trial may cost access collision where more than two “conflicting” nodes transmit at the same time, causing signal fidelity degradation at destinations. Collision can happen in any two-hop neighborhood of a node. While collision among one-hop neighbors can be greatly reduced by carrier sensing before transmission, carrier sensing does not work beyond one hop. This problem, called the hidden terminal problem, causes serious throughput degradation, especially in high data rate sensor applications. Although RTS/CTS can alleviate the hidden terminal problem, it incurs high overhead, 40–75% of the channel capacity in WSNs, because data packets are typically very small (Polastre et al. 2004). TDMA, on the other hand, can solve the hidden terminal problem without extra message overhead because it can schedule transmission times of neighboring nodes to occur at different times. However, TDMA has many other disadvantages (Ye et al. 2004): • Finding an efficient time schedule in a scalable fashion is not trivial. It often requires a centralized node to find a collision-free schedule. Furthermore, developing an efficient schedule with a high degree of concurrency or channel reuse is very hard; the optimal solution is NP-hard (Ramanathan 1999). • TDMA needs clock synchronization. Although clock synchronization is an essential feature of many sensor applications, tight synchronization incurs high energy overhead because it requires frequent message exchanges. • WSNs may undergo frequent topology changes because of time-varying channel conditions, physical environmental changes, battery outage, and node failures. Handling dynamic topology changes is expensive, possibly requiring a global change. • It is difficult to ascertain the interference relation among neighboring nodes because radio interference ranges are different from communication ranges, and some interfering nodes may not be in a direct communication range. This phenomenon is known as interference irregularity (Zhou et al. 2004). Therefore, any channel assignment that uses the communication ranges, in place of interference ranges, for building the conflict relations does not necessarily yield an interference-free schedule. Furthermore, as interference ranges and channel conditions are highly time varying, it is unlikely that one fixed schedule is sufficient to prevent collision all of the time.

4.1 Duty-Cycling Approach Taxonomy

231

• During low contention, TDMA gives much lower channel utilization and higher delays than CSMA because in TDMA a node can transmit only during its scheduled time slots, whereas in CSMA nodes can transmit at any time as long as there is no contention. These difficulties with TDMA suggest that a stand-alone TDMA scheme is not practical. Even with an efficient TDMA schedule, the other factors such as interference irregularity, time-varying channel conditions, and clock synchronization errors would diminish the benefits of TDMA. Nevertheless, the information provided by an efficient TDMA schedule, in particular, the independent sets of nodes that can transmit concurrently, can be used in curtailing occurrences of collision, especially under high contention. This position greatly motivated Z-MAC (section “A Hybrid MAC for WSNs (Z-MAC)”). S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”) and T-MAC (section “An Adaptive Energy-Efficient MAC Protocol for WSNs (T-MAC)”) are a hybrid of CSMA and TDMA in that they also maintain the synchronized time slots. Unlike TDMA, their slots can be much bigger than normal TDMA slots and synchronization failures do not necessarily lead into communication failure because they employ RTS/CTS. Nodes maintain periodic duty-cycle to listen for channel activities and transmit data. As these protocols use RTS/CTS, the overhead of the protocols is quite high because most data packets in WSNs are small. T-MAC improves the energy efficiency of S-MAC by forcing all transmitting nodes to start transmission at the beginning of each active period. A Hybrid MAC for WSNs (Z-MAC) Zebra MAC (Z-MAC) is a hybrid MAC protocol for WSNs that combines the strengths of TDMA and CSMA while offsetting their weaknesses (Rhee et al. 2008). The main feature of Z-MAC is its adaptability to the level of contention in the network; under low contention, it behaves like CSMA, and under high contention it is like TDMA. It is also robust for dynamic topology changes and time synchronization failures commonly occurring in WSNs. Z-MAC is useful for applications where expected data rates and two-hop contention are medium to high. Z-MAC design principles involve conjoint ideas: • Z-MAC uses CSMA as the baseline MAC scheme and uses a TDMA schedule as a “hint” to enhance contention resolution. • A time slot assignment is performed at the time of deployment, and at the beginning the overhead is higher. The high initial overhead is amortized over the long period of network operation; eventually, it is compensated by improved throughput and energy efficiency. An efficient scalable channel-scheduling algorithm is used, DRAND (Rhee et al. 2006). • After slot assignment, each node reuses its assigned slot periodically in every predetermined period, called frame. A node assigned to a time slot is its owner, and the others are the non-owners of that slot. There can be more than one

232

4 Energy Management Techniques for WSNs …

owner per slot because DRAND allows any two nodes beyond their two-hop neighborhoods to own the same slot. • Unlike TDMA, a node may transmit during any time slot in Z-MAC. Before transmitting during a slot (not necessarily at the beginning), a node always performs carrier sensing and transmits a packet when the channel is clear. However, an owner of that slot always has higher priority over its non-owners in accessing the channel. The priority is implemented by adjusting the initial contention window size in such a way that the owners are always given earlier chances to transmit than non-owners. The goal is that, during the slots where owners have data to transmit, Z-MAC reduces the chance of collision since owners are given earlier chances to transmit and their slots are scheduled a priori to avoid collision, but when a slot is not in use by its owners, non-owners can steal the slot. This priority scheme affects the simplicity of switching between CSMA and TDMA depending on the level of contention. An important feature of this scheme is that the probability of owners accessing the channel can be adjusted independently from that of non-owners. This contributes to increasing the robustness of the protocol to synchronization and slot assignment failures while enhancing its scalability to contention. • By mixing CSMA and TDMA, Z-MAC becomes more robust to timing failures, time-varying channel conditions, slot assignment failures, and topology changes than a stand-alone TDMA; in the worst case, it always falls back to CSMA. • Since Z-MAC needs only local synchronization among senders in two-hop neighborhoods, a simple local synchronization scheme is devised where each sending node adjusts its synchronization frequency based on its current data rate and resource budget. Next, the design, implementation, and performance of Z-MAC will be presented in detail. The setup phase in Z-MAC has several operations to be run in sequence: typically, neighbor discovery and slot assignment, local frame exchange, and global time synchronization. These operations run only once during the setup phase and do not run until a significant change in the network topology occurs, such as physical relocation of sensors. The idea is that the initial upfront costs for running these operations are reduced by improved throughput and energy efficiency during data transmission. Z-MAC setup phase operations are integrated with transmission control scheme as afterward described. Several operations are included in Z-MAC setup phase: exclusively, neighbor discovery and slot assignment, local frame exchange, and global time synchronization. Details are provided in what follows: • Neighbor discovery and slot assignment. This operation goes through several essentials: – As a node starts up, it first runs a simple neighbor discovery protocol where it periodically broadcasts a ping to its one-hop neighbors to gather its one-hop neighbor list. A ping message contains the current list of its one-hop

4.1 Duty-Cycling Approach Taxonomy

233

neighbors. Each node sends one ping message at a random time in each second for 30 s. Through this process, each node gathers the information received from the pings from its one-hop neighbors, which essentially constitutes its two-hop neighbor information. – The two-hop neighbor list is used as input to a time slot assignment algorithm. Z-MAC uses DRAND (Rhee et al. 2006), a distributed implementation of RAND (Ramanathan 1999), to assign time slots to every node in the network. DRAND ensures a broadcast schedule where no two nodes within a two-hop communication neighborhood are assigned to the same slot. This assignment guarantees that no transmission by a node to any of its one-hop neighbors interferes with any transmission by its two-hop neighbors. A broadcast schedule can handle any routing changes among its one-hop neighbors. – DRAND is scalable because it does not depend on the network size, but on the local neighborhood size of each node. Z-MAC produces a very efficient time schedule where the slot number assigned to a node does not exceed the size of its local two-hop neighborhood (d). The running time and message complexity of DRAND are also bounded by O(d). Thus, its energy cost is linearly proportional to the size of the local neighborhood. • Local frame exchange. Once a node picks a time slot, it has to decide on the period in which it can use the time slot for transmission; this period is called the time frame of the node. Conventionally, all nodes must keep the same time frame, while synchronizing to have their time slots 0 at the same time. This process requires though to propagate the maximum slot number (MSN) to the entire network and is not adaptive to local time slot changes. When new nodes are added to the network, DRAND can run local slot assignment while maintaining the existing assignment. If this assignment causes the MSN to be changed, the change must be propagated again to the entire network, which incurs extra cost for adapting to a small change in the network topology. In Z-MAC, each node maintains its own local time frame that fits its local neighborhood size, while avoiding any conflict with its contending neighbors. The basic idea is as follows: – Time frame rule (TF rule). Let a node i be assigned to a slot according to DRAND and the MSN within its two-hop neighborhood be Fi. The ith time frame is set to be Li = 2a where a is a positive integer chosen to satisfy condition 2a1
Fi \2a  1; that is, i uses the sith slot in every Li time frame (i slots are l * Li + si, for all i = 1, 2, 3, …). The right-hand side of the inequality constrains the set of feasible values for a to avoid conflict, while the left-hand side of the inequality forces picking the minimum of these feasible values. The right-hand side of the inequality avoids conflict among contending neighbors. The TF rule allows nodes to pick their own time frame sizes based on their local two-hop information. This rule makes DRAND adaptive to dynamic

4 Energy Management Techniques for WSNs …

234 C 2(5)

0 1 2 3 4 5 6 7

E 1(5)

A 0(2) B 1(2)

D 0(5) H 5(5)

(a) Network topology

F 3(5)

G 4(5)

A B C D E F G H

(b) Slot schedule of all nodes

Legend: • In a) the numbers indicate the slot numbers assigned by DRAND, and the number in parenthesis are Fi. • In b) are slots where each node transmits, are the empty slots that are not used by one-hop and two-hop neighbors.

Fig. 4.53 Time frame rule (Rhee et al. 2008)

time frame changes, caused by local topology changes, without incurring any global changes. Figure 4.53 shows an example of a TDMA schedule obtained by the TF rule. If the global time frame is used, then 6 will be the time frame size. Then, nodes A and B can use their slots only once every 6 slots although their frame sizes are 2 each. But, if the TF rule is used, nodes are allowed to use frame size 4. This increases the concurrency in the channel usage and reduces the message delays for nodes A and B. However, slots 6 and 7 are not assigned to any node in the neighborhood. This is a tradeoff; when the network is uniformly dense, the global time frame would create a smaller number of empty slots. But, if the network contains many sparse areas with only a few dense areas, then the local framing would be more preferable. In Z-MAC, since empty slots are available for CSMA, they are not necessarily wasted. • Global time synchronization. Synchronizing on slot 0 is worth some attention; the local framing rule implicitly assumes that all nodes start their time slot 0 at the same time. This can be achieved without any communication, if clocks are synchronized, by fixing a predetermined absolute time to synchronize slot 0. For instance, the beginning of the real time, i.e., when the synchronized clock value is zero, can be set to be the beginning of slot 0. New nodes can easily synchronize their slots if they synchronize their clocks to the global clock. Done with the initial setup phase, transmission control of Z-MAC starts at the end of the DRAND phase, where every node forwards its frame size and slot number to its two-hop neighborhood. Thus, a node knows about the slot and frame information of its one-hop and two-hop neighbors at the beginning of the Z-MAC

4.1 Duty-Cycling Approach Taxonomy

235

phase. At this point, every node synchronizes to slot 0 and then nodes get ready to run the transmission control of Z-MAC. In Z-MAC, a node can be in one of two modes, whether low contention level (LCL) or high contention level (HCL): • A node is in HCL only when it receives an explicit contention notification (ECN) message from a two-hop neighbor within the last tECN period. Otherwise, the node is in LCL. A node sends an ECN when it experiences high contention. • In LCL, any node can compete to transmit in any slot, but in HCL, only the owners of the current slot and their one-hop neighbors are allowed to compete for the channel access. • In both modes, the owners have higher priority over non-owners. If a slot does not contain an owner, or its owner does not have data to send, non-owners can steal the slot. This feature achieves high channel utilization even under low contention as a node can transmit as soon as the channel is available. • Z-MAC implements LCL and HCL using the backoff CCA8 and LPL interfaces of B-MAC (section “Versatile Low-Power Media Access for Sensor Networks (B-MAC)”). The transmission process works according to a specific sequence. As a node i acquires data to transmit, it checks whether it is the owner of the current slot: • If it is the owner of the slot, it takes a random backoff within a fixed time period T0. When the backoff timer expires, it runs CCA, and if the channel is clear, it transmits the data. If the channel is not clear, then it waits until the channel is not busy and repeats the process. • If node i is a non-owner of the current slot and it is in LCL, or if it is in HCL and the current slot is not owned by its two-hop neighbors, then it waits for T0 and then performs a random backoff within a contention window [T0,Tn0]. When the backoff timer expires, it runs CCA, and if the channel is clear it starts transmission. If the channel is not clear, it waits until the channel is clear and repeats the process. • If node i is a non-owner of the current slot and is in HCL, meaning that a two-hop neighbor of i has sent an ECN in the last tECN, it postpones its transmission (it may sleep) until it finds a time slot that either is not owned by a two-hop neighbor or is its owner. After waking up, it repeats the above process. According to the above transmission rules, some functionality is worth noting: • In LCL mode, a node can compete in any slot, albeit with different priorities. • In HCL mode, a node can compete in the current slot only if it is the owner of the slot or a one-hop neighbor to the owner of that slot. It is possible for a transmission that has started in the previous slot to cross over to an HCL slot causing collision with the owner of the slot. One way to prevent this is to restrict a transmission not to cross over an HCL slot. Z-MAC does not to support this restriction as it makes the system design more complicated, especially in a network where tight time synchronization is difficult to achieve. Moreover, packets do not come at a regular interval and may not be of the same size.

236

4 Energy Management Techniques for WSNs …

• One more reason for allowing slot “crossing” is due to a tradeoff in channel utilization. If such a crossing is not allowed, then even when there is some remaining time in a slot, it might be unused if a packet transmission by the owner cannot be completed within that slot. On the other hand, if crossing is permitted, then it is possible that a packet transmission by the next owner, which could act as a hidden terminal to the current owner, could cause a collision. Thus, time is wasted for transmitting the packet. The tradeoff is whether to prevent such a collision by not transmitting during the remaining time in the slot and hence wasting that time, or to make the transmission during that time but possibly risking channel wastage due to a packet collision at the next slot. Both cases waste some amount of slot time, but in the first case, time is always wasted, while in the second case time is wasted only when a collision happens. Specific values of T0 and Tno have noticeable performance impact. The choice of T0 determines the robustness of Z-MAC against time synchronization errors or slot assignment failures, which cause some slots to have more than one owner. If the synchronization error is no more than one TDMA slot size, then there can be at most two to three conflicting owners at any time. Slot sizes have perceptible performance implications: • If the slot size is too small, clock synchronization errors will have higher performance impact, as more nodes will be allowed to overlap over slot boundaries. For slot size x ms, as long as the synchronization error is less than x=2 ms, a slot will have no more than two conflicting owners. Since the effect of clock synchronization errors will likely occur around the boundaries of slots, as the slot size increases, the performance impact of such errors asymptotically reduces, because within a unit time, the number of boundaries gets smaller. • Increasing the slot size tends to increase the transmission delay since it increases the frame size. If a node misses its time slot, it takes one frame size before it becomes an owner again. Therefore, the choice of the slot size should be a function of the accuracy of clock synchronization and also the desired network delay in the network. In Z-MAC, the slot size is a system parameter tunable depending on the application. Contention has a significant impact on WSN performance; thus, contention notification, a main Z-MAC attention, is explicitly handled. ECN messages notify two-hop neighbors not to act as hidden terminals to the owner of each slot when contention is high. Each node makes a local decision to send an ECN message based on its local estimate of the contention level. There are two ways to estimate two-hop contention: • Receiving acknowledgment from the one-hop receiver and measuring the packet loss rate. Since two-hop contention causes collision, it is highly related to the loss rate. However, this technique requires the receiver to send feedback and

4.1 Duty-Cycling Approach Taxonomy

237

incurs extra overhead. Unless the acknowledgment feature is enabled by the application, this overhead can unduly reduce the channel utilization. • Measuring the noise level of the channel. When high contention occurs, it tends to increase the noise level. This technique does not require any extra overhead as the noise level can be measured passively at the time of data transmission. In order to measure the noise level passively without actively sampling the channel, a measurement is performed for the average number of noise backoffs that a sender takes before transmitting a packet. A transmitter takes a noise backoff when it senses the channel using CCA before packet transmission (it transmits only when the channel is clear). When the noise level is higher than the CCA threshold, the node takes backoff. MICA2 (Crossbow 2002) experimentation measures the correlation between the noise level at a measurement node and two-hop contention at the sink, where two clusters of nodes transmit to a common receiver called sink. Under low transmission rates, despite increasing the number of senders, the average noise level and two-hop contention are very low. However, increasing the transmission rate to the full rate where all senders always have data to send results in an increased noise level. As a transmitting node detects high contention, it sends a unicast message, one-hop ECN, to a destination to which it is experiencing contention. If multiple destinations experience contention, it sends one broadcast with information about the multiple destinations. Typically, in WSNs, each node has a single transmit destination, one parent. The ECN message is thus maintained as below: • When a node j receives a one-hop ECN message triggered by its one-hop neighbor i, it first checks whether j is the destination of the ECN message. If so, it then broadcasts the ECN to its one-hop neighbors; these ECN messages are called two-hop ECN. If j is not the destination, it simply discards the one-hop ECN. • When a node receives a two-hop ECN, then it sets its HCL flag. The HCL flag is a soft state, meaning that, unless another two-hop ECN message is received within the last period, the flag is reset. Thus, if node i continually experiences contention, it transmits the ECN message periodically; this refresh period, tECN, is set by the system. Typically implementing: • When a node detects contention, it is likely that its neighboring senders will do so at the same time. Therefore, there are many duplicate ECN messages forwarded to routing nodes. To prevent ECN implosion, overhearing is used to suppress ECN. • When a node i detects high contention, it takes random backoffs before the transmission of a one-hop ECN message. In the meantime, if it receives a one-hop ECN intended for another node that has the same destination as i’s ECN, then node i suppresses its ECN and cancels the transmission of the ECN.

238

4 Energy Management Techniques for WSNs …

After tECN, if it still experiences high contention, it schedules another ECN by taking a random backoff and repeats the above process. • The same suppression rule applies to routing nodes. If a routing node receives a one-hop ECN and it has forwarded an ECN within tECN period, it does not forward a two-hop ECN. ECN is similar to RTS/CTS in CSMA/CA. Yet, there are some differences: • HCL uses topology information, i.e., slot information, to avoid two-hop collision. • The cost of ECN is far less than RTS/CTS, since it is triggered only when contention is high. • Using ECN suppression, only a small number of ECN messages need to be forwarded. As the HCL state may last for a much longer term than a single packet transmission, its cost is amortized over many packet transmissions. Z-MAC requires clock synchronization under high contention to implement HCL. Such synchronization is required only among neighboring senders that are under high contention. It is thus possible to optimize the overhead of clock synchronization because synchronization is required only locally among neighboring senders. Moreover, the frequency of synchronization can be adjusted according to the transmission rates of senders, such that those with higher data rates transmit more frequent synchronization messages. In this scheme, receivers passively synchronize their clocks to the senders’ clocks and do not have to send any synchronization messages. Studying a protocol provides an in-depth view to its features and performance metrics. Analytically, a closed-form expression of channel utilization is evaluated for various WSN MAC protocols, namely B-MAC, Sift (Jamieson et al. 2006), probabilistic TDMA (PTDMA) (Ephremides and Mowafi 1982), and Z-MAC, in a one-hop environment where all nodes can sense the transmission of the other nodes. S-MAC and T-MAC are not analyzed as their performance is lower than B-MAC (section “Versatile Low-Power Media Access for Sensor Networks (B-MAC)”). Analytical expressions are validated experimentally thru MICA2/TinyOS and by ns-2 simulation. ns-2 simulation compares Z-MAC performance with protocols whose TinyOS implementation was not available: specifically, PTDMA and Sift. Comparison against B-MAC was via ns-2 and TinyOS. The packet format of B-MAC is adopted as Z-MAC is built atop. The default initial and congestion backoff window sizes of B-MAC are 32 and 16 slots, respectively (each slot is 400 ls). Three benchmark setups are built for experimentation: namely, one hop, two hops, and multihop: • One-hop benchmark. In this benchmark, as in B-MAC, nodes placed in a circle, equidistant from a receiver, transmit as quickly as possible with full transmission power. Nodes are in a one-hop distance to each other so that there are no hidden terminals. The benchmark is used to measure the achievable throughput

4.1 Duty-Cycling Approach Taxonomy Table 4.8 Default settings of Z-MAC parameters (Rhee et al. 2008)

239

TinyOS and ns-2 experimentation parameters

Default

Owner contention window size (T0) Non-owner contention window size (Tn0) Contention window per-slot duration ECN refresh period (tECN) Z-MAC TDMA slot size Communication range (ns-2) Interference range (ns-2) Communication bandwidth (TinyOS, ns-2)

8 slots 32 slots 400 ls 10 s 50 ms 200 ft 300 ft 19.2 Kbps

of different MAC protocols at different levels of contention within this one-hop neighborhood. • Two-hop benchmark. This benchmark tests the performance of different protocols when hidden terminals are present. Nodes are organized into two clusters such that seven and eight sending nodes are located in each cluster. A receiver node (or routing node) is placed in the middle of the two clusters. • Multihop benchmark. Two multihop topologies are considered: a ten-hop chain topology and a full-edged WSN testbed comprising 42 MICA2 nodes. The ten-hop chain experiment, where 11 nodes are lined to create a line topology (Fig. 4.31), is reproduced as in B-MAC to measure the latency of different protocols. The intermediate nodes forward the messages to the sink. The routing paths are taken from one run of Mint (Woo et al. 2003), the default routing protocol of TinyOS. Table 4.8 lists some default settings of Z-MAC parameters. Experimentation and simulation are performed to analyze throughput, fairness, latency, and energy efficiency: • Throughput. In this experiment, the effective channel utilization of the considered MAC protocols is measured and compared under different scenarios: – One-hop benchmark. All senders are transmitting at their full transmission power, and the receiver has its radio ON always, i.e., no duty-cycle. The effective maximum data throughput on MICA2 is 15.6 Kbps (excluding preamble and sync bytes). For Z-MAC tests, the frame size is fixed at 20 slots for all experiments. HCL is disabled because the performance of HCL and LCL is the same when all nodes are at a one-hop distance to each other. Before running Z-MAC, DRAND and TPSN are run to get slot assignments and to synchronize the clocks of the senders. The findings were as itemized below (Fig. 4.54): The data throughput of Z-MAC with one sender is about 40% less than that of B-MAC with window sizes (0.16 slots) and (32.16 slots), since Z-MAC uses a larger congestion backoff window size. With one source, it sends as non-owner most of the time except for its own slot; therefore, it incurs the cost of waiting for the non-owner slots.

4 Energy Management Techniques for WSNs …

Fig. 4.54 Throughput comparison thru the one-hop MICA2 benchmark (Rhee et al. 2008)

Throughput (Kbps)

240

B-MAC Initial= 0, Congestion=16 B-MAC Initial=16, Congestion= 64 B-MAC Initial=32, Congestion=16 Z-MAC LCL, T0= 8, Tn0 =32 Z-MAC LCL unsynched T0= 8, Tn0 =32

Number of sources (contenders)

The throughput of Z-MAC is almost independent of the number of senders. When the number of senders is small, most senders are sending as non-owners; thus, they can utilize the unused slots that belong to other nodes. As the number of senders increases, so does the number of senders transmitting during their own slots. Thus, when contention is high, it can maintain good throughput since it works more like TDMA. The throughput with 20 senders is much higher than that of B-MAC. For Z-MAC with no clock synchronization, clock values of all nodes are randomized, and the local clock synchronization protocol is turned OFF. This allows some slots to be overlapped with each other so that several nodes consider themselves as owners at the same time. This scenario essentially emulates slot assignment failures. It is noticed that although the Z-MAC performance drops in the presence of time synchronization errors, it is not worse than CSMA, and under high contention it gives comparable performance to Z-MAC with synchronized clocks. This is because T0 is sufficiently large to handle multiple owners within a slot. Hence, when the information about interference relation and synchrony is inaccurate, the performance of Z-MAC gracefully degrades to that of CSMA; but, when this information is accurate, Z-MAC performs well under high contention. ns-2 simulation results involving PTDMA, Sift, B-MAC, and Z-MAC disclose higher steady Z-MAC utilization in the order of 0.6 for 1–20 sources. – Two-hop benchmark. The data throughput is measured when hidden terminals are present. The number of senders is varied, while the number of neighbors is fixed. As in the one-hop benchmark, all senders always have data to send. Each additional sender is chosen from the alternating clusters.

Fig. 4.55 Throughput comparison thru the two-hop MICA2 benchmark (Rhee et al. 2008)

Throughput (Kbps)

4.1 Duty-Cycling Approach Taxonomy

241 B-MAC Initial= 16, Congestion=64 B-MAC Initial=32, Congestion= 16 Z-MAC HCL, T0= 8, Tn0 =32 Z-MAC LCL, T0= 8, Tn0 =32 Z-MAC HCL unsynched T0= 8, Tn0 =32

Number of sources (contenders)

For Z-MAC tests, the frame size is set to 16 slots. Z-MAC is run with HCL disabled (Z-MAC LCL) and with HCL enabled (Z-MAC HCL). For both cases, the local clock synchronization protocol is such that each sender sends one synchronization packet in every 100 packets transmitted. The data throughput reported by Z-MAC includes the overhead of the clock synchronization and ECN: Out of the MICA2 experiment, since the transmission power is as low as 1.3 mW, the maximum achievable throughput also gets lower (Fig. 4.55). As the number of hidden terminals increases along with more senders, the throughput of LCL drops more than that of HCL. On the other hand, HCL performs relatively well by staying around 6 Kbps even under high contention. The protocols with RTS/CTS, such as S-MAC and B-MAC with RTS/CTS, achieve around 2 Kbps throughput, even when no hidden terminals are present (Polastre et al. 2004). This confirms that the overhead of ECN is much lower than that of RTS/CTS. B-MAC shows high sensitivity to hidden terminals as its throughput drops to 1 Kbps under high contention. Z-MAC HCL, when run under unsynchronized clocks, shows a drop in performance but is still better than B-MAC (6 Kbps throughput). The ns-2 runs comparing PTDMA, Sift, B-MAC, and Z-MAC show that Z-MAC HCL has a higher sustained performance independent of the number of senders (a 0.5 utilization). Z-MAC HCL performance degrades slightly from that in the one-hop ns-2 benchmark because nodes can compete only during their own slots and the slots of their one-hop neighbors, and also because of the overhead of ECN messages. Checking the effect of time synchronization errors on the performance of Z-MAC thru ns-2 shows a sustained utilization in the order of 0.6, until the drift error becomes larger than 1 ms/sec even without its local clock synchronization.

4 Energy Management Techniques for WSNs …

242

B-MAC Initial=32, Congestion= 16 Z-MAC HCL, T0= 8, Tn0 =32

Throughput (Kbps)

Fig. 4.56 Throughput comparison thru the multihop MICA2 benchmark (Rhee et al. 2008)

Transmission rate (packets/sec)

– Multihop benchmark. Measuring the data throughput of Z-MAC HCL and B-MAC thru the 42-node MICA2 testbed, it was realized that (Fig. 4.56): Under transmission rates less than 3.12 packets/s, both protocols deliver all the packets and achieve about the same throughput. B-MAC shows slightly better throughput than Z-MAC, because the backoff congestion window of Z-MAC for non-owners is larger than that of B-MAC (16 slots). The backoff value makes this increase since contention is low and most transmissions in Z-MAC are done as non-owners. As the transmission rate increases beyond 3 packets/s, Z-MAC achieves 20– 30% higher throughput than B-MAC. Under the full 50 packets/sec data rate, Z-MAC achieves about 7.2 Kbps, while B-MAC achieves about 5.2 Kbps. These figures are slightly higher than the values from the two-hop benchmark under low contention. This is because, as the network gets densely populated, nodes can sense each other well, so one-hop contention dominates two-hop contention. • Fairness. The fairness index (Jain et al. 1984) is measured for delivered packets of all senders. As the number of packets delivered to the sink is more uniformly distributed among all senders, the index approaches one. The fairness index is computed from the average number of packets delivered per sender within 10 s intervals. For the multihop MICA2 testbed, it was realized that: – Under low transmission rates, both Z-MAC HCL and B-MAC show high fairness. For a 0.5 up to 1.5 packets/sec transmission rate, a fairness index in the range 0.9 down to 0.8 is obtained.

4.1 Duty-Cycling Approach Taxonomy

243

– As the transmission rate increases, the fairness index drops more precipitously for B-MAC. Z-MAC shows about 40% higher fairness index than B-MAC, under the full transmission rate. For a 2 up to 32 packets/sec transmission rate, a fairness index in the range 0.8 down to 0.45 is obtained for Z-MAC, while for B-MAC it is in the range 0.75 down to 0.32. • Latency. The latency experiments in B-MAC are replicated using the multihop benchmark at a sending rate of one packet every 10 s. For Z-MAC, HCL is enabled; but at this source rate, ECN is never sent, giving the same result as Z-MAC LCL. TPSN is run at the beginning to synchronize the clocks of all nodes and to perform the latency measurement of each packet using the time stamps at the source and sink. Both B-MAC and Z-MAC are tested under LPL with 100 ms check interval and full duty-cycle. Both B-MAC and Z-MAC show very similar latencies in all tests. This indicates that the protocol overhead of Z-MAC is quite comparable to that of B-MAC. • Energy efficiency. The energy cost of the Z-MAC setup phase operations in the multihop testbed is listed in Table 4.9. A total of 7.22 J/node on average is consumed for the setup phase constituting about 0.03% of the total energy available per node having 2500 mAh and 3 V battery. Although DRAND and the other operations are not optimized for energy saving, this is still a substantial amount of energy consumption compared to the per transmission energy cost. However, this upfront energy cost is later compensated by increased energy efficiency during the regular transmission of Z-MAC. From the MICA2 testbeds, by comparing B-MAC and Z-MAC it was found that: – Z-MAC uses the CCA and LPL features of B-MAC; thus, its energy efficiency is not better under low-data applications. Using the one-hop MICA2 testbed and varying the transmission rate, the optimal check interval for the traffic pattern is computed. The power consumption of Z-MAC is slightly worse than that of B-MAC. This is because in Z-MAC, nodes tend to wake up longer for transmission since their backoff window sizes are larger; moreover, clock synchronization messages are periodically sent. In this test, as data rates are low, all nodes are in LCL and no overhead for ECN is incurred.

Table 4.9 Average energy consumption during the setup operations in the multihop MICA2 testbed (Rhee et al. 2008)

Operation Neighbor discovery DRAND Local frame exchange TPSN Total Legend Identifiers on each operation acquiring a radio sample in Fig. 4.47

Average (J) 0.73 4.88 1.33 0.28 7.22 map to the activities of

244

4 Energy Management Techniques for WSNs …

– From the 42-node multihop MICA2 testbed, for each sending rate, the duty-cycle was varied from 20 to 60% and the energy efficiency measured in terms of throughput over power. Several findings were achieved: Under low data rates (up to 2 packets/s), B-MAC has slightly higher throughput. Also, B-MAC has slightly less power consumption (up to 10%). Explicably, as B-MAC has a smaller contention window size than Z-MAC, its idle time is less under low transmission rates. As the transmission rate increases beyond 3 packets/s, Z-MAC energy efficiency improves and exceeds that of B-MAC by about 40% at full rate. This higher energy efficiency is attributable to the efficiency in the contention resolution of Z-MAC HCL.

Wrapping up, Z-MAC is a hybrid MAC protocol for WSNs that combines the strengths of TDMA and CSMA while offsetting their weaknesses. Like CSMA, Z-MAC achieves high channel utilization and low latency under low contention; like TDMA, it achieves high channel utilization under high contention and reduces collision among two-hop neighbors at a low cost. Z-MAC uses the knowledge of topology and loosely synchronized clocks as hints to improve MAC performance under high contention. Z-MAC is useful for applications where expected data rates and two-hop contention are medium to high. Appraisal of MAC Protocols with Low Duty-Cycle MAC protocols for WSNs can be broadly divided into TDMA-based and contention-based protocols, and in between there are the hybrid MAC protocols. Going through the TDMA, contention, and hybrid protocols presented in Sect. 4.1.2.2, the following features are weighted. TDMA-based MAC protocols are centered around reservation and scheduling. Several characteristics are accorded to these protocols: • They have a natural advantage of energy conservation compared to contention protocols, because the duty-cycle of the radio is reduced and there is no contention-introduced overhead and collisions. They are inherently energy-efficient, as nodes turn ON their radio only during their own slots and sleep for the rest of the time. By an appropriate design of the slot assignment algorithm, and a correct sizing of the protocol parameters, it is possible to minimize energy consumption. • However, using TDMA protocol usually requires the nodes to form real communication clusters, like Bluetooth (Haartsen 2000) and LEACH (Heinzelman et al. 2000). Most nodes in a real cluster are restricted to communicate within the cluster. Managing intercluster communication and interference is not an easy task. Moreover, when the number of nodes within a cluster changes, it is not easy for a TDMA protocol to dynamically change its frame length and time slot

4.1 Duty-Cycling Approach Taxonomy

245

assignment. So, its scalability is normally not as good as that of a contention-based protocol; typically, Bluetooth may have at most eight active nodes in a cluster. • They need tight synchronization and are very sensitive to interference (Anastasi et al. 2005). • TDMA-based protocols perform worse than contention-based protocols in low traffic conditions. Regarding the above downsides, TDMA-based MAC protocols are not very frequently used in practical WSNs (Anastasi et al. 2009). Contention-based MAC protocols revolve around a different concept. The standardized IEEE 802.11 distributed coordination function (DCF) (LAN/MAN Standards Committee 1997) is an example of contention-based protocol and is mainly built on the research protocol MACAW (Bharghavan et al. 1994). It is widely used in WSNs because of its simplicity and robustness to the hidden terminal problem. However, it was shown that the energy consumption using this MAC is very high when nodes are in idle mode, and this is mainly due to the idle listening (Stemm and Katz 1997). IEEE 802.11 has a power saving mode. Power-aware multi-access protocol with signaling (PAMAS) (Singh and Raghavendra 1998) made an improvement on energy savings by trying to avoid the overhearing among neighboring nodes. S-MAC also exploits the same idea (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”). The main difference between S-MAC and PAMAS is that S-MAC does not use any out-of-channel signaling, whereas PAMAS requires two independent radio channels, which in most cases indicate two independent radio systems on each node. PAMAS does not attempt to reduce idle listening. Contention-based MAC protocols have their specific features: • They are robust and scalable. • They generally introduce a lower delay than TDMA-based protocols and can easily adapt to traffic conditions. • Unluckily, their energy expenditure is higher than TDMA-based MAC protocols because of contention and collisions. Duty-cycle mechanisms can help reducing the energy wastage, but they need to be designed carefully to be adaptive and low in latency. An emerging area of interest consists in using features of industry standards such as the IEEE 802.15.4. The core functions provided by the standard can be used as a basis for developing extensions targeted to a specific scenario, but based on the same specifications. Finally, hybrid protocols combine the strengths of TDMA-based and contention-based MAC protocols, while offsetting their weaknesses. However, these techniques seem to be too complex to be feasible in deployments with a high number of nodes. Solutions such as those presented in Zheng et al. (2005) and Halkes and Langendoen (2007) provide simple slot allocation mechanisms, and a low protocol overhead represents a promising direction in the field of energyefficient MAC protocols for WSNs.

246

4 Energy Management Techniques for WSNs …

Although a low duty-cycle MAC protocol is energy-efficient, it has three side effects: • Increasing the packet delivery latency. At a source node, a sampling reading may occur during the sleep period and has to be queued until the active period. An intermediate node may have to wait until the receiver wakes up before it can forward a packet received from its previous hop. This is called sleep latency in S-MAC (section “Medium Access Control with Coordinated Adaptive Sleeping for WSNs (S-MAC)”), and it increases proportionally with hop length by a slope of schedule length (active period plus sleep period). • Non-adaptivity to the varying traffic rate in sensor network. A fixed duty-cycle for the highest traffic load results in significant energy wastage when traffic is low, while a duty-cycle for low traffic load results in low message data delivery and long queuing delay. Therefore, it is desirable to adapt the duty-cycle under variant traffic load. • Increasing the possibility of collision in fixed synchronous duty-cycle. If neighboring nodes turn to active state at the same time; they all may contend for the channel, making a collision very likely.

4.2

Conclusion for Longer Duty-Cycling

Plenteous classifications and protocols build this chapter, section by section, paragraph following paragraph, figure after figure. To have it done, I had to devote more time, to focus; I augmented my work hours. My duty-cycle was lengthened to be eight hours of work per day, five hours of them for this chapter. The time-consuming self-talking about what to add, how to reshuffle, and organize was excluded from the duty-cycle, from the eight hours!! As abundantly detailed, duty-cycling can be achieved through two different and yet complementary approaches: • Topology control by finding the optimal subset of nodes that guarantee connectivity. This scheme exploits node redundancy, which is typical in WSNs, and adaptively selects only a minimum subset of nodes to remain active for maintaining connectivity. Nodes that are not currently needed for ensuring connectivity can go to sleep and save energy. Therefore, the basic idea behind topology control is to exploit the network redundancy to prolong the network longevity, typically increasing the network lifetime by a factor of 2–3 with respect to a network with all nodes always ON. Topology control protocols encompass location-driven protocols and connectivity-driven protocols. • Power management by operating duty-cycling on active nodes. Active nodes are those selected by the topology control protocol; they do not need to maintain their radio continuously ON. They can switch OFF the radio, by putting it in the low-power sleep mode, when there is no network activity, thus alternating between sleep and wakeup periods. Power management techniques can be

4.2 Conclusion for Longer Duty-Cycling

247

further subdivided into two broad categories depending on the layer of the network architecture they work on: – Independent sleep/wakeup protocols running atop of a MAC protocol, typically at the network or application layer. This permits a greater flexibility as they can be tailored to the application needs and can be used with any MAC protocol. – Protocols strictly integrated with the MAC protocol itself, which permits to optimize medium access functions based on the specific sleep/wakeup pattern used for power management. The introduction of energy harvesting capabilities into WSNs introduces many design questions about the construction of such networks. For example, how can a node use its harvesting abilities intelligently to increase its task performance and lifetime? Can a node adapt its power consumption profile online so as to subsist indefinitely on a given energy source without compromising its task performance? Such questions are meaningful for dynamically adapting the duty-cycle of a node so as to maximize both lifetime and performance. The concept of “energy neutral operation” is introduced to be the condition where the energy consumed by the node is always less than or equal to the energy harvested from the environment (Kansal et al. 2007). It is a non-trivial task to design duty-cycling algorithms that achieve energy neutral operation, since the dynamics of the energy sources being harvested may not be easily predictable. Approaches to the dynamic duty-cycling of nodes with energy harvesting capabilities attempt to model the energy source and adjust the node duty-cycle in anticipation of expected incoming energy or lack thereof (Kansal et al. 2007). These methods also attempt to handle stochasticity in the energy profile by comparing observed energy input with expected input given by the model and adjusting the duty-cycle accordingly. There is also a model-free approach using techniques from adaptive control theory (Vigorito et al. 2007). In addition to energy neutral operation, some WSN applications require the duty-cycling behavior of the node to be as “stable” as possible, meaning that it should have low variance over time. For instance, nodes involved in event monitoring tasks must minimize their sleep time in order to detect fleeting and unpredictable events and report them with low latency (Dutta et al. 2005); in such cases, one would ideally want the node to be awake at any given point in time. The duty-cycle of a node may also be important for communication between sensor nodes; for instance, B-MAC (section “Versatile Low-Power Media Access for Sensor Networks (B-MAC)”) uses the characteristics of a neighboring node duty-cycle to determine the length of a packet preamble. Along this chapter, duty-cycling is methodically described and categorized, and protocols are analyzed and compared. Table 4.10 abundantly compiles and compares the techniques presented all over. The ending exercises unfold and enhance research topics, besides instigating implementation projects.

Duty-cycling techniques (Sect. 4.1)

• Location-driven (Sect. 4.1.1.1): – GAF (conserves energy by identifying nodes that are equivalent from a routing perspective and turning OFF unnecessary nodes, keeping thus a constant level of routing fidelity. It is independent of the routing protocol) – GeRaF (transmission scheme based on geographical routing where packets are relayed on a best effort. Each node has some knowledge of its own position and of the position of the sink) • Connectivity-driven (Sect. 4.1.1.2): – Span (runs above the link and MAC layers and interacts with the routing protocol) – ASCENT (independent of the routing protocol. It limits the packet loss due to collisions because the node density is explicitly taken into account as a parameter. Also, it has good scalability properties) – Naps (decentralized topology management protocol based on a periodic sleep/wakeup scheme) – Uncoordinated power saving mechanisms with latency considerations (asynchronous sleep/wakeup protocol with less restrictive connectivity for time-critical applications) – DDEMA (a fully distributed energy management protocol for largescale WSNs; it allows each sensor to schedule its own activity based on its node degree, without knowledge of global network parameters)

Taxonomy Topology control protocols (Sect. 4.1.1)

Table 4.10 Duty-cycling techniques classified Power management protocols (Sect. 4.1.2)

(continued)

• Sleep/wakeup protocols (Sect. 4.1.2.1): – On-demand schemes: STEM (allows to efficiently trade one design constraint (energy) for the other (latency). It may also be combined with techniques that leverage increased network density to obtain energy savings. It uses two different radio channels, one for wakeup and the other for data) PTW (achieves a tradeoff between energy saving and wakeup latency. It relies on two different radio channels for transmitting wakeup signals and data packets, and uses a wakeup tone to awake neighboring nodes) – Scheduled rendezvous schemes: Wakeup scheduling patterns in WSNs (minimizes the worst-case end-to-end overall delay, which includes both transmission delay and detection delay under different wakeup patterns) Optimal wakeup scheduling of data gathering trees for WSNs (assigns a different wakeup frequency for each node to minimize the total power consumption in the data gathering tree, and ensures delivery delays from any node to the gateway node) – Asynchronous schemes: AWP (based on the optimal results obtained from the block design problem, AWP is an asynchronous wakeup protocol that detects neighboring nodes in finite time without requiring slot alignment. AWP is also resilient to packet collision and network dynamics) RAW (it consists of a routing protocol combined with a random wakeup scheme. The routing protocol is a variant of geographic routing. While in geographic routing a packet is sent to a neighbor that is closest to the destination, in RAW the packet is sent to any of the active neighbors in the FCS, i.e., the set of active neighbors that meet a prespecified criterion)

248 4 Energy Management Techniques for WSNs …

Taxonomy Topology control protocols (Sect. 4.1.1)

Table 4.10 (continued)

(continued)

• MAC protocols with low duty-cycle (Sect. 4.1.2.2): – TDMA-based MAC protocols: TRAMA (a schedule-based MAC protocol that reduces energy consumption by ensuring that unicast and broadcast transmissions incur no collisions, and by allowing nodes to assume a low-power, idle state, whenever they are not transmitting or receiving. Adequate throughput and fairness are achieved by means of an inherently fair transmitter-election algorithm that promotes channel reuse as a function of the competing traffic around any given source or receiver) L-MAC (designed to account for the physical layer properties in WSN nodes. Its goal is minimizing the number of transceiver switches, to make the sleep interval for sensor nodes adaptive to the amount of data traffic and to limit the complexity of implementation) FLAMA (a schedule-based MAC protocol that leverages traffic predictability in WSN applications. As the message ow in periodic reporting applications is stable, FLAMA sets up ows and then uses a pull-based mechanism, so that data are transferred only after being explicitly requested) – Contention-based protocols: S-MAC (MAC reduces energy waste, while tolerating some performance reduction in both per-hop fairnessand latency. S-MAC establishes low duty-cycle operation on nodes in a multihop network; it reduces idle listening by periodically putting nodes into sleep state) T-MAC (based on reducing idle listening by transmitting all messages in bursts of variable length, and sleeping between bursts. T-MAC improves on S-MAC energy usage by using a very short

Power management protocols (Sect. 4.1.2)

4.2 Conclusion for Longer Duty-Cycling 249

Taxonomy Topology control protocols (Sect. 4.1.1)

Table 4.10 (continued)

listening window at the beginning of each active period. After the SYNC section of the active period, there is a short window to send or receive RTS and CTS packets. If no activity occurs in that period, the node returns to sleep. By changing the protocol to have an adaptive duty-cycle, T-MAC saves power at the cost of reduced throughput and additional latency) D-MAC (an energy-efficient and low-latency MAC protocol designed and optimized for unidirectional data gathering trees. DMAC solves the interruption problem and adjusts node duty-cycles adaptively according to the traffic load in the network; it also proposes a data prediction mechanism and the use of MTS packets in order to alleviate problems pertaining to channel contention and collisions) B-MAC (a exible MAC protocol, shipped with TinyOS. It features a simple, predictable, yet scalable implementation and is tolerant to network changes. B-MAC effectively performs clear channel estimation. At its core, it exceeds the performance of other protocols through reconfiguration, feedback, and bidirectional interfaces for higher-layer services. B-MAC may be configured to run at extremely low duty-cycles and does not force applications to incur the overhead of synchronization and state maintenance) – Hybrid MAC protocols: Z-MAC (a hybrid MAC protocol for WSNs that combines the strengths of TDMA and CSMA while offsetting their weaknesses. Z-MAC uses the knowledge of topology and loosely synchronized clocks as hints to improve MAC performance under high contention. Z-MAC is useful for applications where expected data rates and two-hop contention are medium to high)

Power management protocols (Sect. 4.1.2)

250 4 Energy Management Techniques for WSNs …

4.3 Exercises

4.3 1. 2. 3. 4. 5. 6. 7.

8. 9.

10.

11. 12. 13. 14. 15. 16. 17.

251

Exercises Write a technical report on the topology control protocols (Sect. 4.1.1). Write a technical report on the power management protocols (Sect. 4.1.2). Identify and compare the approaches based on the percolation theory. How an energy management scheme can be routing protocol dependent or independent? Identify and compare the energy conservation schemes that are routing protocol dependent. Identify and compare the energy conservation schemes that are routing protocol independent. Scheduled rendezvous schemes require nodes to be synchronized in order to wake up at the same time (section “Scheduled Rendezvous Schemes”). Clock synchronization in WSNs is a significant research and project topic. Search the literature for such clock synchronization schemes, and write a comparative survey report. Based on Fig. 4.19, propose an asynchronous wakeup schedule. Referring to Sect. “4.1.2.2” identify and compare contention-based and scheduled-based MAC protocols, with special emphasis on energy management issues. Compare D-MAC (section “An Adaptive Energy-Efficient and Low-Latency MAC for Data Gathering in WSNs (D-MAC)”) and B-MAC (section “Versatile Low-Power Media Access for Sensor Networks (B-MAC)”) protocols. Based on section “A Hybrid MAC for WSNs (Z-MAC),” write a technical report on Sift. Based on section “A Hybrid MAC for WSNs (Z-MAC),” write a technical report on PTDMA. Based on section “A Hybrid MAC for WSNs (Z-MAC),” write a technical report on the fairness index and its use. Based on Sect. “4.2,” write a technical report on dynamic duty-cycling. Based on Sect. “4.2,” write a technical report o energy neutral operation. Dig the literature for more protocols that may fit in the energy management taxonomy elaborated in this chapter. Write a technical report on the duty-cycling approach taxonomy (Sect. “4.1”).

References Al-Karaki, J.N., and A.E. Kamal. 2004. “Routing Techniques in Wireless Sensor Networks: A Survey.” Wireless Communications (IEEE) 11 (6): 6–28. Anastasi, G., E. Borgia, M. Conti, E. Gregori, and A. Passarella. 2005. “Understanding the Real Behavior of Mote and 802.11 Ad Hoc Networks: An Experimental Approach.” Pervasive and Mobile Computing 1 (2): 237–256. Anastasi, G., M. Conti, M. Di Francesco, and A. Passarella. 2009. “Energy Conservation in Wireless Sensor Networks: A Survey.” Ad Hoc Networks 7 (3): 537–568.

252

4 Energy Management Techniques for WSNs …

Ansari, J., D. Pankin, and P. Mähönen. 2009. “Radio-Triggered Wake-Ups with Addressing Capabilities for Extremely Low Power Sensor Network Applications.” International Journal of Wireless Information Networks 16 (3): 118. Arisha, K., M. Youssef, and M. Younis. 2002. “Energy-Aware TDMA-Based MAC for Sensor Networks.” In System-Level Power Optimization for Wireless Multimedia Communication, ed. R. Karri and D. Goodman, 21–40. Boston, MA: Springer. Atmel. 8-bit Atmel Microcontroller with 128KBytes in-System Programmable Flash. Atmel. January 1, 2011. http://www.atmel.com/Images/doc2467.pdf. Accessed 21 Mar 2014. Bachir, A., M. Dohler, T. Watteyne, and K.K. Leung. 2010. “MAC Essentials for Wireless Sensor Networks.” IEEE Communications Surveys & Tutorials (IEEE) 12 (2): 222–248. Bao, L., and L.L. Garcia-Luna-Aceves. 2001. “A New Approach to Channel Access Scheduling for Ad Hoc Networks.” In The 7th Annual International Conference on Mobile Computing and Networking (MobiCom), 210–221. Rome, Italy: ACM. Bao, L., and J.J. Garcia-Luna-Aceves. 2003. “Topology Management in Ad Hoc Networks.” In The 4th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 129–140. Annapolis, MD: ACM. Bharghavan, V., A. Demers, S. Shenker, and L. Zhang. 1994. “MACAW: A Media Access Protocol for Wireless LAN’s.” In The Conference on Communications Architectures, Protocols and Applications (SIGCOMM), 212–225. London, UK: ACM. Cao, Q., T. Abdelzaher, T. He, and J. Stankovic. 2005. “Towards Optimal Sleep Scheduling in Sensor Networks for Rare-Event Detection.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN), Los Angeles, CA: ACM. Casari, P., A. Marcucci, M. Nati, C. Petrioli, and M. Zorzi. 2005. “A Detailed Simulation Study of Geographic Random Forwarding (GeRaF) in Wireless Sensor Networks.” In Military Communications Conference (MILCOM), 59–68. Atlantic City, NJ: IEEE. Cerpa, A., and D. Estrin. 2004. “ASCENT: Adaptive Self-Configuring Sensor Networks Topologies.” Transactions on Mobile Computing (IEEE) 3 (3): 272–285. Chen, B., K. Jamieson, H. Balakrishnan, and R. Morris. 2002. “Span: An Energy-Efficient Coordination Algorithm for Topology Maintenance in Ad Hoc Wireless Networks.” Wireless Networks 8 (5): 481–494. Cohen, R., and B. Kapchits. 2009. “An Optimal Wake-Up Scheduling Algorithm for Minimizing Energy Consumption While Limiting Maximum Delay in a Mesh Sensor Network.” Transactions on Networking (TON) 17 (2): 570–581. Cooklev, T. 2004. Wireless Communication Standards: A Study of IEEE 802.11, 802.15, and 802.16. NY: Standards Information Network - IEEE Press. Crossbow. January 1, 2002. MICA2. http://www.eol.ucar.edu/isf/facilities/isa/internal/CrossBow/ DataSheets/mica2.pdf. Accessed 3 Feb 2014. Crossbow. January 1, 2002. MICA2DOT. http://www.eol.ucar.edu/isf/facilities/isa/internal/ CrossBow/DataSheets/mica2dot.pdf. Accessed 5 Feb 2014. Dai, H., and R. Han. 2004. “TSync: A Lightweight Bidirectional Time Synchronization Service for Wireless Sensor Networks.” ACM SIGMOBILE Mobile Computing and Communications Review (ACM) 8 (1): 125–139. Demirkol, I., C. Ersoy, and F. Alagoz. 2006. “MAC Protocols for Wireless Sensor Networks: A Survey.” Communications Magazine (IEEE) 44 (4): 115–121. Dictionary.com. January 1, 2017. Anisotropic. Dictionary.com, LLC. http://www.dictionary.com/ browse/anisotropy. Accessed 8 Apr 2017. Dousse, O., P. Mannersalo, and P. Thiran. 2004. “Latency of Wireless Sensor Networks with Uncoordinated Power Saving Mechanisms.” In The 5th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 109–120. Tokyo, Japan: ACM. Dousse, O., F. Baccelli, and P. Thiran. 2005. “Impact of Interferences on Connectivity in Ad Hoc Networks.” Transactions on Networking (TON) 13 (2): 425–436. Dousse, O., M. Franceschetti, N. Macris, and R. Meester. 2006. “Percolation in the Signal to Interference Ratio Graph.” Journal of Applied Probability 43 (2): 552–562.

References

253

Dulman, S.O., and P.J.M Havinga. 2003. A Simulation Template for Wireless Sensor Networks, vol. 6. Enschede, The Netherlands: University of Twente, Centre for Telematica and Information Technology (CTIT). Dutta, P., M. Grimmer, A. Arora, S. Bibyk, and D. Culler. 2005. “Design of a Wireless Sensor Network Platform for Detecting Rare, Random, and Ephemeral Events.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN). Los Angeles, CA: ACM SIGBED and IEEE Signal Processing Society. Elson, J., L. Girod, and D. Estrin. 2002. “Fine-Grained Network Time Synchronization using Reference Broadcasts.” In The 5th Symposium on Operating Systems Design and Implementation (OSDI), 147–163. Boston, MA: ACM. Engstrom, J.C., and C. Gray. 2008. “Clear channel assessment in wireless sensor networks.” In The 46th Annual Southeast Regional Conference (ACM-SE), 464–468. Auburn, AL: ACM. Ephremides, A., and O.A. Mowafi. 1982. “Analysis of A Hybrid Access Scheme for Buffered Users-Probabilistic Time Division.” Transactions on Software Engineering SE-8 (1): 52–61. Fahmy, H.M.A. 2016. “Simulators and Emulators for WSNs.” In Wireless Sensor Networks: Concepts, Applications, Experimentation and Analysis, 381–491. Singapore: Springer. Faizulkhakov, Y.R. 2007. “Time Synchronization Methods for Wireless Sensor Networks: A Survey.” Programming and Computer Software 33 (4): 214–226. Fang, Q., J. Gao, and L.J. Guibas. 2006. “Locating and Bypassing Holes in Sensor Networks.” Mobile Networks and Applications 11 (2): 187–200. Ganeriwal, S., R. Kumar, and M.B. Srivastava. 2003. “Timing-Sync Protocol for Sensor Networks.” In The 1st International Conference on Embedded Networked Sensor Systems (SenSys), 138–149. Los Angeles, CA: ACM. Ganesan, D., D. Estrin, A. Woo, and D. Culler. 2002. Complex Behavior at Scale: An Experimental Study of Low-Power Wireless Sensor Networks. Los Angeles, CA: Technical, Computer Science Department, University of California Los Angeles (UCLA). Ganesan, D., A. Cerpa, W. Ye, Y. Yu, J. Zhao, and D. Estrin. 2004. “Networking Issues in Wireless Sensor Networks.” Journal of Parallel and Distributed Computing 64 (7): 799–814. Garey, M.R., and D.S. Johnson. 1979. Computers and Intractability: A Guide to the Theory of NP-Completeness. New York: W.H. Freeman and Company. Godfrey, P.B., and D. Ratajczak. 2004. “Naps: Scalable, Robust Topology Management in Wireless Ad Hoc Networks.” In The 3rd International Symposium on Information Processing in Sensor Networks (IPSN), 443–451. Berkeley, CA: ACM. Grimmett, G. 1999. “What is Percolation?” In Percolation, ed. G. Grimmett, 1–31. Berlin: Springer. Gu, L., and J.A. Stankovic. 2005. “Radio-Triggered Wake-Up for Wireless Sensor Networks”. Real-Time Systems 29 (2): 157–182. Gupta, P., and P.R. Kumar. 1999. “Critical power for asymptotic connectivity in wireless networks.” In Stochastic Analysis, Control, Optimization and Applications, ed. W.M. McEneaney, G.G. Yin, and Q. Zhang, 547–566. Cambridge, MA: Birkhäuser Boston. Gupta, P., and P.R. Kumar. 2000. “The Capacity of Wireless Networks.” Transactions on Information Theory (IEEE) 46 (2): 388–404. Haartsen, J.C. 2000. “The Bluetooth Radio System.” Personal Communications 7 (1): 28–36. Halkes, G.P., and K.G. Langendoen. 2007. “Crankshaft: An Energy-Efficient MAC-Protocol for Dense Wireless Sensor Networks.” In The 4th European Conference on Wireless Sensor Networks (EWSN), 228–244. Delft, The Netherlands: Springer. Havinga, P., et al. 2003. “EYES—Energy Efficient Sensor Networks.” In Lecture Notes in Computer Science: Personal Wireless Communications (PWC), vol. 2775, ed. M. Conti, S. Giordano, E. Gregori, and S. Olariu, 198–201. Berlin, Heidelberg: Springer. He, T., J.A. Stankovic, C. Lu, and T. Abdelzaher. 2003. “SPEED: A Stateless Protocol for Real-Time Communication in Sensor Networks.” In The 23rd International Conference on Distributed Computing Systems (ICDCS), 46–55. Providence, RI: IEEE.

254

4 Energy Management Techniques for WSNs …

Heinzelman, W.R., A. Chandrakasan, and H. Balakrishnan. 2000. “Energy-Efficient Communication Protocol for Wireless Microsensor Networks.” In The 33rd Annual Hawaii International Conference on System Sciences (HICSS). Maui, Hawaii: IEEE. Hill, J.L., and D.E. Culler. 2002. “Mica: A Wireless Platform for Deeply Embedded Networks.” IEEE Micro 22 (6): 12–24. Intanagonwiwat, C., D. Estrin, R. Govindan, and J. Heidemann. 2002. “Impact of Network Density on Data Aggregation in Wireless Sensor Networks.” In The 22nd International Conference on Distributed Computing Systems, 457–458. Vienna, Austria: IEEE. Jain, R.K., D.-M. Chiu, and W.R. Hawe. 1984. A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer Systems. Hudson, MA: Technical, Eastern Research Lab., Digital Equipment Corporation. Jamieson, K., H. Balakrishnan, and Y.C. Tay. 2006. “Sift: A MAC protocol for event-driven wireless sensor networks.” In Lecture Notes in Computer Science: European Workshop on Wireless Sensor Networks (EWSN), vol. 3868, ed. K. Römer, H. Karl, and F. Mattern, 260– 275. Berlin, Heidelberg: Springer. Johnson, D.B., and D.A. Maltz. 1996. “Dynamic Source Routing in Ad Hoc Wireless Networks.” In Mobile Computing, vol. 353, ed. T. Imielinski and H.F. Korth, 153–181. Berlin: Springer. Kansal, A., J. Hsu, S. Zahedi, and M.B. Srivastava. 2007. “Power Management in Energy Harvesting Sensor Networks.” In ACM Transactions on Embedded Computing Systems (TECS) —Special Section LCTES’05 (p. 6, no. 4). ACM. Karl, H., and A. Willig. 2007. Protocols and Architectures for Wireless Sensor Networks. Chichester, West Sussex: Wiley. Karp, B., and H.T. Kung. 2000. “GPSR: Greedy Perimeter Stateless Routing for Wireless Networks.” In The 6th Annual International Conference on Mobile Computing and Networking (MobiCom), 243–254. Boston, MA: ACM. Keshavarzian, A., H. Lee, and L. Venkatraman. 2006. “Wakeup Scheduling in Wireless Sensor Networks.” In The 7th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 322–333. Florence, Italy: ACM. Koushanfar, F., N. Taft, and M. Potkonjak. 2006. “Sleeping Coordination for Comprehensive Sensing Using Isotonic Regression and Domatic Partitions.” In The 25th IEEE International Conference on Computer Communications (INFOCOM), 1–13. Barcelona, Spain: IEEE. Krishnamachari, L., D. Estrin, and S. Wicker. 2002. “The Impact of Data Aggregation in Wireless Sensor Networks.” In The 22nd International Conference on Distributed Computing Systems Workshops (ICDCS), 575–578. Vienna, Austria: IEEE. Lai, W., and I. Paschalidis. 2006. “Routing Through Noise and Sleeping Nodes in Sensor Networks: Latency vs. Energy Trade-Offs.” In The 45th IEEE Conference on Decision and Control (CDC), 2716–2721. San Diego, CA: IEEE. LAN MAN Standards Committee. 1997. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specification. New York, NY: IEEE Std 802.11, LAN MAN Standards Committee, IEEE Computer Society. Langendoen, K., and N. Reijers. 2003. “Distributed Localization in Wireless Sensor Networks: A Quantitative Comparison.” Computer Networks 43 (4): 499–518. Lee, M., and V.W.S. Wong. 2005. “An Energy-Aware Spanning Tree Algorithm for Data Aggregation in Wireless Sensor Networks.” In The Pacific Rim Conference on Communications, Computers and Signal Processing (PACRIM), 300–303. Victoria, B.C., Canada: IEEE. Li, J., and G.Y. Lazarou. 2004. “A Bit-Map-Assisted Energy-Efficient MAC Scheme for Wireless Sensor Networks.” In The 3rd International Symposium on Information Processing in Sensor Networks (IPSN), 55–60. Berkeley, CA: ACM. Li, J., C. Blake, D.S.J. De Couto, H.I. Lee, and R. Morris. 2001. “Capacity of Ad Hoc Wireless Networks.” In The 7th Annual International Conference on Mobile Computing and Networking (MobiCom), 61–69. Rome, Italy: ACM.

References

255

Li, Y., W. Ye, and J. Heidemann. 2005. “Energy and Latency Control in Low Duty Cycle MAC Protocols.” In: Wireless Communications and Networking Conference (WCNM). New Orleans, LA: IEEE. Liggett, T.M. 1985. “An Improved Subadditive Ergodic Theorem.” The Annals of Probability 13 (4): 1279–1285. Liu, X. 2012. “A Survey on Clustering Routing Protocols in Wireless Sensor Networks.” Sensors (MDPI) 12 (8): 11113–11153. Lu, G., B. Krishnamachari, and C.S. Raghavendra. 2004. “An Adaptive Energy-Efficient and Low-Latency MAC for Data Gathering in Wireless Sensor Networks.” In The 18th International Parallel and Distributed Processing Symposium (IPDPS). Santa Fe, NM: IEEE. Lu, G., N. Sadagopan, B. Krishnamachari, and A. Goel. 2005. “Delay Efficient Sleep Scheduling in Wireless Sensor Networks.” In The 24th Annual Joint Conference of the IEEE Computer and Communications Societies (Infocom), 2470–2481. Miami, FL: IEEE. Mainwaring, A., D. Culler, J. Polastre, R. Szewczyk, and J. Anderson. 2002. “Wireless Sensor Networks for Habitat Monitoring.” In The 1st ACM International Workshop on Wireless Sensor Networks and Applications (WSNA), 88–97. Atlanta, GA: ACM. Malesci, U., and S. Madden. 2006. “A measurement-based analysis of the interaction between network layers in TinyOS.” In Lecture Notes in Computer Science: Wireless Sensor Networks (EWSN), vol. 3868, ed. K. Römer, H. Karl, and F. Mattern, 292–309. Berlin, Heidelberg: Springer. Mangione-Smith, W., P. Ghang, S. Nazareth, P. Lettieri, W. Boring, and R. Jain. 1996. “A Low Power Architecture for Wireless Multimedia Systems: Lessons Learned from Building A Power Hog.” In International Symposium on Low Power Electronics and Design (ISLPED), 23–28. Monterey, CA: ACM. Mao, G., B. Fidan, and B.D.O. Anderson. 2007. “Wireless Sensor Network Localization Techniques.” Computer Networks 51 (10): 2529–2553. McGlynn, M.J., and S.A. Borbash. 2001. “Birthday Protocols for Low Energy Deployment and Flexible Neighbor Discovery in Ad Hoc Wireless Networks.” In The 2nd ACM International Symposium on Mobile Ad Hoc Networking & Computing (MobiHoc), 137–145. Long Beach, CA: ACM. Meester, R., and R. Roy. 1996. Continuum Percolation. New York, NY: Cambridge University Press. Melodia, T., D. Pompili, and I.F. Akyildiz. 2005. “On the Interdependence of Distributed Topology Control and Geographical Routing in Ad Hoc and Sensor Networks.” Journal on Selected Areas in Communications (IEEE) 23 (3): 520–533. Miller, M.J., and N.H. Vaidya. 2004. “Power Save Mechanisms for Multi-Hop Wireless Networks.” In First International Conference on Broadband Networks, 518–526. San Jose, CA: IEEE. Murata Electronics. December 31, 2014. TR1001: 868.35 MHz Hybrid Transceiver. Murata Electronics N.A., Inc. http://wireless.murata.com/RFM/data/tr1001.pdf. Accessed 27 Sept 2017. Murata Electronics. December 31, 2015a. TR3000: 433.92 MHz Hybrid Transceiver. Murata Electronics N.A., Inc. http://wireless.murata.com/datasheet?/RFM/data/tr3000.pdf?ref=rfm. com. Accessed 25 Sept 2017. Murata Electronics. January 1, 2015b. TR1000: 916.50 MHz Hybrid Transceiver. Murata Electronics N.A., Inc. http://wireless.murata.com/RFM/data/tr1000.pdf. Accessed 17 Mar 2017. Nieberg, T., S. Dulman, P. Havinga, L. van Hoesel, and J. Wu. 2003. “Collaborative Algorithms for Communication in Wireless Sensor Networks.” In Ambient Intelligence: Impact on Embedded System Design, ed. T. Basten, M. Geilen, and H. de Groot, 271–294. Boston, MA: Springer. NIST/SEMATECH. October 30, 2013. What Are Outliers in the Data? NIST/SEMATECH. http:// www.itl.nist.gov/div898/handbook/prc/section1/prc16.htm. Accessed 24 Oct 2017.

256

4 Energy Management Techniques for WSNs …

Pantazis, N.A., and D.D. Vergados. 2007. “A Survey on Power Control Issues in Wireless Sensor Networks.” Communications Surveys & Tutorials 9 (4): 86–107. Paruchuri, V., S. Basavaraju, A. Durresi, R. Kannan, and S.S. Iyengar. 2004. “Random Asynchronous Wakeup Protocol for Sensor Networks.” In First International Conference on Broadband Networks (BroadNets), 710–717. San Jose, CA: IEEE. Penrose, M. 2003. Random Geometric Graphs. New York: Oxford University Press. Perkins, C., E. Belding-Royer, and S. Das July 1, 2003. RFC 3561-Ad Hoc On-Demand Distance Vector (AODV) routing. Network Working Group. http://www.rfc-editor.org/rfc/pdfrfc/ rfc3561.txt.pdf. Accessed 15 July 2015. Polastre, J., J. Hill, and D. Culler. 2004. “Versatile Low Power Media Access for Wireless Sensor Networks.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 95–107. Baltimore, MD: ACM. Proakis, J.G., and M. Salehi. 2007. Digital Communications. New York, NY: McGraw-Hill Education. Raghunandan, G.H., and B.N. Lakshmi. 2011. “A Comparative Analysis of Routing Techniques for Wireless Sensor Networks.” In National Conference on Innovations in Emerging Technology (NCOIET), 17–22. Erode, India: IEEE. Raghunathan, V., C. Schurgers, S. Park, and M.B. Srivastava. 2002. “Energy-Aware Wireless Microsensor Networks.” Signal Processing Magazine 19 (2): 40–50. Rajendran, V., J.J. Garcia-Luna-Aveces, and K. Obraczka. 2005. “Energy-Efficient, Application-Aware Medium Access for Sensor Networks.” In International Conference on Mobile Ad Hoc and Sensor Systems Conference (MASS), 623–630. Washington, DC: IEEE. Rajendran, V., K. Obraczka, and J.J. Garcia-Luna-Aceves. 2006. “Energy-Efficient, Collision-Free Medium Access Control for Wireless Sensor Networks”. Wireless Networks 12 (1): 63–78. Ramanathan, S. 1999. “A Unified Framework and Algorithm for Channel Assignment in Wireless Networks.” Wireless Networks 5 (2): 81–94. Rao, A., S. Ratnasamy, C. Papadimitriou, S. Shenker, and I. Stoica. 2003. “Geographic Routing Without Location Information.” In The 9th Annual International Conference on Mobile Computing and Networking (MobiHoc), 96–108. San Diego, CA: ACM. Raynal, M. 2013. Concurrent Programming: Algorithms, Principles, and Foundations. Berlin, Heidelberg: Springer. RFM. May 19, 2004. ASH Transceiver Designer’s Guide. https://wireless.murata.com/RFM/data/ tr_des24.pdf. Accessed 23 Nov 2017. Rhee, I., A. Warrier, J. Min, and L. Xu. 2006. “DRAND: Distributed Randomized Tdma Scheduling for Wireless Ad-Hoc Networks.” In The 7th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 190–201. Florence, Italy: ACM. Rhee, I., A. Warrier, M. Aia, J. Min, and L.S. Mihail. 2008. “Z-MAC: A Hybrid MAC for Wireless Sensor Networks.” IEEE/ACM Transactions on Networking (TON) 16 (3): 511–524. Santi, P. 2005. “Topology Control in Wireless Ad Hoc and Sensor Networks.” Computing Surveys (CSUR) 37 (2): 164–194. Schurgers, C., V. Tsiatsis, S. Ganeriwal, and M. Srivastava. 2002. “Optimizing Sensor Networks in The Energy-Latency-Density Design Space.” Transactions on Mobile Computing 99 (1): 70–80. Shih, E., P. Bahl, and M.J. Sinclair. 2002. “Wake on Wireless: An Event Driven Energy Saving Strategy for Battery Operated Devices.” The 8th Annual International Conference on Mobile Computing and Networking (MobiCom), 160–171. Atlanta, GA: ACM. Singh, S., and C.S. Raghavendra. 1998. “PAMAS: Power Aware Multi-Access Protocol with Signalling for Ad Hoc Networks.” Computer Communication 28 (3): 5–26. Sinha, A., and A. Chandrakasan. 2001. “Dynamic Power Management in Wireless Sensor Networks.” Design & Test of Computers 18 (2): 62–74. Sivrikaya, F., and B. Yener. 2004. “Time Synchronization in Sensor Networks: A Survey.” Network 18 (4): 45–50. SparkFun Electronics. January 1, 2016. The Basics of GPS. SparkFun Electronics. https://learn. sparkfun.com/tutorials/gps-basics. Accessed 16 Aug 2016.

References

257

Stemm, M., and R.H. Katz. 1997. “Measuring and Reducing Energy Consumption of Network Interfaces in Hand-Held Devices.” Transactions on Communications E80-B (8): 1125–1131. Tang, L., Y. Sun, O. Gurewitz, and D.B. Johnson. 2011. “PW-MAC: An Energy-Efficient Predictive-Wakeup MAC Protocol for Wireless Sensor Networks.” In The 30th IEEE International Conference on Computer Communications (INFOCOM), 1305–1313. Shanghai, China: IEEE. Texas Instruments. 2007. CC1000: Single Chip Very Low Power RF Transceiver. Texas Instruments. http://www.ti.com/lit/ds/symlink/cc1000.pdf. Accessed 23 Dec 2013. Texas Instruments. January 1, 2005. CC1100: Low-Cost Low-Power Sub-1 GHz RF Transceiver. Texas Instruments. http://datasheet.octopart.com/CC1100RTKR-Texas-Instruments-datasheet10422951.pdf. Accessed 3 Feb 2014. Texas Instruments. August 1, 2004. MSP430x11x2, MSP430x12x2 Mixed Signal Microcontroller. Texas Instruments, Inc. http://www.ti.com/lit/ds/symlink/msp430f1122.pdf. Accessed 8 Oct 2016. TinyOS. August 20, 2012. TinyOS. http://www.tinyos.net. Accessed 9 Oct 2013. Tseng, Y.-C., C.-S. Hsu, and C.-S. Hsu. 2003. “Power-Saving Protocols for IEEE 802.11-Based Multi-Hop Ad Hoc Networks.” Computer Networks 43 (3): 317–337. Ungjin, J., L. Sunggu, and S. Yoo. 2012. “Optimal Wake-Up Scheduling of Data Gathering Trees for Wireless Sensor Networks.” Journal of Parallel and Distributed Computing 72 (4): 536– 546. Van Dam, T., and K. Langendoen. 2003. “An Adaptive Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In The 1st International Conference on Embedded Networked Sensor Systems (SenSys), 171–180. Los Angeles, CA: ACM. van Hoesel, L.F.W., and P.J.M. Havinga. 2004. “A Lightweight Medium Access Protocol (LMAC) for Wireless Sensor Networks: Reducing Preamble Transmissions and Transceiver State Switches.” In The 1st International Workshop on Networked Sensing Systems (INSS), 205–208. Tokyo, Japan: Society of Instrument and Control Engineers (SICE). Vasudevan, S., J. Kurose, and D. Towsley. 2005. “On Neighbor Discovery in Wireless Networks with Directional Antennas.” In The 24th Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), 2502–2512. Miami, FL: IEEE. Vigorito, C.M., D. Ganesan, and A.G. Barto. 2007. “Adaptive Control of Duty Cycling in Energy-Harvesting Wireless Sensor Networks.” In The 4th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and Networks, 21–30. San Diego, CA: IEEE. Want, R. 2006. “An Introduction to RFID Technology.” Pervasive Computing 5 (1): 25–33. Warrier, A., S. Park, J. Min, and I. Rhee. 2007. “How Much Energy Saving Does Topology Control Offer for Wireless Sensor Networks?—A Practical Study.” Computer Communications 30 (14/15): 2867–2879. Woo, A., T. Tong, and D. Culler. 2003. “Taming the Underlying Challenges of Reliable Multihop Routing in Sensor Networks. In The 1st International Conference on Embedded Networked Sensor Systems (SenSys), 14–27. Los Angeles, CA: ACM. Xiao, Y., and J. Rosdahl. 2002. “Throughput and Delay Limits of IEEE 802.11.” Communications Letters 6 (8): 355–357. Xu, Y., J. Heidemann, and D. Estrin. 2000. Adaptive Energy-Conserving Routing for Multihop Ad Hoc Networks. Marina del Rey, CA: Information Sciences Institute, University of Southern California. Xu, Y., J. Heidemann, and D. Estrin. 2001. “Geography-Informed Energy Conservation for Ad Hoc Routing.” In The 7th Annual International Conference on Mobile Computing and Networking (MobiCom), 70–84. Rome, Italy: ACM. Xu, Y., S. Bien, Y. Mori, J. Heidemann, and D. Estrin. 2003. Topology Control Protocols to Conserve Energy in Wireless Ad Hoc Networks. Los Angeles, CA: Technical, Center for Embedded Network Sensing (CENS), University of California.

258

4 Energy Management Techniques for WSNs …

Yang, X., and N.H. Vaidya. 2004. “A Wakeup Scheme for Sensor Networks: Achieving Balance Between Energy Saving and End-To-End Delay.” In The 10th IEEE Real-Time and Embedded Technology and Applications Symposium (RTAS), 19–26. Toronto, Canada: IEEE. Ye, W., and J. Heidemann. 2004. “Medium Access Control in Wireless Sensor Networks.” In Wireless Sensor Networks, ed. C.S. Raghavendra, K.M. Sivalingam, and T. Znati, 73–91. Boston, MA: Springer. Ye, W., J. Heidemann, and D. Estrin. 2002a. A Flexible and Reliable Radio Communication Stack on Motes. USC/ISI Technical Report ISI-TR-565. Los Angeles, CA: Information Sciences Institute, University of Southern California. Ye, W., J. Heidemann, and D. Estrin. 2002b. “An Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In The 21st Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), 1567–1576. New York, NY: IEEE. Ye, W., J. Heidemann, and D. Estrin. 2004. “Medium Access Control with Coordinated Adaptive Sleeping for Wireless Sensor Networks.” IEEE/ACM Transactions on Networking 12 (3): 493– 506. Zheng, R., J.C. Hou, and L. Sha. 2003. “Asynchronous Wakeup for Ad Hoc Networks.” In The 4th ACM International Symposium on Mobile Ad hoc Networking and Computing (MobiHoc), 35– 45. Annapolis, MD: ACM SIGMOBILE. Zheng, T., S. Radhakrishnan, and V. Sarangan. 2005. “PMAC: An Adaptive Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In 19th IEEE International Parallel and Distributed Processing Symposium. Denver, CO: IEEE. Zheng, T., S. Radhakrishnan, and V. Sarangan. 2012. “A Switch Agent for Wireless Sensor Nodes with Dual Interfaces: Implementation and Evaluation.” Tsinghua Science and Technology 17 (5): 586–598. Zhou, G., T. He, S. Krishnamurthy, and J.A. Stankovic. 2004. “Impact of Radio Irregularity on Wireless Sensor Networks.” In The 2nd International Conference on Mobile Systems, Applications, and Services (MobiSys), 125–138. Boston, MA: ACM. Zorzi, M., and R.R. Rao. 2003a. “Geographic Random Forwarding (Geraf) for Ad Hoc and Sensor Networks: Energy and Latency Performance.” Transactions on Mobile Computing 2 (4): 349– 365. Zorzi, M., and R.R. Rao. 2003b. “Geographic Random Forwarding (Geraf) for Ad Hoc and Sensor Networks: Multihop Performance.” Transactions on Mobile Computing 2 (4): 337–348.

Chapter 5

Energy Management Techniques for WSNs (2): Data-Driven Approach

Management is a rule of prolonged success …

5.1

Data-Driven Approach Taxonomy

Datasheets of commercial tiny sensor nodes show that data communication is very expensive in terms of energy consumption, whereas data processing consumes significantly less energy. The energy cost of receiving or transmitting a single bit of information is approximately the same as that required by the processing unit for executing a thousand operations. On the other hand, the energy consumption of the sensing unit depends on the specific sensor type. Thus, to extend the lifetime of a WSN, most of the energy conservation schemes proposed in the literature aim to minimize the energy consumption of the communication unit. In this context, data-driven approaches can be divided according to the problem they address (Fig. 5.1). Specifically classifying (Anastasi et al. 2009): • Data reduction schemes that address the scenario of unneeded samples; they target reducing the amount of data to be delivered to the sink node, but via different principles. Explicitly, these are in-network processing, data compression, and data prediction. • Energy-efficient data acquisition schemes aiming to reduce the energy spent by the sensing subsystem. However, some of them can reduce the energy spent for communication as well. Data reduction and energy-efficient data acquisition techniques are meticulously detailed in Sects. 5.1.1 and 5.1.2, respectively. A comparative assembly of the techniques offered throughout this chapter is made available in Table 5.11.

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_5

259

260

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Data-driven approach

Data reduction Sect. 5.1.1

Energy-efficient data acquisition Sect. 5.1.2

Fig. 5.1 Data-driven approach taxonomy (Anastasi et al. 2009)

5.1.1

Data Reduction Protocols

Data reduction schemes address the scenario of unneeded samples; via different principles, they target reducing the amount of data to be delivered to the sink node (Fig. 5.2): • In-network processing performs data aggregation, such as computing the average of some values, at intermediate nodes between the sources and the sink (Sect. 5.1.1.1). Hence, the amount of data are reduced while traversing the network toward the sink. Aggregation techniques can be roughly classified into two categories; specifically, structure-based techniques, in which all nodes periodically report to the sink (Madden et al. 2002), or structure-free techniques, in which data aggregation is performed without explicit maintenance of a structure (Fan et al. 2007). A discussion and classification of aggregation approaches can be found in Fasolo et al. (2007) and Jesus et al. (2015). The most appropriate in-network processing technique depends on the specific application and must be tailored to its requirements. • Data compression can be applied to reduce the amount of information sent by source nodes; this scheme involves encoding information at the nodes generating data and decoding it at the sink (Sect. 5.1.1.2). Nonetheless, compressing

Data reduction

In-network processing

Data compression

Fig. 5.2 Data reduction approach taxonomy (Anastasi et al. 2009)

Data prediction

5.1 Data-Driven Approach Taxonomy

261

data can be a valuable help in power saving only if the execution of compression algorithms does not require an amount of energy greater than the one saved in reducing transmission. Compression prior to transmission in wireless battery-powered devices may actually cause an overall increase of power consumption, if no energy awareness is introduced, as compression algorithms are aimed at saving storage and not energy (Barr and Asanović 2006). • Data prediction builds an abstraction of a sensed phenomenon, a model describing data evolution (Sect. 5.1.1.3). The model can predict, within certain error bounds, the values the sensor nodes obtain; it resides both at the sensors and at the sink. If the needed accuracy is satisfied, queries issued by users can be evaluated at the sink through the model without the need to get the exact data from nodes. On the other hand, explicit communication between sensor nodes and the sink is needed when the model is not accurate enough; i.e., the actual sample has to be retrieved, and/or the model has to be updated. Data prediction reduces the number of information sent by source nodes and the energy needed for communication as well.

5.1.1.1

In-Network Processing Protocols

Battery power is the most limiting factor in designing WSN protocols. Therefore, to reduce power consumption, several mechanisms are made available throughout the literature and comprehensively described in this book, such as duty-cycling, data-driven, and mobility approaches. In-network processing through data aggregation protocols aims to combine and summarize data packets of several sensor nodes to reduce data transmission. An example of data aggregation scheme is presented in Fig. 5.3 where a group of sensor nodes collect information from a target region. When the basestation queries the network, instead of sending each sensor node data to the basestation, one of the sensor nodes, called data aggregator, collects the information from its neighboring nodes, aggregates them, e.g., computes the average, and sends the aggregated data to the basestation over a multihop

Data aggregator

Target region Fig. 5.3 Data aggregation in a WSN (Ozdemir and Xiao 2009)

Basestation

262

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

path. As illustrated, data aggregation reduces the number of data transmissions, thereby improving the bandwidth and energy utilization in the network. In WSNs, the benefit of data aggregation increases if the intermediate sensor nodes perform data aggregation incrementally when data are being forwarded to the basestation. While increasing network lifetime, data aggregation protocols may degrade important QoS metrics in WSNs, such as data accuracy, latency, fault tolerance, and security. Therefore, the design of an efficient data aggregation protocol is an inherently challenging task because the protocol designer must weight energy efficiency, data accuracy, latency, and fault tolerance, against security (Akkaya et al. 2008). In order to achieve this tradeoff, data aggregation techniques are tightly coupled with how packets are routed through the network. Hence, the architecture of the WSN plays a vital role in the performance of different data aggregation protocols. The majority of WSN applications require a certain level of security; thus, it is infeasible to aggregate data at the expense of security. In addition, there is a tough conflict between security and data aggregation protocols. Security protocols require sensor nodes to encrypt and authenticate any sensed data prior to transmission and push toward data decryption at the basestation (Hu and Evans 2003). On the other hand, data aggregation protocols need plain data to implement data aggregation at every intermediate node so that energy efficiency is maximized. Moreover, data aggregation causes alterations in sensor data, and therefore, it is a challenging task to provide security and data authentication along with data aggregation. Due to these conflicting goals, data aggregation and security protocols must be designed together so that data aggregation can be performed without sacrificing security. The necessity of implementing data aggregation and security together have led many researchers to work on secure data aggregation (Castelluccia et al. 2009; Ozdemir and Xiao 2009; Rezvani et al. 2015). Several protocols allow routing and aggregation of data packets simultaneously. These protocols can be categorized as tree-based, cluster-based, hybrid tree/ cluster-based, multi-path-based, and hybrid tree/multipath-based as to be detailed in Sections “Tree-Based Data Aggregation Protocols”, “Cluster-Based Data Aggregation Protocols”, “Hybrid Tree/Cluster-Based Data Aggregation Protocols”, “MultipathBased Data Aggregation Protocols”, and “Hybrid Tree/Multipath-Based Data Aggregation Protocols”, respectively.

Tree-Based Data Aggregation Protocols The simplest way to achieve distributed data aggregation is to determine some data aggregator nodes in the network and ensure that the data paths of sensor nodes include these data aggregator nodes (Xu et al. 2001; Lindsey et al. 2002; Intanagonwiwat et al. 2003). As Fig. 5.4 illustrates, the main issue of tree-based data aggregation protocols is the construction of an energy-efficient data aggregation tree. According to the tree-based approach, a shortest-path tree (SPT), a spanning tree, rooted at the sink is first constructed. Subsequently, such a structure is exploited in answering queries generated by the sink through performing

5.1 Data-Driven Approach Taxonomy

263

Basestation

Fig. 5.4 Tree-based data aggregation (Ozdemir and Xiao 2009)

in-network aggregation along the aggregation tree by proceeding level by level from its leaves to its root. Thus, as two or more messages get to a given node, their aggregate can be computed exactly. Many routing strategies are based on a hierarchical organization of the network nodes (Al-Karaki and Kamal 2004; Akkaya and Younis 2005). The simplest way to aggregate data flowing from the sources to the sink is electing some special nodes to function as aggregation points, and defining a preferred direction to be followed when forwarding data. In addition, a node may be marked as special depending on many factors like its position within the data gathering tree (Solis and Obraczka 2005), its resources (Erramilli et al. 2004), the type of data stored in its queue (Ding et al. 2003), or the processing cost due to aggregation procedures (Luo et al. 2006). However, this method of aggregation has some drawbacks, as actual WSNs are not free from failures. More precisely, when a packet is lost at a given level of the tree, e.g., due to channel impairments, the data coming from the related subtree are lost as well. In spite of the potentially high cost of maintaining a hierarchical structure in dynamic networks and the scarce robustness of the system in case of link/device failures, these approaches are particularly suitable to design optimal aggregation functions and perform efficient energy management. In some research, the sink organizes routing paths to evenly and optimally distribute the energy consumption while favoring the aggregation of data at the intermediate nodes (Ding et al. 2003; Harris et al. 2007; Gupta et al. 2008; Al-Karaki et al. 2009). Several techniques are embraced in the tree-based data aggregation approach: • Greedy incremental tree (GIT) (Intanagonwiwat et al. 2002) is a data-centric routing protocol that allows data aggregation using directed diffusion (Intanagonwiwat et al. 2000, 2003). In Krishnamachari et al. (2002), GIT is compared with two other data-centric routing schemes, namely center at nearest

264

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

source (CNS) (Akkaya et al. 2008) and shortest-path tree (SPT) (Intanagonwiwat et al. 2000). The simulation results have shown that GIT performs best in terms of average number of transmissions. • Tiny aggregation service (TAG) is a data-centric data aggregation framework built upon SPT routing (Madden et al. 2002). TAG is specifically designed for monitoring applications and allows an adjustable sleep schedule for sensor nodes; by permitting parent nodes, let their children know about the waiting time for transmission. Also, parent nodes cache their children data to prevent from data loss. TAG performs data aggregation in two phases: – Distribution phase, where the basestation queries are disseminated to the sensor nodes. The basestation broadcasts a message asking the sensor nodes to organize a routing tree so that it can send the queries. Each message has a field that specifies the level or distance from the root of the sending node (the level of the root is equal to zero). When a node that does not belong to any level receives this message, it sets its own level by incrementing the current level in the message by one and assigns the sender as its parent. This process continues until all sensor nodes in the network join the tree and have a parent. This messaging is repeated periodically to keep the tree structure updated. Once the tree is formed, the basestation queries the network via the aggregation tree. Sensor nodes use their parents when replying to basestation queries. – Collection phase, where the aggregated sensor readings are routed up the aggregation tree to the basestation. TAG employs an SQL-like language to query the network. Each query specifies the quantity that needs to be collected, the aggregation function and the sensor nodes that need to perform the data collection. Another SPT-based data aggregation protocol that promotes the parent energy awareness is proposed in Ding et al. (2003). In this protocol, parent selection depends on sensor nodes distance to the basestation and their residual energy level. • Directed diffusion is a reactive data-centric protocol that encompasses three phases: typically, interest dissemination, gradient setup, and path reinforcement and forwarding (Intanagonwiwat et al. 2003). The directed diffusion protocol as below described is illustrated in Fig. 5.5: – Interest dissemination phase. The basestation propagates an interest message describing the type of data that needs to be collected and the operational mode for the collection. Upon reception of the interest message, each sensor node rebroadcasts it to its neighbors. – Gradient setup phase. Sensor nodes also prepare interest gradients, which are basically the vectors containing the next hop that propagates the result of the query back to the basestation. For each type of data, a different gradient may be set up. – Path reinforcement and forwarding phase. At the end of gradient setup phase for a certain type of data, only a single path is used to route packets toward

5.1 Data-Driven Approach Taxonomy Basestation

Basestation

Source

265 Basestation

Source

Source

Legend: If the basestation sends an interest that reaches sensor nodes A and B, and both forward the interest to sensor node C, then node C sets up two vectors indicating that the data matching the interest must be returned to A and/or B.

Fig. 5.5 Directed diffusion (Ozdemir and Xiao 2009)

the sink. Data aggregation is performed during data forwarding phase. The basestation periodically refreshes the data gathering tree created by the reinforced paths. However, this is an expensive operation and it may overcome the data aggregation gain if the network topology is dynamic. A modified version of directed diffusion, called enhanced directed diffusion (EDD), integrates directed diffusion with a cluster-based architecture so that the efficiency of the local interactions during gradient setup phase increases (Zhou et al. 2004). Along the same way goes another protocol (Lee and Wong 2005). • Power-efficient gathering in sensor information systems (PEGASIS) organizes sensor nodes in a chain for the purpose of data aggregation (Lindsey et al. 2002). In PEGASIS, each data aggregation chain has a leader responsible for transmitting aggregated data to the basestation. In order to evenly distribute the energy expenditure in the network, sensor nodes take turns acting as the chain leader. The chain forming can be achieved either in centralized manner by the basestation or in a decentralized manner by using a greedy algorithm at each sensor node. Both approaches require the global knowledge of the network. The chain-building process starts from the sensor node farthest from the basestation and continues toward the basestation. When a node stops, the chain is reconstructed to bypass the worn-out node. In a sensor node chain, each sensor node receives data from a neighbor and aggregates it with its own reading, by generating with the data it received, a single packet that has the same length. This process is repeated along the chain, and the leader adds its own data into the packet and sends it directly to the basestation. It should be noted that node i will be in some random position j on the chain. Thus, the leader in each round of communication will be at a random position on the chain, which is essential to make the sensor network robust to node failures. Two major drawbacks of PEGASIS have been observed. First, PEGASIS requires each sensor node to have a complete view of the network topology so that chains can be formed properly, and all nodes must be able to transmit directly to the basestation. Second, if the distances between sensor nodes in a chain are too large, then the energy expenditure of sensor nodes can be significantly high.

266

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

• Energy-aware distributed aggregation tree (EADAT) relies on an energy-aware distributed heuristic (Ding et al. 2003). The basestation is the root of the aggregation tree; hence, it initiates the tree forming by broadcasting a control message composed of five fields, explicitly stating, ID, parent, power, status, and hop-count. The control message is forwarded among sensor nodes until each node broadcasts the message once and the outcome is an aggregation tree rooted at the basestation. By considering the energy level of sensor nodes, the algorithm gives higher chance to sensor nodes with higher residual power to become a non-leaf tree node. Therefore, the data forwarding task is performed by the sensor nodes that have high energy levels. Simulation results have shown that EADAT prolongs network lifetime and saves more energy in comparison with routing methods without aggregation. It is also observed that the average energy level of sensor nodes decreases much more slowly compared to the scenario without data aggregation. • The delay-bounded medium access control (DBMAC) integrates routing and MAC protocols to perform data aggregation (Di Bacco et al. 2004). The main objective of the proposed DBMAC scheme is both to minimize the latency for delay-bounded applications and to increase energy efficiency by taking advantage of data aggregation mechanisms. DBMAC employs a carrier sense multiple access with collision avoidance (CSMA/CA) that depends on a request-to-send/ clear-to-send/data/ acknowledgment (RTS/CTS/DATA/ACK) handshake. By taking advantage of CTS messages of other nodes, sensor nodes can select the relay node among those nodes that already have some packets to transmit in their queue. This process increases the data aggregation efficiency in the network as all the information stored along the path is aggregated into a single data packet. DBMAC is a brilliant instance of how routing and data aggregation may influence each other by showing that energy-efficient data aggregation solutions are obtained by a cross-layer design.

Cluster-Based Data Aggregation Protocols In cluster-based data aggregation protocols, sensor nodes are subdivided into clusters. In each cluster, a cluster head is elected in order to aggregate data locally and transmit the aggregation result to the basestation. Cluster heads can communicate with the sink directly via long-range radio transmission. However, this is quite inefficient for energy-constrained sensor nodes. Thus, cluster heads usually form a tree structure to transmit aggregated data by multi-hopping through other cluster heads, which results in significant energy savings. Figure 5.6 clarifies how is cluster-based data aggregation implemented. The pros and cons of the cluster-based schemes are close to those in the tree-based approach. Numerous cluster-based data aggregation protocols are presented in the literature:

5.1 Data-Driven Approach Taxonomy

267

Cluster head Basestation

Cluster head Cluster Cluster head Cluster Cluster Fig. 5.6 Cluster-based data aggregation (Ozdemir and Xiao 2009)

• Low-energy adaptive clustering hierarchy (LEACH) is a self-organizing and adaptive clustering protocol; it takes advantage of randomization to evenly distribute the energy expenditure among the sensor nodes (Heinzelman et al. 2002). LEACH is a clustered approach where cluster heads act as data aggregation points. The protocol consists of two phases; in the first phase, cluster heads are elected and cluster structures are formed, and then, in the second phase, cluster heads aggregate and transmit data to the basestation: – LEACH cluster head election process is based on a distributed probabilistic approach. In each data aggregator selection round, sensor nodes calculate the threshold T(n):
T ð nÞ ¼

P 1PðR mod ð1=pÞÞ

0

if n 2 G; otherwise

ð5:1Þ

where P is the desired percentage of cluster heads, R is the round number, and G is the set of nodes that have not been cluster heads during the last 1=P rounds.

In order to be a cluster head, sensor node n picks a random number between [0,1] and becomes a cluster head if this number is lower than T(n). Cluster head advertisements are broadcasted to sensor nodes, and sensor nodes join the clusters based on the signal strength of the advertisement messages. According to the number of cluster members, each cluster head schedules its cluster based on TDMA to optimally manage the local transmissions. – In the second phase, as displayed in Fig. 5.7, sensor nodes send their data to the cluster heads according to the established schedule. Optionally, sensor nodes may turn OFF their radios until their scheduled TDMA transmission slot.

268

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Sink

Legend: • All source nodes (S) send their data to their cluster heads according to the established schedule. • The use of a TDMA protocol in the data collection phase ensures that there are no collisions within the clusters, thus saving both energy and time. • After cluster heads (CH) have received all the data from the active sources, they send them back to the sink using a single direct transmission (dotted lines).

Fig. 5.7 LEACH clustering protocol (Fasolo et al. 2007)

LEACH requires cluster heads to send their aggregated data to the basestation over a single link. However, this is a disadvantage of LEACH because single link transmission may be quite expensive when the basestation is far away from the cluster head. LEACH is entirely distributed, as it does not require global knowledge regarding network structure. It is also an adaptive protocol in terms of cluster head selection. On the other hand, there may be high control message overhead if the network topology is dynamic due to nodes mobility. LEACH was the inspiration to dozens of protocols that either improved its performance or set it as a comparison base (Younis and Fahmy 2004; Qing et al. 2006; Kumar et al. 2009). • Cougar is a clustering scheme that performs periodic per-hop data aggregation; it is suitable for applications where sensor nodes continuously generate correlated data (Yao and Gehrke 2002). Once cluster heads aggregate their cluster data, they send the local aggregated data to a gateway node. Similar to LEACH, Cougar is negatively affected by dynamic network topologies. However, Cougar has a unique cluster head election procedure; it selects the cluster heads based on more than one metric and allows sensor nodes to be more than one hop away from their cluster heads. This calls for routing algorithms to exchange packets within clusters. Cougar employs the ad hoc on-demand distance vector (AODV) protocol for intercluster relaying. In Cougar, synchronization is used to correctly aggregate data. The cluster head is synchronized with all sensor nodes in the cluster and does not report its aggregated data to the gateway node until all sensor nodes send their data. Therefore, the synchronization mechanism helps improving the correctness of the aggregated data. An enhancement to Cougar is suggested in Considine et al. (2004). • The hybrid energy-efficient distributed clustering (HEED) protocol benefits from the availability of multiple power levels at sensor nodes for cluster head

5.1 Data-Driven Approach Taxonomy

269

selection (Younis and Fahmy 2004). A combined metric composed of the node residual energy and the node proximity to its neighbors is introduced. HEED defines the average of the minimum power level required by all sensor nodes within the cluster to reach the cluster head; this is the average minimum reachability power (AMRP) used to estimate the communication cost in each cluster. In order to select cluster heads, each sensor node computes its probability of becoming the cluster head as: PðCHÞ ¼ C 

Eresidual Emax

ð5:2Þ

where, C Eresidual Emax

is the initial percentage of cluster heads, is the current residual energy of the sensor node, and denotes the initial energy of the sensor node.

Each sensor node broadcasts a cluster head message; sensor nodes select their cluster head to be the node with the lowest AMRP in the set of received cluster head messages. This process recursively continues until every node is assigned to a cluster head. As in LEACH, cluster heads in HEED communicate directly with the basestation. HEED terminates in O(1) iterations and incurs low message overhead, while achieving fairly uniform cluster head distribution across the network. With appropriate bounds on node density and intracluster and intercluster transmission ranges, HEED almost asymptotically guarantees connectivity of clustered networks. Simulation results demonstrated that HEED is effective in prolonging the network lifetime and supporting scalable data aggregation. • A position-based aggregator node election (PANEL) protocol for WSNs outperforms other aggregator node election protocols through its support to asynchronous sensor network applications where the sensor readings are fetched by the basestation after some delay (Buttyán and Schaffer 2010). The motivation for the design of PANEL was to support reliable and persistent data storage applications, such as TinyPEDS (Girao et al. 2007). PANEL ensures load balancing and supports intra- and intercluster routing, allowing sensor-to-aggregator, aggregator-to-aggregator, basestation-to-aggregator, and aggregator-to-basestation communications. Compared with HEED, PANEL creates more cohesive clusters than HEED; on the other hand, PANEL is more energy-efficient than HEED.

Hybrid Tree/Cluster-Based Data Aggregation Protocols In this approach, tree-based techniques are comprised of each cluster, to have the application requirements fulfilled. The noticeable such hybrid techniques are as follows:

270

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

• Clustered diffusion with dynamic data aggregation (CLUDDA) is a hybrid approach that combines clustering with diffusion mechanisms (Chatterjea and Havinga 2003). CLUDDA includes query definitions inside interest messages initiated by the basestation. Each interest message contains the definition of the query that describes the operations that need to be performed on the data components in order to generate a proper response. Interest transformation reduces the processing overhead by utilizing the existing knowledge of queries. CLUDDA combines directed diffusion (Intanagonwiwat et al. 2003) and clustering during the initial phase of interest propagation. Using clustering mechanism, it is ensured that only cluster heads that perform intercluster communication are involved in the transmission of interest messages. As the regular sensor nodes do not transmit any data unless they are capable of servicing a request, CLUDDA conserves energy. In CLUDDA, any cluster head that has the knowledge of query definition can perform data aggregation, and hence, the aggregation points are dynamic. Also, each cluster head maintains a query cache to present the different data components that were aggregated to obtain the final data. Cluster heads also keep a list of the addresses of neighboring nodes where the data messages originated. These addresses are used to propagate interest messages directly to specific nodes instead of broadcasting. • A simple, location-based clustering scheme (LCS) is suggested in Pattem et al. (2008). Given a sensor field and a cluster size, nodes close to each other form clusters. The clusters so formed remain static for the lifetime of the network. Within each cluster, the data from each of the nodes is routed along a shortest-path tree (SPT) to a cluster head node. Data aggregation takes place at each of the intermediate nodes along the SPT. The cluster head then sends the aggregated data from its cluster to the sink (basestation) along a multihop path with no intermediate aggregation. It was argued that, for a given network size, there exists a simple, static clustering scheme that is near-optimal, in terms of energy efficiency, across a wide range of spatial correlations.

Multipath-Based Data Aggregation Protocols Aggregating along a tree is very susceptible to node and transmission failures that are common in WSNs. Since each of these failures loses an entire subtree of readings, a large fraction of the readings are typically unaccounted for in an SPT-based system; this introduces significant error in the query answer. Efforts to reduce losses by retransmitting packets waste significant energy and delay query responses. In order to overcome the robustness problems of aggregation trees, the multi-path approach has been suggested in Manjhi et al. (2005), Chen and Zhang (2006), and Nath et al. (2008). Instead of having an aggregation tree where each node has to send the partial result of its aggregation to a single parent, these solutions send data over multiple paths.

5.1 Data-Driven Approach Taxonomy

271

The main idea in the multi-path approach is that each node can send the data to its, possibly, multiple neighbors by exploiting the broadcast characteristics of the wireless medium. Hence, data may flow from the sources to the sinks along multiple paths and aggregation may be performed by each node. In contrast to the tree-based schemes discussed in section “Tree-Based Data Aggregation Protocols”, multi-path approaches allow to propagate duplicates of the same information. Clearly, such schemes trade a higher robustness, as multiple copies of the same data can be sent along multiple paths, for some extra overhead due to sending duplicates. An aggregation structure that fits well with this methodology is called rings’ topology, where sensor nodes are divided into several levels according to the number of hops separating them from the data sink. As clarified in Fig. 5.8, data aggregation is performed over multiple paths as packets move level by level toward the sink. The synopsis diffusion protocol is introduced as one of the main forerunners of the multi-path approach: • In the synopsis diffusion protocol, data aggregation is performed through a multi-path approach (Nath et al. 2008). The underlying topology for data dissemination is organized into concentric rings around the sink. Synopsis diffusion consists of two phases; explicitly, the distribution of the queries, and the data retrieval. The ring topology is formed when a node sends a query over the network. Two different structures are considered: – Simple ring structure. During the query distribution phase, the network nodes form a set of rings around the querying node q, which is the only sensor belonging to ring R0. A node is in ring Ri if it is i hops away from the querying node. Fig. 5.8 Aggregation paths over a ring structure (Fasolo et al. 2007)

R1 R2 R3 R4

272

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

– Adaptive rings. It is more robust than the simple ring structure and is able to cope with changes in the network. The distribution phase is not changing, but this time a node u in ring i keeps track of the number of times, nov, the transmissions from any node ni−1 in ring i − 1 included its own data during the last few epochs. That is, node u checks whether its data are aggregated with the information sent by any node in ring i − 1. If nov is small, u tries to find a better ring in order to have more of its own data included in the subsequent transmissions. The main feature of synopsis diffusion is that as dataflow over multiple paths, a node may receive duplicates of the same information. This may affect the aggregation result, especially when aggregation functions are duplicate sensitive. This issue is resolved through proposing proper aggregation functions and data structures. On the upside, multi-path schemes are suitable for networks with frequent packet losses due to mobility or channel impairments, as the extra overhead (duplicates) pays off in terms of robustness; so, if a link fails, the data may still reach the sink through a different path.

Hybrid Tree/Multipath-Based Data Aggregation Protocols In order to benefit from the advantages of tree-based and multi-path schemes, a hybrid approach is proposed to adaptively tune their data aggregation structure for optimal performance. Tributaries1 (National Geographic Society 2018b) and deltas2 (National Geographic Society 2018a) is a typical such approach: • The tributaries’ (National Geographic Society 2018b) and deltas’ protocols attempt to overcome the glitches of both the tree- and multi-path-based structures, by combining the best features of both schemes (Manjhi et al. 2005). In the resulting hybrid algorithm, both data aggregation structures may simultaneously run in different regions of the network. The protocol considers low packet loss rates and high packet loss rates: – Under low packet loss rates, a data aggregation tree is the most suitable structure due to the possibility of implementing efficient sleeping modes and for efficient representation and compression of the data.

1

A tributary is a freshwater stream that feeds into a larger stream or river. The larger, or parent, river is called the mainstream. The point where a tributary meets the mainstream is called the confluence. Most large rivers are formed from many tributaries. Tributaries, also called affluents, do not flow directly into the ocean. 2 Deltas are wetlands that form as rivers empty their water and sediment into another body of water, such as an ocean, lake, or another river. Deltas can also empty into land, although this is less common.

5.1 Data-Driven Approach Taxonomy

273

Sink Sink

Delta region

Expanded delta region

Legend: Node T5 is switched to an M vertex (diagram on the right) and, as a consequence, can transmit the aggregated data flow also to nodes M4 and M5. In particular, M5 can contribute to the data aggregation by passing the data coming from node M to lower levels.

Fig. 5.9 Tributaries’ and deltas’ protocol (Fasolo et al. 2007)

– For high loss rates or when transmitting partial results which are accumulated from many sensor readings, a multi-path approach may be the best choice due to its increased robustness. Hence, nodes are divided into two categories: nodes using a tree-based approach to forward packets (T nodes) and nodes using a multi-path scheme (M nodes). The network is thus organized in regions implementing one of the two schemes. The main difficulty of linking regions running different data aggregation structures may be overcome through the following rules, as portrayed in Fig. 5.9: – Edge correctness. An edge originating from an M node can never be incident on a T node; hence, the aggregation outcome in a multi-path region can only be received by an M node. – Path correctness. M nodes form a subgraph including the sink node, which is fed by trees composed of T nodes. According to the above rules, the sink is surrounded by M nodes only. These nodes form the so-called delta region which can be expanded or shrunk by switching nodes from the tree mode (T) to the multi-path mode (M) and vice versa. In practice, only the nodes lying along the boundary between the two regions are allowed to change their operating mode. Expanding the delta region corresponds to increasing the number of paths toward the sink, which is suitable when the packet loss probability is high. On the other hand, shrinking the region is beneficial in case the network is static and the packet loss probability is small. The user can set a threshold to specify the minimum percentage of nodes that should contribute to the aggregation operation. Noticeably, this percentage increases in case of a wider delta region. This implies that more multi-path nodes are available, consequently leading to a higher robustness against failures and, in turn, to more nodes actively contributing to the aggregation outcome (Fig. 5.9). The opposite holds when the delta region is shrunk.

274

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Appraisal of In-Network Processing Protocols WSNs as designated for environmental monitoring are mainly event-based systems. A sensor network consists of one or more sinks which request specific data streams by expressing interests or queries. The sensors in the network act as sources that detect environmental events and push relevant data to the appropriate requesting sinks. For example, there may be a sink that is interested in a particular spatiotemporal phenomenon, such as “does the temperature ever exceed 30 °C in area A between 11 am and 12 pm?” During the given time interval, all sensors in the corresponding spatial portion of the network act as event-based reporters. They report information toward the requesting sink if and when they detect the indicated phenomenon. Due to the nature of unattended operation in remote or potentially hostile locations, WSNs are extremely short in energy. However since various sensor nodes often detect common phenomena, there is likely some redundancy in the data the various sources communicate to a particular sink. In-network filtering and processing techniques can help conserve the scarce energy resources. Data aggregation has been put forward as an essential paradigm for routing in WSNs. The scheme is to combine the data coming from different sources en-route, eliminating redundancy, minimizing the number of transmissions, and thus saving energy. This paradigm shifts the focus from the traditional address-centric approaches for networking (finding short routes between pairs of addressable end nodes) to a more data-centric approach (finding routes from multiple sources to a single destination that allows in-network consolidation of redundant data). Several techniques have been put into action to aggregate data, namely tree-based, cluster-based, hybrid tree/cluster-based, multi-path-based, and hybrid tree/multipath-based. There is no way to judge which technique is better, since they all are application-dependent; a technique may suit a deployed WSN for a typical application but cannot be of efficiency for another application.

5.1.1.2

Data Compression Protocols

Data compression algorithms fall into two broad classes: lossless and lossy algorithms. Lossless and lossy compression are terms that describe whether or not, in the compression of a file, all original data can be recovered when the file is uncompressed (Said and Pearlman 1996). Specifically speaking: • With lossless compression, every single bit of data that was originally in the file remains after the file is uncompressed; all of the information is completely restored. This is generally the technique of choice for text or spreadsheet files, where losing words or financial data could pose a problem. The graphics interchange file (GIF), an image format used on the Web, provides lossless compression. • Lossy compression reduces a file by permanently eliminating certain information, especially redundant information. When the file is uncompressed, only a

5.1 Data-Driven Approach Taxonomy

275

part of the original information is still there. In lossy compression generally used for video and sound, a certain amount of information loss will not be detected by most users. The joint photographic experts group (JPEG) image file, commonly used for photographs and other complex still images on the Web, is an image that has lossy compression. Using JPEG compression, the creator can decide how much loss to introduce and make a tradeoff between file size and image quality. In WSNs, data are collected by sensors, which, due to noise, produce different readings even when they are sampling an unchanging phenomenon. For this reason, sensor manufactures specify not only the sensor operating range but also the sensor accuracy. Datasheets express accuracy by providing a margin of error, but typically do not include a probability distribution for this error. Thus, when a value is measured by a sensor, it is certain that the actual value is within the error margin, but how far it is from the real value? In this context, lossy compression algorithms may convolute the original error distribution when that distribution is not uniform (Schoellhammer et al. 2004). On the other hand, the criticality of some application domains demands sensors with high accuracy, as measures being corrupted by the compression process are intolerable. In body area networks (BANs), for instance, sensor nodes permanently monitor and log vital signs. Each small variation of these signs has to be captured because it might provide crucial information to make a diagnosis. Thus, lossless compression algorithms might be suitable for WSNs. Since sensor nodes are typically equipped with a few KBytes of memory and a 4–8 MHz microprocessor, embedding classical data compression schemes in these tiny nodes is practically unachievable (Barr and Asanović 2006; Sadler and Martonosi 2006). Lossless data compression can achieve very high compression ratios despite non-negligible memory occupation and computational effort requirements. Due to the limited resources available in tiny sensor nodes, to apply data compression in WSNs requires specifically designed algorithms. Two approaches have been followed, distributing the computational cost on the overall network or exploiting the statistical features of the data under monitoring: • Distributing the computational cost on the overall network (Ganesan et al. 2003; Chen et al. 2004; Guestrin et al. 2004; Girod et al. 2005). This approach is natural in cooperative and dense WSNs, where nodes can collaborate with each other so as to carry out tasks they could not execute alone. By the particular topology of these WSNs, data measured by neighboring nodes are correlated both in space and time. Thus, distributed transforms or estimating distributed models can be applied, which allow decorrelating the data measured by sensors, thus representing these data through fewer bits. For instance, by adopting model estimation, data can be represented by only the parameters of the model. Obviously, the model is generally only an approximation of the data. Therefore, distributed compression algorithms are intrinsically lossy. Due to the popularity of dense sensor networks, several works have been proposed both for distributed quantization and transform (Ganesan et al. 2003) and distributed model estimation (Chen et al. 2004; Guestrin et al. 2004).

276

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

• Exploiting the statistical features of the data under monitoring so as to adapt some existing algorithms to the constraints imposed by the limited resources available on the sensor nodes (Schoellhammer et al. 2004; Sadler and Martonosi 2006; Lynch 2007). Actually, this approach would result in very effective management of energy consumption in WSNs consisting of a number of static sensor nodes, placed in several different prefixed points in an area under monitoring. Also, one or more data collectors, called data mules, come into contact with the static sensors at approximately regular intervals to collect measured values (Shah et al. 2003). This model is characterized by a number of advantages in comparison with the traditional approach based on multihop communication; explicitly, the network lifetime is longer as the number of exchanged packets decreases, the packet loss probability decreases since the number of hops decreases, and the network capacity increases and node synchronization error decreases as the number of hops is smaller than in the multihop approach. On the other hand, the data latency and the costs of the network infrastructure might increase (Somasundara et al. 2006). Since the model with data mules is typically applied in environmental monitoring applications, latency in general is not an issue. Further, the additional cost for data mules can be maintained very low if mobility of external agents available in the environment is exploited. Obviously, the second approach can be a valuable help in power saving only if the execution of compression algorithms does not require an amount of energy larger than the one saved in reducing transmission. After analyzing several classic compression algorithms, it was concluded that compression prior to transmission in wireless battery-powered devices might actually cause an overall increase of power consumption, if no energy awareness is introduced (Barr and Asanović 2006). On the other hand, standard compression algorithms are aimed at saving storage and not energy. Thus, appropriate strategies have to be adopted (Razzaque et al. 2013).

An Efficient Lossless Compression Algorithm for Tiny Nodes of Monitoring WSNs (LEC) Lossless entropy compression (LEC) algorithm exploits the natural correlation that exists in data collected by WSNs and the principles of entropy compression (Marcelloni and Vecchio 2009). LEC low complexity and the small amount of memory required for its execution make it particularly suitable to be used on commercial tiny sensor nodes. Further, it is able to compute a compressed version of each value on the fly, thus reducing storage occupation. LEC exploits a very short fixed dictionary, whose size depends on the precision of the analog-to-digital converter (ADC). Thus, with the dictionary size being fixed a priori, LEC does not suffer from the growing dictionary problem, which might affect other approaches specifically proposed for WSNs.

5.1 Data-Driven Approach Taxonomy

277

LEC follows a scheme similar to the one used in the baseline JPEGy(k) algorithm for compressing the DC coefficients of a digital image (Pennebaker and Mitchell 1992). Such coefficients are characterized by a high correlation, similar to that characterizing data collected by WSNs. LEC exploits a modified version of the exponential-Golomb code (Exp-Golomb) of order 0 (Teuhola 1978), which is a type of universal code. The basic idea is to divide the alphabet of numbers into groups whose sizes increase exponentially. Like Golomb coding (Golomb 1996) and Elias coding (Elias 1975), a codeword is a hybrid of unary and binary codes. In particular, the unary code, a variable-length code, specifies the group, while the binary code, a fixed-length code, represents the index within the group. In LEC, the alphabet of numbers is divided into groups whose sizes increase exponentially, but groups are entropy coded3 (Merriam-Webster 2018c), rather than unary coded. Such modification introduces the possibility of specifying prefix-free codes for the groups. This is an advantage since these codes can be recalculated to best fit the particular probability distribution of the inputs being compressed. Moreover, since the original version of Exp-Golomb manages only nonnegative integers, a minor modification to the original scheme was the introduction of a bijection to map the actual values onto a nonnegative domain. In the sensing unit of a sensor node, each measure mi acquired by a sensor is converted by an ADC to a binary representation ri on R bits, where R is the resolution of the ADC, that is, the number 2R of discrete values the ADC can produce over the range of analog values. Figure 5.10 shows the block scheme of LEC approach. For each new acquisition mi, LEC computes the difference di = ri – ri-1, which is input to an entropy encoder. In order to compute d0, it is assumed that r−1 is equal to the central value among the 2R possible discrete values. The entropy encoder performs compression losslessly by encoding differences di more compactly based on their statistical characteristics. Each nonzero di value is represented as a bit sequence bsi composed of two parts si|ai, where si codifies the number ni of bits needed to represent di (i.e., the group to which di belongs) and ai is the

3

– In Thermodynamics: A measure of the unavailable energy in a closed thermodynamic system that is also usually considered to be a measure of the system disorder. It is a property of the system state that varies directly with any reversible change in heat in the system and inversely with the temperature of the system. Broadly, it is the degree of disorder or uncertainty in a system. – The degradation of the matter and energy in the universe to an ultimate state of inert uniformity is a process of degradation or running down or a trend to disorder. – Chaos, disorganization, randomness. – In Statistical Mechanics: A factor or quantity that is a function of the physical state of a mechanical system and is equal to the logarithm of the probability for the occurrence of the particular molecular arrangement in that state. – In Communication Theory: A measure of the efficiency of a system (such as a code or a language) in transmitting information, being equal to the logarithm of the number of different messages that can be sent by selection from the same set of symbols and thus indicating the degree of initial uncertainty that can be resolved by any one message.

278

5 Energy Management Techniques for WSNs (2): Data-Driven Approach Compressor ri

+

di

bsi

Encoder

ri-1

Delay

Prefix-free table Transmission/Storage Uncompressor ri

+

di

Decoder

bsi

+ ri-1

Delay

Fig. 5.10 Block diagram of the compressor/uncompressor schemes (Marcelloni and Vecchio 2009)

representation of di (i.e., the index position in the group). When di is equal to 0, the corresponding group size is equal to 1; therefore, there is no need to codify the index position in the group; and it follows that ai is not represented. For any nonzero di, ni is trivially computed as dlog2 ðjdi jÞe, at most ni is equal to R. Thus, in order to encode ni a prefix-free table of R + 1 entries has to be specified. This table depends on the distribution of the differences di; more frequent differences have to be associated with shorter codes. In typical data collected by WSNs, the most frequent differences are those close to 0. This occurs for smooth signals, but also for non-smooth signals. For non-smooth signals, to model the drastic changes in the data, sampling intervals are typically very short. Thus, in order to avoid the cost of computing frequencies on sensor nodes, the amount of work already carried out on JPEG algorithm is exploited. In Table 5.1, the first 11 lines coincide with the table used in the baseline JPEG algorithm for compressing the DC coefficients (Pennebaker and Mitchell 1992). These coefficients have statistical characteristics similar to the measures acquired by the sensing unit. If the resolution of the ADC is larger than 14 bits, the table has to be appropriately extended; this extension will produce a higher memory occupation, but will not perceptibly affect the compression ratios. Differences corresponding to longer codes have a probability very close to 0, thus letting practically unchanged the compression ratios obtainable using LEC.

5.1 Data-Driven Approach Taxonomy Table 5.1 Dictionary used (Marcelloni and Vecchio 2009)

279

ni

si

di

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

00 010 011 100 101 110 1110 11110 111110 1111110 11111110 111111110 1111111110 11111111110 111111111110

0 −1,+1 −3,−2, +2, +3 −7,…,−4, +4,…, +7 −15,…,−8, +8,…, +15 −31,…,−16, +16,…, +31 −63,…,−32, +32,…, +63 −127,…,−64, +64,…,+127 −255,…,−128, +128,…, +255 −511,…,−256, +256,…, +511 −1023,…,−512, +512,…, +1023 −2047,…,−1024, +1024,…, +2047 −4095,…,−2048, +2048,…, +4095 −8191,…,−-4096, +4096,…, +8191 −16,383,…,−8192, +8192,…, +16,383

In order to manage negative di, LEC maps the input differences onto nonnegative indexes, using the following bijection:
index ¼

di ; 2ni  1  jdi j;

di ;  0; di \0

ð5:3Þ

Finally, si is equal to the value at entry ni in the prefix-free table and ai is the binary representation of index over ni bits. Since di is typically represented in two’s complement notation, when di < 0, ai is equal to the ni low-order bits of di − 1. The procedure used to generate ai guarantees that all possible values have different codes. For instance, using Table 5.1, di = 0, di = +1, di = −1, di = +255, and di = −255 are encoded as 00, 010|1, 010|0, 111110|11111111, and 111110| 00000000, respectively. Once bsi is generated, it is appended to the bitstream that forms the compressed version of the sequence of measures mi. The compression algorithm is very simple, since it can be implemented in a few lines of code and requires only maintaining in memory the column si of Table 5.1. Thus, it reduces the amount of data generated by the sensing unit’s onboard sensor nodes and consequently decreases the number of packets to be transmitted, thus saving energy. In order to show the effectiveness and validity of LEC, it was tested against various real-world datasets. First, using smooth signals like temperature and relative humidity, which are particularly suitable for WSNs. Then, the performance of LEC was assessed against real non-smooth signals, like solar radiation, and seismic, and ECG signals. The obtained findings were as below detailed:

280

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Table 5.2 Number of samples of the four smooth datasets (Marcelloni and Vecchio 2009)

Deployment name

Number of samples

HES-SO FishNet LUCE Grand-St-Bernard Le Génépi

12,652 64,913 23,813 21,523

• Smooth signals. Temperature and relative humidity measurements from four SensorScope deployments are used (LCAV 2017): specifically, HES-SO FishNet, LUCE, Grand-St-Bernard, and Le Génépi Deployments. The WSNs adopted in the deployments employ a TinyNode node, which uses a TI MSP430 microcontroller (Texas Instruments 2004), an XE1205 radio (Semtech 2018), and a Sensirion SHT75 sensor module (Sensirion 2018a). This module is a single chip that includes a capacitive polymer sensing element for relative humidity and a bandgap temperature sensor4 (AZoSensors 2018). Both sensors are seamlessly coupled to a 14-bit ADC and a serial interface circuit on the same chip. The Sensirion SHT75 can sense air temperature in the −20 °C, +60 °C range and relative humidity in the 0%, 100% range. The outputs raw_t and raw_h of the ADC for temperature and relative humidity are represented with resolutions of 14 and 12 bits, respectively. The outputs raw_t and raw_h are converted into measures t and h expressed, respectively, in Celsius degrees (°C) and percentage (%) as described in Sensirion (2018b). The datasets corresponding to the four deployments store measures t and h. On the other hand, the LEC algorithm works on raw_t and raw_h. Before applying LEC, raw_t and raw_h are extracted from t and h, respectively, by using the inverted versions of the conversion functions (Sensirion 2018b). Datasets are built by extracting from the four SensorScope deployment databases the temperature and relative humidity measurements for a randomly extracted sensor node within a specific time interval. Table 5.2 lists the number of samples of the datasets. Table 5.3 displays some statistical characteristics of the datasets. The computations involve: – The mean s and the standard deviation rs of the samples. – The mean d and the standard deviation rd of the differences between consecutive samples. – The information entropy of the original signal as given by:

4

A silicon bandgap temperature sensor is a type of thermometer or temperature detector commonly employed in electronic devices. It has good stability at extreme environmental conditions due to the integral stability of crystalline silicon.

HES-SO FishNet LUCE Grand-St-Bernard Le Génépi

Dataset

14.92 7.21 4.34 4.09

± ± ± ±

3.88 3.16 3.95 4.05

Temperature
s  r
s

3.02E–04 −2.87E–05 −4.28E–04 1.38E–04

d
 rd
± ± ± ± 0.26 0.05 0.21 0.33

10.26 10.07 10.29 10.25

H 5.10 4.05 6.15 6.82

Hd 83.94 87.04 68.17 69.71

± ± ± ± 9.79 8.04 18.18 19.43

Relative humidity
s  r
s

Table 5.3 Statistical characteristics of the four smooth datasets (Marcelloni and Vecchio 2009)

−1.60E–03 1.12E–04 −3.28E–04 −1.70E–04

d
 rd
± ± ± ±

1.26 0.55 1.68 2.77

9.75 10.08 10.78 10.84

H

Hd 5.84 5.85 7.14 7.67

5.1 Data-Driven Approach Taxonomy 281

282

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

H¼

N X

pðxi Þ  log2 pðxi Þ

ð5:4Þ

i¼1

where N is the number of possible values xi (the output of the ADC) and pðxi Þ is the probability mass function of xi . – The information entropy of the differentiated signal is: Hd ¼ 

N X

pðdi Þ  log2 pðdi Þ

ð5:5Þ

i¼1

Clearly, the first two datasets are characterized by lower entropy values than the other two datasets. The performance of a compression algorithm is computed using the compression ratio (CR) defined as:   comp size CR ¼ 100  1  orig size

ð5:6Þ

where comp size and orig size are, respectively, the sizes in bits of the compressed and the uncompressed bitstreams. Considering that uncompressed samples are normally byte-aligned, both temperature and relative humidity samples are represented by 16-bit unsigned integers. Thus, from Table 5.2, orig size may be computed for the given datasets, as listed in Table 5.4. It is noticed that: – As expected, the LEC algorithm achieves higher compression ratios on datasets characterized by a low entropy and, in general, a low variability between consecutive samples, that is, low values of the mean and standard deviation of the differences between consecutive samples. – By comparing the results in Table 5.4 with the statistical descriptions of the datasets shown in Table 5.3, it is observed that the LUCE temperature dataset, being characterized by the lowest entropies, the lowest mean and standard deviation of the differences between consecutive samples, achieves the highest compression ratio. Table 5.4 Compression ratios obtained by LEC on the four smooth datasets (Marcelloni and Vecchio 2009) Dataset

orig size

Temperature comp size

CR%

Relative humidity comp size CR%

HES-SO FishNet LUCE Grand-St-Bernard Le Génépi

202,432 1,038,608 381,008 344,368

70,069 303,194 156,099 158,994

65.39 70.81 59.03 53.83

76,635 396,442 180,192 178,725

62.14 61.83 52.71 48.10

5.1 Data-Driven Approach Taxonomy

283

– On the other hand, samples in LUCE datasets are obtained by measuring temperature and relative humidity at intervals of 30s unlike intervals of 2 min used in the other datasets. All samples are assumed transmitted to the sink using the lowest number of messages so as to save power (Mainwaring et al. 2002). Since each packet can contain at most 29 Bytes of payload (Croce et al. 2008), it is possible to count the number of packets necessary to deliver the uncompressed (orig pkt) and the compressed (comp pkt) bitstreams. The packet compression ratio, PCR, is defined as:   comp pkt PCR ¼ 100  1  orig pkt

ð5:7Þ

Table 5.5 summarizes the results obtained. To assess the performance achieved by LEC, it is compared with S-LZW (Sadler and Martonosi 2006) and with LTC (Schoellhammer et al. 2004). Comparing Table 5.6 that summarizes the results obtained by the S-LZW algorithm and Table 5.4, it is obvious that LEC outperforms considerably in terms of CRs. To evaluate the complexity of LEC, a comparative analysis is conducted on the number of instructions required to compress data. The Sim-It Arm simulator (Sim-It-ARM 2007) is adopted, and it is an instruction-set simulator that runs both system-level and user-level ARM programs. From Table 5.7, it is realized that though LEC achieves higher compression ratio, it requires less instructions. Further comparing LEC and LTC (Schoellhammer et al. 2004) disclosed LEC better performance in terms of CR and lower number of instructions. Table 5.5 Number of packets to deliver the uncompressed and compressed versions of the four datasets (Marcelloni and Vecchio 2009) Dataset

orig pkt

Temperature comp pkt

PCR%

Relative humidity comp pkt PCR%

HES-SO FishNet LUCE Grand-St-Bernard Le Génépi

873 4477 1643 1485

302 1307 673 686

65.41 70.81 59.04 53.80

331 1709 777 771

Table 5.6 Compression ratios obtained by S-LZW on the four smooth datasets (Marcelloni and Vecchio 2009)

Dataset HES-SO FishNet LUCE Grand-St-Bernard Le Génépi

62.08 61.83 52.71 48.08

Temperature

Relative humidity

CR% 30.35 48.99 26.86 22.02

CR% 36.27 31.24 24.92 21.93

284

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Table 5.7 Complexity of LEC against S-LZW (Marcelloni and Vecchio 2009) LEC Temperature Average number of instructions Average number of saved bits Average number of instructions per saved bit

S-LZW Temperature

Relative humidity

Relative humidity

30,549.25

26,610.25

63,207.00

63,207.00

2762.75 11.35

2844.00 10.58

1598.00 68.65

1880.00 184.55

• Non-smooth signals. LEC was also assessed through compressing non-smooth signals, such as solar radiation, seismic, and ECG datasets. Comparing Table 5.3 with Table 5.8, it is noticed that temperature and relative humidity datasets are characterized by higher entropies than solar radiation, seismic, and ECG datasets. Since LEC is an entropy compression algorithm, independently of the smoothness of the signals, its performance increases when the entropy decreases. This difference between the entropies of smooth and non-smooth datasets is justifiable. Temperature and relative humidity signals are characterized by possibly large differences between consecutive samples; however, these differences are quite repetitive. It follows that a larger number of symbols is needed for encoding temperature and relative humidity signals than for solar radiation, seismic, and ECG signals. As for the compression ratio, as listed in Table 5.9, LEC has achieved compression ratios remarkably higher than S–LZW. While LEC demonstrated significantly better compression performance in comparison with S–LZW, one critical challenge of LEC is robustness, an indicator of whether or not a data compression algorithm can be widely and effectively applied in various WSN applications for different sensor data with very different temporal characteristics. The differential predictor in LEC is generic without training/ learning when applied to any sensed data streams, making LEC simple to use and Table 5.8 Statistical characteristics of three non-smooth datasets (Marcelloni and Vecchio 2009) Dataset


s  r
s

d
 rd


H

Hd

Solar radiation Seismic ECG

10.36 ± 15.35 −0.36 ± 26.96 −0.33 ± 0.18

1.26E–04 ± 7.83 −7.78E–04 ± 3.65 −4.83E–06 ± 0.05

5.34 6.79 6.23

4.24 3.91 3.78

Table 5.9 Compression ratios for LEC against S-LZW for three non-smooth datasets (Marcelloni and Vecchio 2009)

Dataset Solar radiation Seismic ECG

LEC

S-LZW

CR% 70.31 69.72 71.53

CR% 58.25 43.33 54.59

5.1 Data-Driven Approach Taxonomy

285

scalable for large-size WSNs. As a result of such a generic and non-adaptive predictor, LEC may show good compression performance for some sensed data streams in a WSN and may also produce poor performance for other data streams in other WSNs. In other words, LEC suffers from its lack of robustness. An improvement over LEC is implemented through sequential lossless entropy compression algorithm for WSNs (SLEC) that achieves highly robust compression performance for various sensor data streams (Liang and Li 2014). At the same time, S-LEC is sufficiently simple to enable its energy-efficient implementation and execution on resource-constrained WSN nodes. Moreover in WSN-based compression algorithms, K-RLE inspired from the basic run-length encoding (RLE) algorithm (Salomon 2004) is proposed to improve on compression outcomes (Capo-Chichi et al. 2009).

5.1.1.3

Data Prediction Protocols

Data prediction techniques build a model describing the sensed phenomenon, so that queries can be answered using the model instead of the actually sensed data. There are two instances of a model in the network, one residing at the sink and the other at source nodes; hence, there are as many pairs of models as sources. The model at the sink can be used to answer queries without requiring any communication, thus reducing the energy consumption. Obviously, this operation can be performed only if the model is a valid representation of the phenomenon at a given instant. Here comes the role of the model residing at the source nodes that is used to ensure the model effectiveness. Functionally, sensor nodes just sample data as usual and compare the actual data against the prediction. If the sensed value falls within an application-dependent tolerance, then the model is considered valid. Otherwise, the source node may transmit the sampled data and/or start a model update procedure involving the sink as well. The features of a specific data prediction technique depend on the way the model is built. Data prediction techniques can be split into three main classes (Fig. 5.11): • Deriving a stochastic characterization of the phenomenon, i.e., in terms of probabilities and/or statistical properties. Two main approaches are available: – Mapping data into a random process described in terms of a probability density function (PDF). Data prediction is then obtained by combining the computed PDFs with the observed samples. – Deriving a state space representation of the phenomenon, so that forthcoming samples can be guessed by filtering out a non-predictable component modeled as noise. • Time-series forecasting, where a set of historical values (the time-series) obtained by periodical samplings are used to predict a future value in the same series. The main difference with respect to other statistical, or probabilistic approaches, is that time-series analysis explicitly considers the internal structure of data. Generally, a time-series can be represented as a combination of a pattern

286

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Data prediction

Stochastic approaches

Time-series forecasting

Algorithmic approaches

Fig. 5.11 Data prediction approach taxonomy (Anastasi et al. 2009)

and a random error. The pattern, in turn, is characterized by its trend (its long-term variation) and its seasonality (its periodical fluctuation). Once the pattern is fully characterized, the resulting model can be used to predict future values in the time-series. • Heuristic or state transition model describing the sensed phenomenon. Such algorithmic approaches derive methods or procedures to build and update the model on the basis of the chosen characterization. Based on the aforementioned taxonomy, the following sections comprehensively present the most relevant data prediction schemes (Sections “Stochastic Approaches”, “Time-Series Forecasting Approaches”, and “Algorithmic Approaches”).

Stochastic Approaches Approximate Data Collection in Sensor Networks Using Probabilistic Models (Ken) A robust approximate technique is suggested, Ken, based on using replicated dynamic probabilistic models to minimize communication from sensor nodes to the network PC basestation (Chu et al. 2006). In addition to data collection, Ken5 (Merriam-Webster 2018d) is well suited to anomaly and event detection applications. The usual core challenge in WSNs is to minimize energy consumption. Prior to Ken approach, database research has proposed to achieve this goal by pushing data-reducing operators like aggregation and selection down into the network. Nonetheless, this approach has proven unpopular with early designers of WSN technology, who typically want to extract complete “dumps” of the sensor readings, i.e., to run SELECT queries. However, because these queries do not include

5

– The range of perception, understanding, or knowledge. – The range of vision, sight, view.

5.1 Data-Driven Approach Taxonomy

287

data reduction, they consume significant energy in WSN query processors. In Ken, the SELECT problem for WSNs is tackled. WSNs can provide innovative new data sources for a variety of applications; however because of the limited battery resources on each sensor device, it is challenging to extract these data. In practice, WSN deployments are valuable only if they can run unattended for months or even years. Hence, any WSN technology has to be cautious in its energy consumption. Among the various tasks performed by a wireless sensor node, radio transmissions are by far the most expensive in terms of energy consumption. On a typical sensor node, the Telos mote, message transmission, and receipt expend an order of magnitude more energy than CPU computations over an equivalent length of time (Moteiv 2004). Poor energy consumption can be dramatic in practice; for example, a software bug may keep the radios active to the extent of exhausting sensor batteries in few days. One way to significantly reduce communication cost in WSNs is to perform in-network aggregation such as AVG and MIN (Madden et al. 2002; Guestrin et al. 2004; Silberstein and Yang 2007), data reduction via wavelets or distributed regression (Hellerstein and Wang 2004; Nath et al. 2008). However, these techniques do not provide the fine data granularity desired by many WSN users. The proposed Ken approach offers meaningful features (Fig. 5.12): • It is based on a form of compression using replicated dynamic probabilistic models. The basic idea is to maintain a pair of dynamic probabilistic models over the WSN attributes, with one copy distributed in the network and the other at a PC basestation. At every time instance with a frequency f, the basestation simply computes the expected values of the WSN attributes according to the model and uses it as the answer to the “SELECT * FREQ f”query; this scheme

Source (sensor nodes)

Sink (basestation)

Legend: • A pair of dynamic probabilistic models, at the sensor nodes and at the basestation, are maintained over X, the WSN attributes. • Information is communicated from the sensor nodes to the basestation only if the predictions are not within bounds.

Fig. 5.12 Ken approach (Chu et al. 2006)

288

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

requires no communication. The sensor nodes always possess the ground truth, and whenever they sense anomalous data that was not predicted by the model within the required error bound, they proactively route the data back toward the basestation. As data are routed toward the basestation, spatial correlations among the reported data are used to further lower communication. Using these techniques, all user-visible readings are guaranteed to be within a fixed error bound from the measured readings, even though very few readings are communicated to the basestation. • An attractive feature of the Ken architecture is that it naturally accommodates applications that are based on event reporting or anomaly detection; these include fire alert and response (Lin et al. 2014) and vehicle tracking (Sharp et al. 2005). In these scenarios, the sample rate is typically quite high, but the communication rate should remain quite low under most circumstances. The model reflects the expected “normal” state of the environment being monitored; anomalies result in reports being pushed to the basestation for urgent handling by infrastructure logic. In addition to naturally supporting these applications, Ken enhances them with the added functionality of being able to support interactive query results with well-bounded approximate answers. In essence, approximate data collection and event detection become isomorphic6 (Meriam-Webster 2018a). • A key technical challenge was the degree to which spatial correlations can be exploited among distributed sensor nodes. A rich model of the spatial correlations can significantly improve compression of the data communicated to the basestation. On the other hand, sophisticated models may require significant communication among the sensor nodes for coordination and in-network model maintenance. Tradeoffs are exploited considering how the distributed model maintenance is mapped onto the communication topology of a WSN. There is a focus on a natural class of distributed models based on spatial partitions of the network; within this class of models, choosing the optimum partitioning is NP-hard. Two algorithms for partitioning are proposed, dynamic programming and greedy heuristic. Experimentally, it was shown that the heuristic algorithm performs effectively on real-world network traces. As a start, the problem that Ken attempts to solve is formalized and then follows a discussion on how replicated dynamic probabilistic models can be used to solve this problem. Also provided, a detailed description of Ken design, and a formalization of the model selection. The problem definition shows out as:

6

Being of identical or similar form, shape, or structure.

5.1 Data-Driven Approach Taxonomy

289

• A sensor network, called source, continuously monitors a set of distributed attributes X and generates a data value xt at every time instance t that depends on the application-specific frequency of data collection. b t , of the true • A PC basestation (sink) that requires approximation, X  an 2-loss   t b t data values at all times, i.e., 8i; t; xi  X i \ 2. b t by It is required to design a data collection protocol that optimally achieves X utilizing known temporal and spatial correlations in the WSN attributes. Ken normal operation is mostly carried by the source. The source performs the following steps at time t (Fig. 5.12): 1. Using the transition model, compute the probability distribution function   (PDF) p ¼ p X1t þ 1 ; X2t þ 1 ; :::; Xnt þ 1 over attributes Xit þ 1 ; i ¼ 1; . . .; n. This PDF depends on all observations that have been communicated to the sink so far; other values are ignored since they did not change the model at the sink. 2. Compute the expected values of the attributes according to the PDF: Z   ^it þ 1 ¼ Xit þ 1  p X1t þ 1 ; . . .; Xnt þ 1 dX1t þ 1 . . .dXnt þ 1 X ð5:8Þ 3. If the expected values are sufficiently accurate, i.e.,  tþ1  X ^i  xit þ 1 \ 2; 8i ¼ 1; . . .; n, then stop. There is no need to send anything to the sink. 4. Otherwise: (a) Find the smallest subset of attributes such that conditioning on it makes the predictions accurate; i.e., find the smallest Y ¼ fXi1 ; . . .; Xik g such that the expected values according to the PDF satisfy the accuracy guarantees:   pY ¼ p X1t þ 1 ; . . .; Xnt þ 1 Xit1þ 1 ¼ xit1þ 1 . . .Xitkþ 1 ¼ xtikþ 1

ð5:9Þ

The set of attributes Y ¼ fX1 ; . . .; Xn g always satisfies this condition. (b) Send the values of attributes in X to the sink. The sink carries on the following steps: 1. Similar to the first step followed by the source, it uses the transition model to   compute p ¼ p X1t þ 1 ; X2t þ 1 ; :::; Xnt þ 1 : 2. If the sink receives from the source values of attributes in X ¼ fXi1 ; . . .; Xik g, then condition p using these values, as performed by the source in the fourth bullet above. 3. Compute the expected values of the attributes Xit þ 1 , and use them as the approximation to the true values. The model used for prediction in Ken operation is principal. Regardless of the model, later described, the correctness of data collection is never compromised; i.e., the accuracy guarantees at the sink are always maintained. However, the

290

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

effectiveness of Ken in reducing the data transmission rates is highly dependent on the chosen model. This would insinuate choosing a sophisticated and complex model that captures all the correlations and the patterns in the data. However, because of the distributed nature in which data are generated in a WSN, this is not always useful; in fact, complex models can sometimes result in higher overall communication costs than relatively unsophisticated models such as Model 1 described below. As much as nodes may need to communicate with each other to bring correlated data together in one place, communication can be very significant. For a more specific cost evaluation, the total communication cost incurred during the above described normal operation of Ken consists of two components: • Intra-source. The cost incurred in the process of collecting data generated at each time step to check if the predictions are accurate (Step 3 in the source operation). • Source–sink. The cost experienced while sending a set of values to the sink (Step 4b in the source operation). The suffered high communication costs lead to proposing approximate data collection with accuracy guarantees, as presented in the models below and typically adopted by Ken in Model 4 and Model 5. Thus, given a WSN consisting of a set N of nodes, a set of attributes X observed by these nodes, and a communication topology connecting these nodes, find a dynamic probabilistic model M, and a mapping from f : X ! N specifying where attribute predictions should be checked, such that the cost is the sum of: • Intra-source. The total communication cost incurred in sending the value of attribute Xi to f(Xi) at every time step. • Source–sink. The total communication cost incurred while sending a set of values to the sink as required. Now, it is time to exhibit several forms of the replicated dynamic probabilistic models that might be used for prediction. Basically, both source and sink maintain a dynamic probabilistic model of how data evolve; these models are always kept in sync (Fig. 5.12). The sink uses the data value(s) predicted by the model as approximation to the true data. The source, who knows the predicted value by virtue of running a copy of the model, makes sure that the predicted data values satisfy the required bounded-loss approximation guarantees, by communicating some information to the application user as required. In general, no class of models will work suitably for all types of WSNs and applications. The relevant forms for the dynamic probabilistic models that might be applicable are: • Model 1. Constant data values model. This simplistic prediction model assumes that the data value remains constant over time:

5.1 Data-Driven Approach Taxonomy

^tþ1 ¼ X ^ t ; i:e, X ^it þ 1 ¼ X ^it ; 8i X i i

291

ð5:10Þ

The predicted value according to the model is same as the predicted value at the last time instant. The source, by virtue of “running” a copy of the model, knows the value that the sink uses and sends an update to the sink if the accuracy bound is not satisfied. Since this model is naturally distributed, each sensor node can decide, and if required, send such an update independently of the other sensor nodes. • Model 2. Linear prediction model. Because of the strong temporal correlations present in sensor data, this improved model would be: ^it þ 1 ¼ ai  X ^it þ bi X

ð5:11Þ

^it denotes the approximation computed at where ai and bi are constants, and X ^it þ 1 is sufficiently close time t. In that case, the source at time t checks whether X tþ1 it observes and sends an update to the sink if it is not. to xi The model, though an improvement over the first model, only utilizes the temporal correlations and ignores the spatial correlations across sensors that tend to be very strong in many WSN deployments. To unify treatment of both kinds of correlations, let the prediction model used by Ken be a dynamic probabilistic model can be used to compute a probability density function (PDF),  t that t p X1 ; X2 ; :::; Xnt , over the possible assignments of values to the variables Xi, i = 1,…,n at time t. For simplicity, presentation is restricted to Markovian models, where given the values of all attributes at time t, the values of attributes at time t + 1 are independent of those for any time earlier than t. This assumption leads to a very simple model for representing such a dynamic system consisting of:   – A PDF for the initial state of the system, p X1t¼0 ; X2t¼0 ; :::; Xnt¼0 .   tþ1 tþ1  – A transition model, p X1 ; X2 ; :::; Xnt þ 1 X1t ; X2t ; :::; Xnt ; that can be used to compute the PDF for time t + 1 from the PDF at time t. As noted before, the sink uses the expected values of the variables according to the PDF computed at time t as the approximation to the true variable values. However, depending on the approximation errors and the specified accuracy bounds, the source may need to communicate some information to the sink that is incorporated in the model so that the accuracy bounds are met. For the Markovian dynamic models, this information takes the form of observed variables values, Xit ¼ xti , at time t, and the process by which this information is incorporated in the model is by conditioning. If the observations reported to the sink at time t are denoted by Ot , the PDF at time t + 1 can be computed given all the observations communicated to the sink till time t:

292

(a)

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

(b)

(c)

Fig. 5.13 Two-dimensional Gaussian linear model (Chu et al. 2006)

   p X1t þ 1 ; X2t þ 1 ; :::; Xnt þ 1 O1...t Z    ¼ p X1t þ 1 ; X2t þ 1 ; :::; Xnt þ 1 X1t ; X2t ; :::; Xnt     p X1t ; X2t ; :::; Xnt O1...t dX1t dX2t . . .dXnt

ð5:12Þ

• Model 3. Two-dimensional linear Gaussian model. A system with two variables, X1 and X2, is considered where the data evolution is modeled using a two-dimensional linear Gaussian. Figure 5.13a denotes the model at time t,   ^1t and X ^2t , respectively. When transitioning to p X1t ; X2t , with expected values X the next time instance,  t + 1, both the source and the sink use  the transition model p X1t þ 1 ; X2t þ 1 X1t ; X2t to obtain a model p X1t þ 1 ; X2t þ 1 over the two ^ t þ 1; X ^ t þ 1 (Fig. 5.13b). variables at time t + 1 with expected values X 1 2 The source checks whether the expected values that will be used, as approximation to the true values by the sink, are sufficiently accurate. If yes, no data needs to be transmitted to the sink (Fig. 5.13c). On the other hand, if the predictions are not accurate, the source communicates a subset of the values to the sink (in this example, the observed value of X1t þ 1 ), and both the source and sink update the model using this information to obtain   p X1t þ 1 ; X2t þ 1 X1t þ 1 ¼ x . The expected values of the variables will thus fulfill the required accuracy guarantees. • Model 4. Disjoint-Cliques model (DjC). A proper choice of M to reduce intra-source cost and also to utilize spatial correlations between attributes is to partition the sensor attributes in multiple localized clusters and use a multi-dimensional model for each of the clusters. These clusters of attributes are called cliques, and the sensor node at which the inference is done is the clique root. The clique root is not required to be a part of the clique (Fig. 5.14). Disjoint-Cliques models naturally localize and distribute the in-network computation required by Ken and may be appropriate for use in many different WSN

5.1 Data-Driven Approach Taxonomy

293

Runs 3-dimensional model over {X4,X5,X6}

Runs two 3-dimensional models over {X1,X2,X3} and {X4,X5,X6} Basestation

Runs 3-dimensional model over {X1,X2,X3} Legend: Example with cliques {X1,X2,X3} and {X4,X5,X6}.

Fig. 5.14 Disjoint-Cliques model (Chu et al. 2006)

environments. A Disjoint-Cliques model is denoted as M ¼ M1 ðC1 Þ; . . .; Mk ðCk Þ, where k is the number of partitions, and Ci  X wherei 2 f1; k g is a partitioning of X. Specifically, separate k multi-dimensional models will be run, one for each clique Ci . The data reduction factor, mi ; is defined to be the expected number of communicated values for Ci . Noticeably, the expected cost incurred by the group of sensor nodes Ci is independent of the rest of the sensor network. Precisely, the total communication cost of collecting the values of variables in Ci at one node, checking whether the prediction is accurate and communicating a subset of values to the basestation as needed, depends only on the members of Ci and the communication topology of the network. Hence, if Ciroot was chosen to be the sensor node at which the data corresponding to Ci is collected, then:



Intra-source ¼

X

  comm x; Ciroot

x2Ci



  Source-sink ¼ mi  comm Ciroot where Ciroot can be simply computed as: Ciroot ¼ argminroot2X

X x2Ci

commðx; rootÞ þ mi  commðroot; 0Þ

ð5:13Þ

294

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

(c) (a)

(b)

Fig. 5.15 Average model (Chu et al. 2006)

This independence between costs of different cliques allows devising algorithms to find effective Disjoint-Cliques model solutions. The problem of finding the optimal Disjoint-Cliques model is NP-hard; exhaustive and heuristic algorithms for finding good Disjoint-Cliques solutions are discussed later. Model 5. Average model (Avg). It is a class of models which could potentially P ^ ¼ ni¼1 Xi =n can predict work very well if the average of all the variables X perfectly the individual values (Fig. 5.15). In this case, f : X ! N maps every attribute to the node that produces it, whereas M consists of n models, Mi , each
Though this model might seem to incur a considerably over two variables Xi ; X.
and communicating it to all the nodes in the WSN. As high cost in computing X
can be done efficiently using shown in the figure, the computation of X in-network aggregation requiring only O(n) messages in the process.
to the sensor nodes also takes at most O(n) messages. Disseminating X
always predicts Xi within Hypothetically, if the source–sink cost is zero (i.e., X the approximation bounds), this model can reduce the total communication cost by a factor of O(n) over the naive approach of communicating all values generated by the WSN to the basestation. In Ken, the solutions to find Disjoint-Cliques follow two approaches: • Exhaustive algorithm for finding an optimal solution. A dynamic programming-based algorithm that finds the optimal solution for a given instance of the problem is implemented. The algorithm proceeds by bottom-up finding the optimal solution for each subset of attributes, using the principle of optimality to reduce unnecessary computation. The complexity of this algorithm is Oðn  4n þ S  2n Þ, where n is the number of attributes in X, and S is the cost of computing the data reduction factor for a given clique Ci , which makes it prohibitively expensive except in the simplest WSNs.

5.1 Data-Driven Approach Taxonomy

295

• Greedy heuristic algorithm for finding a Disjoint-Cliques model. The algorithm begins by choosing an arbitrary attribute Xx from the set of all attributes, finding the clique containing Xx that has the largest per-attribute data reduction factor, and then greedily choosing this clique to be in the final solution. It then removes the attributes contained in this clique, including Xx , from the set of all attributes, and repeats the procedure till all attributes are included in some clique in the solution. The algorithm also avoids computing data reduction factors for cliques that contain attributes too far from each other in the WSN. Because of the high intra-source communication required to collect such attributes at one location, and also because spatial correlations tend to be inversely proportional to distance, such cliques are unlikely to be part of the optimal solution. The running cost of this algorithm is further reduced by restricting the sizes of cliques considered in the final solution through the parameter k. The complexity of this n greedy algorithm is upper-bounded by Oðð Þ  SÞ where n is the number of k attributes in X, and S is the cost of finding a data reduction factor for a given clique. To put Ken into action making use of Model 4 and Model 5, an extensive experimentation was conducted. Ken was implemented in MATLAB, and its performance was assessed through datasets from two real-world WSN deployments. A deployment at the Intel Research Lab in Berkeley (Madden, Intel Lab Data 2004) consisting of 49 MICA2 motes (Crossbow 2002), and a deployment at the UC Berkeley Botanical Gardens comprising 11 MICA2 motes. Three attributes are sensed by these motes: typically, temperature, humidity, and voltage. Ken showed effectiveness in reducing the communication cost and the energy consumption of the system without unjustified sacrifice in the quality of results or frequency. For meaningful evaluation, several schemes are applied to the obtained datasets and compared: • TinyDB (Madden et al. 2003). This scheme always reports all sensor values to the basestation, and the guarantees it provides are the strongest as the data errors are always zero. • Approximate caching (ApC) (Olston et al. 2001). This approach caches the last reported reading at the sink and source; sources do not report if the cached reading is within the threshold of the current reading. In modeling terms, this is a degenerate Markov model. Approximate caching reporting threshold was set to match that of Ken (e.g., 0.5 °C for temperature). • Ken with Disjoint-Cliques (DjC) and Average (Avg) models. The Disjoint-Cliques models permit a finely tuned study of the impact of spatial correlations via restricting the maximum possible clique size. Greedy-k heuristic algorithm is used to find the Disjoint-Cliques model used in Ken; DjCk denotes the model found by Greedy-k. • Single-node dual models (Jain et al. 2004). This technique is equivalent to Ken with DjC1 (maximum clique size restricted to one).

296

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Disregarding network topology-dependent characteristics, there is an investigation of the effectiveness of these schemes on reducing the amount of data reported to the basestation. Figures 5.16 and 5.17 plot the experimentation outcomes for the Intel Research Lab and Botanical Gardens data, respectively. From the figures, several facets are distinguished:

Temperature (%)

Humidity (%)

• Ken and approximate caching approach achieve significant savings over TinyDB. This is because error tolerance is relaxed in exchange for communication savings. While the magnitude of these savings will clearly vary with the desired error bounds, the figures indicate that even for a modest bound of 0.5 ° C, 65% data reduction is readily attainable; TinyDB pays a hefty premium for exactness. • Ken with average model reports at a higher rate than Ken with Disjoint-Cliques model with max clique size restricted to 2 (DjC2). Given that Ken with Average model also incurs a high fixed aggregation and dissemination cost, it will not outperform DjCk, k  2, though it could conceivably outperform ApC and DjC1.

Minutes since initial deployment

(a) Measured dataset

Data reported (%)

ApC: Approximate Caching Avg: Ken with Average model DjCN: Ken with DisjointCliques model of clique size N

TinyDB ApC

Avg

DjC1

DjC2

DjC3

DjC4

DjC5

(b) Percentage data reported under various schemes Fig. 5.16 Experimentation outcomes from Intel Research Lab (Chu et al. 2006)

297

Temperature (%)

Humidity (%)

5.1 Data-Driven Approach Taxonomy

Minutes since initial deployment

(a) Measured dataset

Data reported (%)

ApC: Approximate Caching Avg: Ken with Average model DjCN: Ken with DisjointCliques model of clique size N

TinyDB ApC Avg

DjC1 DjC2 DjC3 DjC4

DjC5 DjC6

(b) Percentage data reported under various schemes Fig. 5.17 Experimentation outcomes from Botanical Gardens (Chu et al. 2006)

• Ken with Disjoint-Cliques at a max clique size of 1 (DjC1) and approximate caching misses at comparable rates of about 65% in these datasets. This suggests that capturing and modeling temporal correlations alone may not be sufficient to outperform approximate caching. However, the definitive advantage of larger cliques is noticed as the data reported falls rapidly since Ken utilizes spatial correlations. Ken often has the opportunity to select and report those few nodes, which serve to strongly indicate the readings of other nodes. Approximate caching fundamentally cannot take advantage of spatial correlations and, as a result, suffers in such environments. • The Botanical Gardens dataset stronger spatial and temporal correlations yield more data reduction (21% data reported at cliques of size 5, DjC5) compared to that of Intel Research Lab (36% data reported).

298

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Considering topology-dependent costs, experimentation also explores the effect of intra-source and source–sink communication costs on the performance of the above-presented schemes. Ken, like the proposals in Deshpande et al. (2004); Jain et al. (2004), focuses on using probabilistic models to provide approximate answers efficiently. This is significant for database research in general and is particularly well suited to WSNs, since sensor data are by nature noisy and uncertain, but often drawn from fairly smooth distributions (Deshpande et al. 2005). Ken and (Deshpande et al. 2004) present two complementary views in the design space for approximate WSN queries using probabilistic models. Ken is advantageous in ensuring the faithfulness of the approximate answer to the actual sampled value, without assuming model correctness, while significantly reducing the volume of data transmitted. An extension of Ken is given in Kanagal and Deshpande (2008), where a dynamic probabilistic model (DPM) is exploited to implement a probabilistic database view, i.e., a consistent snapshot of data coming from a model with a user-friendly interface. The concept of bringing to the user such a hidden state of the sensor database can be actually implemented through model-based views (Deshpande et al. 2004). The proposed solution obtains these views by means of a DPM. An interesting application of DPMs derives the internal (hidden) state of the sensed phenomenon through the available sampled data. For example, it is possible to get the operational state of a node, i.e., if it is working or failing, based on its readings, even though a specific variable is not available in the system. The suggested model uses a particle filter approach to store the output of a DPM as a set of weighted samples (the particles). The querying system converts queries referring to the DPM view into queries suitable to the particle-based representation. The resulting queries can also be optimized to perform aggregates over requested data. Particles are updated to match the incoming data stream by performing particle filtering, and a Monte Carlo algorithm allows to estimate the state of DPMs. Experimental evaluation of the prototype implementation is tested over sensor data from the Intel Lab dataset (Madden, Intel Lab Data 2004) to demonstrate the feasibility of online modeling of streaming data and to establish the advantages of such tight integration between dynamic probabilistic models and database systems.

Time-Series Forecasting Approaches Many everyday WSN applications require sensor nodes to report approximations of their readings at regular time intervals. Time-series prediction techniques have shown to effectively reduce the communication effort for such applications, while guaranteeing user-specified accuracy requirements on collected data. The achievable communication savings offered by time-series prediction strongly depend on the type of signal sensed; nevertheless, an inadequate a priori choice of a prediction model can in practice lead to poor prediction performance. In numerous WSN deployments, sensor nodes are distributed at various locations over a region of interest and collect data at regular time intervals (Szewczyk;

5.1 Data-Driven Approach Taxonomy

299

et al. 2004; Buonadonna et al. 2005). Each sensor on a node captures a time-series representing the development of the sensed physical variable over space and time. Reporting these time-series to a sink, through the sensor nodes onboard radio, represents a significant communication overhead. As stressed throughout, the radio channel has limited capacity, and radio communication is the dominant factor of energy consumption in WSNs (Polastre et al. 2005). Hence, the development of adequate data gathering techniques able to reduce the amount of data sent throughout the network is a key factor for allowing long-term, unattended network operation. Time-series forecasting was subject of interest in the WSN research community (Lazaridis and Mehrotra 2003; Deshpande et al. 2004; Jain et al. 2004; Tulone and Madden 2006b). Sections “Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ)” and “Adaptive Model Selection for Time-Series Prediction in WSNs (AMS)” put under focus the time-series concept and techniques that aim at energy management in WSNs. Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ) The probabilistic adaptable query system (PAQ) is proposed to approximate the values of sensors in WSNs, based on time-series forecasting (Tulone and Madden 2006b). This method relies on autoregressive models (AR) built at each sensor to predict local readings. Nodes transmit these local models to a sink node, which uses them to predict sensor values without directly communicating with sensors. When needed, nodes send information about outlier readings and model updates to the sink. Such an approach can drastically reduce the amount of communication required to monitor the readings of all sensors in a network, and it also provides provably correct, user-controllable error bounds on the predicted values of each sensor. Autoregressive time-series models have been widely used to approximate and summarize time-series for applications in finance, communication, weather prediction, and a variety of other non-WSN domains (Brockwell and Davis 2016). Most of WSN deployments have a similar characteristic; data are collected at a regular rate to some centralized basestation (sink), where it is stored on disk and analyzed using conventional data processing tools, e.g., databases, mathematical analysis packages, and GIS software. One major focus of WSNs research has been on building tools to facilitate the collection of data, such as database query languages like TinyDB (Madden et al. 2005), parallel programming systems as regions (Welsh and Mainland 2004), and power-conserving and failure-masking network layers like directed diffusion (Intanagonwiwat et al. 2000). PAQ focuses on improving the performance of these data collection tools using a probabilistic approach, a class of statistical techniques broadly known as time-series forecasting. These techniques apply to phenomena evolving over time and use the recent history of readings to predict the most likely future values. A general framework is suggested to efficiently answer queries at the sink based on the simple type of time-series autoregressive (AR) model. This model is adopted because it is

300

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

computationally tractable on WSNs, unlike the fully general ARMA models (Brockwell and Davis 2016), and because it can offer a substantial reduction in communication and improvement in loss rates over other data collection approaches. The suggested AR model is evaluated both analytically and through simulation and revealed to properly model physical phenomena and accurately predict future values. A combination of AR models is used in PAQ to probabilistically answer queries; low computation cost and memory usage advocate PAQ suitability for wide range of WSN hardware. An AR model is used both globally at the sink and locally at each sensor. At the sink, it predicts the readings of individual sensors. At each sensor, the model detects when the sensor produces outlier readings, or when the model ceases to properly fit the data, thus permitting the sensor to relearn the model and notify the sink of the new model parameters. PAQ has several advantages over deterministic query systems: • Significantly reducing the amount of communication required to report the value of every sensor at the sink. • Allowing the detection of sensor readings that are outliers, in the sense that they are not consistent with recent history or have malfunctioned. • Adapting to dynamic changes in the distribution of data produced by sensors, and tolerating missing sensor data. • Abstaining from requiring a large amount of training data or a priori knowledge of the probability distribution of sensor values, and running with any of the previously mentioned abstractions, such as TinyDB, and directed diffusion for data collection. The use of probabilistic and time-series models in WSNs has its roots and was under focus in more than a scheme: • As in PAQ, the work in Jain et al. (2004); Kotidis (2005); Cohen and Kapchits (2009) relies on a combination of local and global probabilistic models which are kept in sync to reduce communication between sensor nodes and the network sink. A query framework based on Kalman filters is proposed in Jain et al. (2004), and both the sink and sensors activate a Kalman filter with user-specified accuracy when a new query is received. However, this strategy does not support multiple queries with variable precision or clustering, and the local models do not adjust to nonlinear phenomena. Similar to PAQ, the approach in Cohen and Kapchits (2009) exploits temporal/spatial correlation; however, it has a weighty learning phase that does not work well for non-stationary data. Neither Jain et al. (2004) nor Cohen and Kapchits (2009) provides a provable bound on the maximum error or on the error probability of answers provided at the sink. The snapshot queries’ approach proposed in Kotidis (2005) is also similar to PAQ in that it exploits local models and correlations, but it provides weaker guarantees. • The approaches in Cheng et al. (2003) and Deshpande et al. (2004) have shown that generative model-based approaches can significantly reduce the communication burden in WSNs. Yet, these approaches require a relatively sophisticated user who can describe the appropriate model for his domain, and usually,

5.1 Data-Driven Approach Taxonomy

•

• •

•

301

they involve a complex centralized learning phase that must be rerun if the data distribution changes. In contrast, PAQ is built on lightweight models that can be rapidly learned by the individual nodes when confronted with non-stationary distributions. How to approximate answers to queries in distributed environments with a fixed bound on the error is tackled in Olston and Widom (2002). This method, though simple, has the potential to offer far less reduction in communication than model-based approaches such as PAQ and the models discussed above. Alike PAQ, the work in Tulone (2004) uses AR models built at each sensor node to reduce communications in the context of time synchronization. This work did not focus though on querying or clustering issues. A different time-series scheme creates a piecewise linear approximation of signals generated by sensors and sends the approximations out of the network (Lazaridis and Mehrotra 2003). Different from PAQ, this scheme captures a large time-series and approximates it, rather than building a model that can be used for prediction outside of the network. Other time-series approximation methods, based on wavelets (Ganesan et al. 2003; Chen et al. 2004), have been proposed, but while reducing communication they are not as predictive as PAQ.

Importantly, there are substantial functional differences between PAQ and other probabilistic approaches of query answering: • Such approaches typically build a model centrally at the sink, using an expensive learning phase where each sensor transmits many readings to the sink. This is because they use relatively complex probabilistic models, e.g., multivariate Gaussians (Deshpande et al. 2004) or generalized graphical models, which are too complex to build or maintain on many classes of WSN hardware such as Berkeley motes (Crossbow 2002). Consequently, adapting to changes in the underlying distribution of sensor data cannot be accomplished without rerunning such an expensive learning phase. • Contrarily, PAQ relies mostly on local probabilistic models computed and maintained at each sensor. In order to adapt the local model to variations in the data distribution, each sensor continuously maintains its local model and notifies the sink only of significant changes. Communication from the sink to each of the nodes is further limited by exploiting data similarities between sensors that are geographically nearby. PAQ system relies on geographic clusters of sensors that are similar at a given point in time and that are computed by sensors that are near each other. Therefore, the sink maintains only the models’ coefficients of a few designated sensors, called cluster leaders, and uses them for prediction. To illustrate PAQ, an overview of terminology and time-series techniques is first to be introduced. The network under study comprises a dynamic set S of sensor nodes and one or more sink nodes. Each node performs sensor readings on m physical phenomena (metrics), F1, F2, …, Fm, each of which evolves over time. For example, F1 may be

302

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

temperature, and F2 is light, etc. Each sensor performs a reading on each Fi every C time units. PAQ is designed to work with sensor nodes, like Berkeley motes (Crossbow 2002), equipped with just a few kilobytes of memory, and slow 8-bit processors without floating point or dedicated signal processing hardware. Queries submitted at the sink have the form: SELECT sensorlist WHERE PðF1 ; F2 ; . . .; Fm Þ ERROR CONFIDENCE k% where P(F1, F2, …, Fm) is a predicate over F1, F2, …, Fm consisting of atoms Fi 2 ½a; b; Fi [ a; and Fi \b, such that a, b are user-specified. For instance, a valid predicate could be [(F2 2 ½a; b _ F3 2 ½c; d Þ ^ Fm [ g]. ERROR x indicates that the user is tolerant to a maximum absolute error in the query result. The CONFIDENCE clause indicates that at least k% of the readings should be within x of their true value. For instance, the user might issue the query “SELECT nodeid, temp WHERE temp > 25 °C ERROR 0.1 °C CONFIDENCE 95%”, which would report the temperature at each node to within 0.1 °C, a property that would be satisfied by 95% of the readings. As for time-series forecasting, a time-series is a set of observations xt, each of which is recorded at time t. Important in the analysis of time-series is the description of a suitable uncertainty model for the data. To allow for the possibly unpredictable nature of future observations, it is usual to suppose that each observation xt is a sample of a random variable Xt (denoted as X(t)). Two definitions are to be introduced: • Definition 1 A time-series model for the observed data {xt} is a specification of the joint distributions of the random variables {Xt} of which {xt} is a sample. Obviously, to make predictions, it is assumed that some part of this model does not vary with time. Hence, an important component of time-series modeling is to remove trend and seasonal components to get a weakly stationary time-series. Informally, a time-series {Xt} is stationary if it has statistical properties similar to those of the time-shifted series {Xt+h} for each integer h. More precisely, {Xt} is weakly stationary if its mean function lx ðtÞ and its covariance function cx ðt þ h; tÞ are independent of t for each h. Linear time-series models, which include the class of AR moving-average (ARMA) models, provide a general framework for studying stationary processes. The ARMA processes are defined by linear difference equations with constant coefficients. One of the key properties is the existence and uniqueness of stationary solutions of the defining equations (Brockwell and Davis 2016). • Definition 2 {Xt} is an ARMA (p, q) process if {Xt} is stationary and for all t, Xt  /1  Xt1  . . .  /p  Xtp ¼ Zt þ h1  Zt1 þ . . . þ hq  Ztq

ð5:14Þ

5.1 Data-Driven Approach Taxonomy

303

  where fZt g  WN ð0; r2 Þ and the polynomials 1  /1  z  . . .  /p  zp and   1  h1  z  . . .  hq  zq have no common factors, fZt g is a series of uncorrelated random variables, each with zero mean and r2 variance. Such a sequence is referred as white noise and denoted by WN ð0; r2 Þ. An autoregressive model of degree p, denoted by AR(p), is a particular type of ARMA model with q = 0 (i.e., the right-hand side of Eq. 5.14 contains just a single term). Such models are referred to as autoregressive or AR models. This model is adopted in PAQ to predict the value of Fi a sensor reads at time t. This choice was motivated by its simplicity leading to lower computational cost and memory requirements, which typically suits WSNs. Noticeably, there is an inherent tradeoff between efficiency and precision since there exist more sophisticated models, e.g., hidden Markov chains (Cowell et al. 2006), or recursive neural networks (Michalski et al. 2013) that are able to capture nonlinear data distributions but are much more computationally expensive. AR models provide the good balance between simplicity and accuracy in WSNs. A closer look at PAQ principles and functionalities is now due to: • PAQ employs a combination of statistical models with live data acquisition. Each sensor in PAQ maintains a local AR model and samples its values once every C seconds; it uses recent readings to predict future local readings. When a reading is not properly predicted by the model, the node marks the reading as an outlier or chooses to relearn the model, depending on the extent to which the reading disagrees with the model and the number of recent outliers that have been detected. This design is motivated by the need to monitor changes in the physical phenomena and detect outliers and also to reduce communication between the local sensors and the sink. Except for a cluster leader node, as below described, a node does not need to communicate while monitoring the model or during its learning phase, but only when computing and adjusting its cluster. The sink maintains one AR model per geographic cluster. A cluster is a subset of sensors that are in communication range and whose values differ by at most a constant value h; explicably, close-by sensors will often produce similar readings. Since the local models are dynamic, clusters are dynamic sets that can vary in number. One sensor in each cluster is designated as leader. It is responsible for communicating with the sink on behalf of its cluster. The leader AR model, called the cluster model, is used to predict the values of all sensors in the cluster with an error of at most h over the member sensor local models and with the same confidence. The sink maintains the coefficients associated with each of the leaders models and receives periodic readings from them; it also maintains a list of the current clusters. The leaders’ models and the cluster sets stored at the sink allow the sink to answer queries over all sensors using just the cluster models, thus avoiding a large number of message transmissions. To reduce communications, clusters are computed locally by the clusters of leaders and members of each cluster. Each leader is responsible for notifying the sink of changes in its

304

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

model coefficients, in its cluster members, and for transmitting periodic readings to the sink. • Local AR model. Because of sensors limitations, the probabilistic model maintained by the sensor must be lightweight, in terms of both the computational and storage requirements. Local models are designed with energy restrictions in mind. Since physical phenomena are typically not stationary, a time-series is usually decomposed into three components: specifically, a trend component that grows very slowly over time, a seasonal component with some periodicity, and a stationary component. However, maintaining the trend and seasonal components substantially increases the complexity of learning and adapting the model. In order to simplify PAQ model and ignore trend and seasonal components, an autoregressive model AR(p) is considered with a narrow prediction window, such as p = 3 (Tulone 2004). If the elapsed time since the last reading is relatively short, it is reasonable to neglect those components (Brockwell and Davis 2016). However, an AR(p) model is unlikely to be a good fit for nonlinear physical phenomena. In particular, WSN data are typically locally linear, but there are periodic nonlinearities that are not well-predicted by AR(p) models. To solve this hitch, PAQ linear models are enhanced with dynamic updates that are detected and performed locally. The clue is to detect when the model is no longer a good fit for the data being sensed and dynamically relearn the model coefficients on such occurrence. The efficiency of this approach comes from the fact that learning and updating the AR model are cheap compared to the costs of learning and maintaining a nonlinear model. Multivariate or univariate models, which to adopt? In WSNs, each sensor device typically has multiple sensors. To handle multiple sensors, a multivariate AR model with m components is computed; one for each physical measurement, or m univariate models, can be created. For m > 2, the computational cost of learning the multivariate model is higher than the cost associated with m univariate models. In the sensing model comprised in PAQ, each measurement reads F every C time units, where the history of these values up to time t is denoted as v1 ; . . .; vi ; ::; vt . Although the proposed approach is independent of the size of the parameter vector q, q = 3 is chosen because it simplifies the model and ignores seasonal and trend components and has low computational and storage costs as well. Therefore, each sensor Sj models F as a dynamic AR(3) time-series with Gaussian white noise of zero mean and standard deviation bðxÞ. In case of a time-series F(t) with nonzero mean g, the time-series X ðtÞ ¼ F ðtÞ  g is studied (Brockwell and Davis 2016): X ðtÞ ¼ a  X ðt  1Þ þ b  X ðt  2Þ þ c  X ðt  3Þ þ bðxÞ  N ð0; 1Þ where a; b; c 2 R, the set of real numbers.

ð5:15Þ

5.1 Data-Driven Approach Taxonomy

305

Therefore, the predictor P(t) of F at time t is given by its mean g plus a linear combination of the increments or decrements of the last three readings with respect to g. Precisely, the prediction at time t > ti–1 is given as: PðtÞ ¼ g þ a  ðvi1  gÞ þ b  ðvi2  gÞ þ c  ðvi3  gÞ

ð5:16Þ

The function b(x) represents the standard deviation of the white noise. The distribution of the noise is assumed to be invariant over time. The following lemma computes the error bound associated with the prediction P(t) of F(t) at time t. Lemma 1. Let P(t) be the prediction of F at time t associated with the model in Eq. (5.15), and let e ¼ m  bðxÞ, where m is a real-valued constant larger than 1. Then, the actual value at time t is contained in ½PðtÞ  e; PðtÞ þ e; with error probability at most 1=m2 : • Learning phase. There are two parameters that mainly affect the efficiency and accuracy of the learning phase; exclusively, the number of readings, N, collected during the learning phase, and ; C; the time interval between two consecutive readings. Given these parameters, each sensor builds the following data structures during the learning phase: – An initially empty queue V, containing the most recent N readings. – The coefficients a; b; c and the mean g. – The standard deviation b(C) of the white noise during C time units. During the learning phase, a sensor performs a reading every C time units and inserts it into V. After performing N readings, it computes the mean g from the N readings, and the coefficients a; b; c are computed at the end for efficiency reasons. The coefficients a; b; c are computed by calculating the minimum squared error between the readings contained in V and the predicted values via least-squares regression. In least-squares regression, it is supposed that v1 ; . . .; vN are the values read during the learning phase, and that
v1 ; . . .;
vN are such that
v1 ¼ vi  g for i = 1, …, N. Then, a; b; c correspond to the coefficients of the best linear predictor and are obtained by minimizing the function: Qða; b; cÞ ¼

N X ð
vi  ða 
vi1 þ b 
vi2 þ c 
vi3 ÞÞ2

ð5:17Þ

i¼4

The coefficients a; b; c can be computed by setting the partial derivatives of the minimum squared error to zero and solving a linear system of three equations (Golub and Van Loan 2012). After computing the mean g and coefficients a; b; c, the sensor computes the variance of the white noise during C time units by computing the prediction error ei ¼ Pi 
vi for i = 1,…, N, and the mean
e of e1,…,eN. Hence, the variance of the white noise during C time units is:

306

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi PN 2
ð e  e Þ i¼1 i bðCÞ ¼ ð N  1Þ

ð5:18Þ

Thus, the parameters g; a; b; c and bðCÞ uniquely describe the AR model for a given set of learning data fv1 ; . . .; vN g. The computational cost of the learning phase is the cost of reading N values, plus the cost of computing matrices A and B of the linear system A*X = B in three unknowns, which involves 12 * (N–3) sum and product operations, plus the cost of solving the linear system. As discussed throughout this section, the local AR model must be dynamic; hence, a sensor periodically monitors its local model and updates it as needed. Knowing that the sink, or another specialized node, monitors the validity of the sensor model, this design offers several benefits: • It allows the sensor to detect data anomalies compared to previous history. A data anomaly is a sensor value that the model does not predict to within a user-specified error bound. These anomalies can be classified into outlier values, which are transient mispredictions that the model does not account for or distribution changes that are persistent mispredictions advocating the model needs to be relearned either because of a faulty sensor or a fundamental shift in the data being sensed. • It requires no communication during learning and updating, in contrast with the approach taken in BBQ (Deshpande et al. 2004), as described in Section “Model-Driven Data Acquisition in Sensor Networks (BBQ)”. This is possible because of the simplicity of PAQ local model, which requires relatively small learning history and low computational cost and memory storage. However, the efficiency of this approach is related to the efficiency of monitoring and updating the AR model. To save energy, the model should be updated only when its readings diverge consistently from its model. Data anomalies are classified based on the model prediction error. Two thresholds are set d and the maximum error e ¼ m  bðCÞ (as defined in Lemma 1). These thresholds are chosen such that: • If the absolute value of the prediction error falls in ½0; d, then the model is a good predictor of the data. • If it falls in ½0; d, the data are still within the user-specified error bound, but the model might need to be updated. • If the error prediction exceeds e, then the data are an outlier because of Lemma 1. Since m is chosen such that fewer than a fraction m2 of the values will be mispredicted (if the stationarity assumption holds), a single outlier value can be neglected while still satisfying the user-specified confidence bound. However, though this might be an isolated anomaly that requires no action, it might correspond to an abrupt change in the data distribution, signifying that the data were not

5.1 Data-Driven Approach Taxonomy

307

stationary, which implies that the node should update the model and send the updated model parameters to the sink. Based on the considerations spelled out in the previous two paragraphs, the monitoring algorithm tracks the quality of the model. Each sensor starts monitoring its model just after the learning phase; it takes a reading every C time units and updates its queue V, which contains the most recent N values. If the prediction error exceeds d, it begins monitoring the next readings. While monitoring, the sensor keeps track of the number of times the prediction error is in the range ½d. . .e, and the number of times the error exceeds e. The sensor sends a notification to the sink as soon as it detects a variation in the data distribution. After K readings, if the sensor has detected variations in the data distribution, it recomputes g; a; b; c based on the values stored in V. Then, it sends the new coefficients to the sink and optionally a list of the outlier values. Out of the dynamic local AR model, several outcomes may be grasped through experimentation and simulation: • The accuracy and efficiency of the dynamic update depend on the parameters N; K; d; e; and a. N represents the length of the history that is used to compute the model. The computational cost of relearning increases linearly with N. However, the accuracy does not necessarily improve as N grows. For instance, if the data distribution is irregular, e.g., not well fit by a linear model, then a larger value of N will not result in a better fit. • The choice of error bound e offers a tradeoff between accuracy and error probability (ability to meet a user-specified confidence bound) which is inversely proportional to m2 . Moreover, the choice of m has also an impact on the number of readings marked as outliers. • The tradeoff between accuracy and efficiency is presented by the threshold d which defines, along with e and a, the conditions under which the model should be updated. Clearly, it is desirable to keep the number of updates low, since updates incur additional learning and communication costs. However, making the interval ½d; e too small will not result in an energy improvement since the model will not properly fit the data and will thus flag more readings as outliers over time. • The duration of the monitor window, K, poses another tradeoff. It is required to keep the monitor window relatively short to update the model as fast as possible, but making it too short can result in excessively frequent updates. An enhancement over PAQ was presented by the same authors, and the similarity-based adaptive framework (SAF) improves in two aspects (Tulone and Madden 2006a): • The AR model is refined so that a trend component is included in the forecast as well. This leads to a better prediction of phenomena with sharp variations of their values. SAF can detect not only outliers, but also inconsistent data. This happens when nodes cannot compute a stationary model. In this case, nodes can improve model stability in two steps:

308

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

– They can filter the data so as to smooth outliers. – They can enlarge the size of used data to decrease the impact of the outliers. When such mechanisms are not enough to get a valid model, nodes can explicitly invalidate the model stored at the sink and start rebuilding a new model from scratch. • A centralized clustering scheme that is optimal in the number of clusters is proposed; it has a complexity of O(n*logn). In SAF, queries are answered using lightweight linear time-series models built by each node from a small number of readings, enabling the models to be quickly relearned, and stored at the sink. Sensor nodes and the sink only communicate occasionally to exchange models or answer queries that require more accuracy than the stored models can provide. SAF is capable of detecting outlier values, periods of data instability and data similarities among sensor nodes. The approach of similarity detection works even under dynamic conditions, e.g., node mobility, data distribution variation, or unstable communication channels. SAF uses simple linear time-series models that consists of a time-varying function Tr, called trend component, and a stationary autoregressive (AR) component X(t) representing the overtime divergence of the phenomenon from Tr. SAF was implemented experimentally and analyzed through simulation. Interesting fallouts were ascertained: • Out of experimentation, the adopted model is capable of predicting the data produced by sensors measuring physical phenomena that evolve slowly over time like ambient light, temperature, and humidity. • It is computationally tractable on WSNs and is cheap to learn. In contrast with other approaches (Deshpande et al. 2004; Chu et al. 2006), nodes learn the models locally, requiring no communication, when a developed monitoring algorithm detects that the local model is no longer a good fit for the data. When this occurs, the sensor relearns the model and transmits its coefficients to the sink. • The stationary AR component X(t) allows to compute bounds that are used by the sink when predicting sensor values. This means that SAF does not require periodic readings to ensure a given accuracy as in PAQ. Moreover, the stationary component allows SAF to provide other interesting properties such as the ability to detect outliers and periods of data instability (highly noisy data). • SAF can detect data similarity between sensors and group them into clusters using the model coefficients stored at the sink. • Unlike previous methods, detecting node similarities requires no additional communication. SAF detects similarities at the sink and not at the sensor nodes and relies on a novel definition of data similarity based on prediction values, not on raw data as in previous approaches. More precisely, node similarity is detected based on the bounds derived from the models stored at the sink.

5.1 Data-Driven Approach Taxonomy

309

Adaptive Model Selection for Time-Series Prediction in WSNs (AMS) The adaptive model selection (AMS) is a lightweight online algorithm that allows sensor nodes to autonomously select the statistically most suitable model among a set of candidate models, differently from the approaches described in Section “Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ)” where a single model was used to represent a given quantity (Le Borgne et al. 2007). AMS extends the time-series forecasting scheme with an adaptive multi-model selection mechanism. Basically, as a priori knowledge of the monitored phenomenon could be unavailable, it would be better to let the system itself choose the right model automatically. Thus, all nodes keep a set of models, but at a given instant only one of them is used for data prediction. As a matter of significance, complex models can lead to a better prediction at the expense of a higher update cost; i.e., they need more parameters to be described properly. In AMS, all models are updated at every sampling instant, but only the current one is used for prediction. If the error between the sensed data and the current model is higher than the allowed threshold, then the current model is switched to the one satisfying the requested accuracy and minimizing the cost of update. An update procedure is performed to ensure that both source and sink nodes are synchronized to the newly selected model. To save nodes’ resources, poorly performing models are discarded over time by using a racing mechanism. Experimental results on AMS were obtained on the basis of 14 real-world sensor time-series and have shown the efficiency and versatility of the proposed AMS framework in improving communication savings. A statistical procedure known as racing (Maron and Moore 1997) is used to discard models that perform poorly over time so as to save sensor nodes computational and memory resources. In multiple WSNs application scenarios, users are interested in observing physical phenomena with a prespecified, application-dependent accuracy. Thus, gathered sensor data are usually accepted to lie within a known error bound, say ½e; þ e; e 2 R, where R is the set of real numbers. A sensor node regularly collecting local measurements can fit a prediction model to the real data and communicates it to the sink, which can then use the model to compute estimates of future sensor readings. The sensor node can then reproduce the same readings estimations and transmit a model update to the sink only if the current measurement differs from the predicted by more than e, thus avoiding unnecessary communication. In the absence of notification from the sensor node, the sink implicitly assumes that the value obtained from the shared prediction model is within the required error bound. This strategy, referred to as dual prediction scheme (DPS), may lead to high communication and energy savings if adequate prediction models are used (Lazaridis and Mehrotra 2003; Jain et al. 2004; Tulone and Madden 2006b). Typically, the model to use, e.g., constant or linear, is fixed a priori, while model parameters are estimated on the basis of incoming data. Figure 5.18 illustrates how the DPS behaves on a temperature time-series obtained from a real-world WSN deployment, when the required data accuracy is set to 0.5 °C and a simple AR model is used (MIDRA 2006). It is seen that the

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Fig. 5.18 Dual prediction scheme with error threshold emax = 0.5 °C (Le Borgne et al. 2007)

Temperature ( )

310

Time step

predicted data are within  0.5 °C of the real data up to the 1261st time step. At time t = 1262, the prediction error exceeds the tolerated threshold and the sample collected at time t = 1262 is sent to the sink. The prediction model is then updated to take into account the new acquired data; from time t = 1263 to t = 1272, the predicted measurements are again close enough to the real ones, making further communication between the sensing node and the sink unnecessary. At t = 1273, the sensor node realizes again that the sink is predicting the sensor measurements with an error higher than e and thus transmits the current reading. The procedure is repeated at t = 1286, when a new update is sent to the sink. In this illustration, out of 35 collected readings, only 3 were effectively transmitted by the sensor node, which amounted to about 90% of communication savings. Obviously, the achievable communication savings depend, among others, on the particular sensed phenomenon, on the data sampling rate and, last but not least, on the used prediction model. Continuously reporting sensor readings to a sink node at regular time intervals is the almost “default” data gathering mechanism in real WSN deployments (Szewczyk et al. 2004; Madden 2004). With respect to this “default monitoring” scheme, the DPS significantly reduces communication between a sensor node and the sink, while guaranteeing the data collected to be within a user-specified accuracy. The gains in communication offered by the DPS, however, depend on the adequacy between the prediction model used and the characteristics of the time-series captured by the sensor. The main task of the DPS is to run an identical prediction model h at both the source and the sink nodes and to use it to produce estimates of the future sensor readings, given some of the previous samples. If the predicted value differs from the actual sensor measurements by more than a given error threshold e, a model update is transmitted to the sink. The simplest implementation of the DPS uses a constant prediction model, referred to as CM, which allows the sink to reconstruct a piecewise constant

5.1 Data-Driven Approach Taxonomy

311

approximation of the real sensor signal. Using a CM, no updates are sent as long as readings collected by the sensor do not diverge by more than e from the last reading sent to the sink. When this difference becomes bigger than e, the current reading is sent; this process is repeated over time (Lazaridis and Mehrotra 2003). Such approach provides attractive communication savings with respect to the “default monitoring” strategy. On selected time-series, a CM can be outperformed by more complex prediction techniques, as shown in Jain et al. (2004); Tulone and Madden (2006b). However, all these methods depend on a number of parameters that cannot be fixed a priori in a generic manner. To get into AMS design concepts, it is worthy to recall that the approaches to perform time-series forecasting in WSNs range from simple heuristics to sophisticated modeling frameworks (Lazaridis and Mehrotra 2003; Jain et al. 2004; Tulone and Madden 2006a, b). These methods typically allow improving upon the simple monitoring approach in which measurements are continuously reported at fixed time intervals. However, they also overlook two relevant issues: • Complex prediction techniques, like Kalman filtering, rely on parameters whose identification proves to be difficult in practical settings, particularly when no a priori knowledge on the signals is available (Jain et al. 2004). These difficulties increase with the flexibility of the model, or, equivalently, with the number of parameters needed to specify it. Therefore, the more flexible the model, the less usable he is in practice. • As the DPS requires the sensor node and the sink to run the same prediction model, all the parameters of the model must be sent each time an update is needed. There exists therefore a tradeoff between the ability of a model to properly fit the signal, so as to lower the number of data transmissions and model updates, and the number of parameters that need to be computed locally at the sensor node and sent to the sink when an update is needed. AMS addresses both issues by introducing a generic procedure for adaptive model selection. The rationale of the approach is to use complex prediction models only if they prove to be efficient both in terms of computation and achievable communication savings and otherwise to rely on simpler models. A set of models of increasing complexity is considered, and the sensor nodes assess their performances online as sensor data are collected, on the basis of a metric that weights the number of updates by their size. It is possible to select, among a set of candidates, the model that offers the highest achievable communication savings. The implementation of AMS is based on AR models, whose parameters can be updated online as new observations become available, and that are computationally economical to maintain (Makridakis 1997; Box et al. 2008). A sensor node running AMS maintains a set of K candidate prediction models:   hi Xhi ;t ; hhi ;t ; 1 i K

ð5:19Þ

For each model hi , a given quality measure is recursively estimated and the model that optimizes this performance indicator is selected as the current model.

312

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

These performance measures may also be used to run the racing mechanism, which allows discarding poorly performing models from the set of candidate models. The main goal of the DPS is to reduce the number of updates between a sensor node and the sink. To measure the performance of the DPS, it is therefore meaningful to consider the relative update rate, i.e., the ratio of the number of updates effectively sent when running the DPS to the number of updates that would have been sent by the “default monitoring” scheme. Let Uhi ;t be the relative update rate for model hi at time t, where Uhi ;1 ¼ 1. Uhi ;t can be recursively computed as: Uhi ;t ¼

ðt  1Þ  Uhi ;t1 þ 1 t

ð5:20Þ

if an update is needed at time t, or otherwise as: Uhi ;t ¼

ðt  1Þ  Uhi ;t1 t

ð5:21Þ

The relative update rate reflects the percentage of transmitted packets with respect to the “default monitoring” scheme. Note that the relative update rate for the “default monitoring” scheme is 1 since it requires the transmission of the entirety of the readings, and that any lower value indicates a gain in the number of transmitted packets. However, this performance indicator does not take into account the fact that while an update in the “default monitoring” mode only consists of the current sensor readings, updating a hi requires the input values Xhi ;t and the model parameters hhi ;t . Consequently, performing a single update may require sending a high number of bytes to the sink, which may become critical in settings characterized by a very limited network bandwidth. To take into account the packet size of a single model update, an alternative performance indicator is suggested, relative data rate, defined to be: Whi ;t ¼ Uhi ;t  Chi

ð5:22Þ

where Chi is the ratio of the number of bytes required to send an update of model hi to the number of bytes required to send an update in the “default monitoring” mode. The relative data rate Whi ;t measures the savings in terms of data rate for model hi at time t with respect to the “default monitoring” mode. The need for the racing mechanism arises as it is likely that some fhi g will perform poorly; it would then be preferable not to maintain them to save computational and memory requirements (Maron and Moore 1997). The racing mechanism determines, on the basis of hypothesis testing (Hamilton 1994), what models among a set of candidate models are significantly outperformed by others. For instance, let hi ¼ argminhi Whi ;t be the model with the lowest relative data rate at time instant t among the set of candidate models fhi g, and let Dhi ;hi ¼ Whi ;t  Whi ;t be the difference between the estimated relative data rates of any model hi and hi .

5.1 Data-Driven Approach Taxonomy

313

Relying on the Hoeffding bound7 (Hoeffding 1963), a distribution-free statistical bound, the racing mechanism assumes with probability 1  d that hi truly outperforms hi if: sffiffiffiffiffiffiffiffiffiffi  ln 1d Dhi ;hi [ R  2t

ð5:23Þ

where R is the range taken by the random variable Dhi ;hi . Luckily, due to the lack of parametric assumptions, the Hoeffding bound requires no other information than the range of values taken by the random variables considered, which is known in advance. As 0 Whi ;t Chi , and 0 Whi ;t Chi ; it follows that R ¼ Chi þ Chi , and the bound for discarding model hi is therefore given by: sffiffiffiffiffiffiffiffiffiffi  ln 1d Dhi ;hi [ ðChi þ Chi Þ  2t

ð5:24Þ

The racing mechanism discards poor performing models from the set of candidates among which the AMS chooses the current model. Since the bound gets tighter as t increases, only one model is eventually maintained on the sensor node. The reasoning of the AMS algorithm can be clarified in the upcoming bullets: • It takes as inputs the error tolerance e, the number of candidate models K, the set of models fhi g, and their corresponding costs fChi g. • The first model sent to the sink is that with the lowest model cost. • When the sensor collects a new reading Xt, AMS runs the function SimulateModel, which estimates the relative update rates Uhi ;t for all candidate or not by model hi . This function first determines whether an  update is necessary  ^ checking if the current reading estimation Xt¼ hi Xhi ;t1 ; hh ;t1 , computed by i

model hi at time t, is more than e of the actual sensor value Xt. • The relative update rate Uhi ;t is then computed as described in Eqs. (5.20) and (5.21). Since the parameters of a candidate model may be updated recursively as new sensor readings become available, the function SimulateModel maintains two sets of parameters for each model hi : hhi ;t and hhi ;t . • Parameters hhi ;t are continuously updated with incoming data so that the model is constantly refined, e.g., using the recursive least-squares (RLS) procedure for AR models. As long as no update is necessary for model hi , parameters hhi ;t remain unchanged since they represent the parameters that would be shared by the sensor node with the sink if hi was the current model.

7

Upper bounds are derived for the probability that the sum S of n independent random variables exceeds its mean ES by a positive number nt. It is assumed that the range of each summand of S is bounded or bounded above.

314

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

• After running to completion, the function SimulateModel returns the control to AMS, which then behaves as a “classical” DPS scheme. It therefore checks whether the absolute value of the difference between the reading estimation   ^t þ 1¼ h Xh ;t ; hh ;t , computed at the sink using the current model h , and the X ^t does not exceed the tolerated error threshold e. If this actual sensor value X threshold is exceeded, the current model h is assigned the model in fhi g that minimizes the chosen performance indicator, and an update composed of the input values Xh ;t and the parameters hh ;t is sent to the sink. Experimental evaluation of AMS was based on AR models. Tests were performed on how AR models, whose parameters can be recursively updated, can improve upon a CM, when running the DPS. Choosing AR models is twofold: • They are both theoretically and experimentally good candidates for time-series predictions (Makridakis et al. 1997; Box et al. 2008). • Model parameters can be estimated by means of the RLS algorithm, which allows to online adapt the parameters to the underlying time-series, without the need of storing large sets of past data (Alexander 2012). Time-series forecasting using AR models is performed by regressing the value Xt of the time-series Ht at time  instant t againstthe elements of the time-series at the previous p time instants Xt1 ; Xt2 ; . . .; Xtp . The prediction at time t + 1 is thus obtained as: ^t þ 1 ¼ h1  Xt þ h2  Xt1 þ . . . þ hp  Xtp þ 1 X

ð5:25Þ

  where h1 ; h2 ; . . .; hp are the AR coefficients and p is the order of the AR model, thus denoted as AR(p).   Following the notation in Eq. (5.19), let hARð pÞ;t ¼ h1 ðtÞ; h2 ðtÞ; . . .; hp ðtÞ be   the row vector of parameters, and XARð pÞ;t ¼ Xt ; Xt1 ; . . .; Xtp þ 1 be the row vector of inputs to a model AR(p). The prediction is thus obtained by the scalar product: ^t þ 1 ¼ hARð pÞ;t  XTARð pÞ;t X

ð5:26Þ

where the superscript T is the transposition operator. The parameters hARð pÞ;t can be computed by means of the RLS algorithm, which consists of a computationally economical set of equations that allows to recursively update the parameters hARð pÞ;t as new observations Xt are available. The computational cost for an update of the vector hARð pÞ;t is of the order 3  p3 þ 5  p2 þ 4  p: The experimental evaluation was based on a set of 14 publicly available datasets, collected in real WSN deployments (Madden, Intel Lab Data 2004; MIDRA 2006; National Data Buoy Center 2018). Performance was measured in terms of gains in update rate, gains in data rate, and the racing mechanism:

5.1 Data-Driven Approach Taxonomy

315

• Gains in update rate. In most cases, models AR(1),…,AR(5) outperformed the CM; i.e., they have lower update rates; also, the performances of AR models are statistically equivalent regardless of the model order. However, the CM performed significantly better than AR models for some time-series. Deficiencies of AR models are due to the nature of those time-series, qualitatively characterized by sudden and sharp changes. These abrupt changes cause the variance in the estimation of AR coefficients to increase, making the models unstable and thus allowing a simple CM to provide better performances in terms of lower update rates. Each of the 14 datasets was studied under CM and AR(1),…,AR(5) models; there is one model for each dataset that yielded the lowest update rate, and this model is adopted for AMS. AR models outperformed CM model in 9 of the 14 datasets. • Gains in data rate. The performance of the DPS is assessed in terms of the weighted update rate Whi ;t ¼ Uhi ;t  Chi , or data rate, introduced in Eq. (5.22). Model costs Chi were computed assuming that each data sample and parameter can be stored in one byte. Accordingly, CM requires one Byte to be sent to the sink, while the update of an AR(p) model requires 2*p Bytes (p Bytes for the initial input values and p Bytes for the parameters). The packet overhead (header and footer) depends on the specific communication protocol. In the experiments, a packet overhead of Poverhead = 24 Bytes corresponds to the average overhead of the IEEE 802.15.4 protocol, a standard for WSNs (Polastre et al. 2005). The size of a packet carrying an update for an AR(p) model is therefore: CARð pÞ ¼

24 þ 2  p 24 þ 1

ð5:27Þ

The performances of CM and AR(p) models are calculated in terms of percentage of bytes sent to the sink with respect to the “default monitoring” mode. The cost of the CM is one. In contrast, there is a performance deterioration of AR models, as the cost of sending their parameters wastes to more simple models their advantage in prediction accuracy. Out of the 14 datasets, AR models only outperformed CM five times. AMS involves the model that yielded the lowest data rate for each dataset. • Racing mechanism. The convergence speed when relying on the racing mechanism is obtained. The weighted update rate Whi ;t was used to evaluate performance of competing models, with a confidence 1  d  0:95%. Efficiency of the racing in terms of rapidity in discarding poorly performing models depends on the nature of the time-series. The convergence to the best model in less that 1000 time instants was obtained in four cases. For other cases, subsets of two or three remaining models were still in competition after 1000 time instants. The performances of remaining models were in those cases ranging from less than 1% up to 5%, and the a posteriori best model was always observed to be part of the remaining set. AR(4) and AR(5) were discarded in all cases due to the overhead incurred by sending their parameters to the sink. For five time-series,

316

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

AR(3) and AR(4) were in the remaining candidate models, while for the other nine time-series, either CM, AR(1), or both were still competing after the 1000th time step. Out of experimentation, several conclusions can be drawn: • Applying the DPS with a CM yielded significant communication savings with respect to the “default monitoring” mode, ranging from about 50%, for a very tight error threshold (1=100 of the sensor signal range), to 96% for rough approximations (1=5 of the sensor signal range). • The communication savings could be further slightly reduced on average by relying on AR models, although only those with low orders were observed to provide improvements due to the additional communication overhead incurred by transmitting their parameters. • Out of the set of initial candidate models, the application of the AMS procedure allowed to select in an online manner the best performing one, which turned out to be CM for some of the signals. • An empirical guideline may be adapted, to scenarios where no a priori knowledge is available on the signal characteristics and is by applying the AMS procedure with a set of about four models, composed of CM and AR models up to the order of three.

Algorithmic Approaches As revealed throughout Sect. 5.1.1.3, several models have been proposed for data prediction in WSNs. The algorithmic approach, as well be shown in Sections “Energy-Efficient Data Collection in Distributed Sensor Environments (EEDC)” and “Buddy”, is used to get predictions based on a heuristic or behavioral characterization of the sensed phenomena. Energy-Efficient Data Collection in Distributed Sensor Environments (EEDC) The proposed energy-efficient data collection (EEDC) in distributed sensor environments considers a series of sensor models that progressively expose increasing number of power saving states. For each of the sensor models considered, quality-aware data collection mechanisms enable quality requirements of the queries to be satisfied while minimizing energy consumption (Han et al. 2004). Motivated by Olston et al. (2001), this work explores data collection protocols for sensor environments that exploit the tradeoff between application quality and energy consumption at the sensors. As a non-stopping reminder, sensors are better when power aware through shutting down components, e.g., radio, when not needed in order to conserve energy.

5.1 Data-Driven Approach Taxonomy

317

Designing a scalable data management solution to drive distributed sensor applications poses significant challenges: • Given the limited computational, communication, and storage resources at the sensors, a traditional distributed database where sensors function as nodes in a distributed system might not be feasible. To facilitate complex query processing and analysis, data might need to be migrated to repositories that reside at more powerful server(s). What a clearly impractical solution. • An alternative resort where sensor data are continuously collected at a logically centralized database might also be infeasible. Since sensor readings may change very frequently/continuously in such highly dynamic environments, blindly transmitting the sensor updates to the server will impose severe network and storage overheads. Furthermore, since communication constitutes a major source of power drain in battery-operated sensors, expensive energy cost would be incurred. The proposed EEDC considers a series of sensor models that progressively expose increasing number of power saving states. For each of the sensor models considered, quality-aware data collection mechanisms enable satisfying quality requirements of the queries while minimizing energy consumption. The system and query models, the data collection framework, and the formal characterization of the sensor data collection problem are to be illustrated in what follows: • System and query model. The considered system consists of a set of n sensors and one server residing at a resource abundant node that maintains a database. Each sensor is assumed to communicate directly to the server. The proposed solution can serve as a building block for large-scale distributed sensor system. Each incoming query Qi is associated with an accuracy constraint Ai indicating its tolerance to error in reply precision. Furthermore, a query has a latency bound D, which requires that each query be answered within D time units. Each sensor node has a processor with limited memory, an embedded sensor, an analog-to-digital converter, and radio circuitry. A micro-operating system controls each component. Different radio modes are considered, while all other components are assumed always ON. There are three sensor states: explicitly, active (a), listening (l), and sleeping (s). While the sensor is in the active/listening mode, the transmitter/receiver is ON; when the sensor goes to the sleeping state, its radio is turned OFF completely. There exist two types of micro-sensor devices. A Berkeley MICA mote that has only one radio, either transmitting or receiving data (Crossbow 2002), and an MIT AMPS that has two radios, and can transmit and receive data simultaneously (Sinha and Chandrakasan 2001). These two types of sensor models are illustrated in Fig. 5.19. When a radio is in the idle mode, it is capable of detecting an incoming packet, but not in the process of receiving a packet. This mode is classified as the listening state since Tx is OFF and Rx is ON. Even in the sleeping state, the changes in sensor values can still be detected, since the sensor and processor are always ON. In EEDC, it is assumed that a sensor node has two radios.

318

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Fig. 5.19 Sensor radio modes (Han et al. 2004)

Radio modes 1-radio node 2-radio node Tx ON, Rx OFF Tx ON, Rx ON Tx OFF, Rx ON Tx OFF, Rx OFF

Sensor state Active (a) Listening (l) Sleeping (s)

• Data collection framework. An effective methodology to exploit the tradeoff between application quality and data imprecision is through having the server maintain an approximate value of the data, whose divergence from the true value is guaranteed to be error bounded at any time. Precisely, let S ¼ fs1 ; . . .; sn g be the set of sensors. Each sensor hosts its exact value that may change frequently. For each, si 2 S let vi denote the value stored at sensor si . The approximation of vi is represented by a range ri with lower bound li and upper bound ui : ri ¼ ½li ; ui  stored in the database at the server. A query for the value of sensor si is answered in the format of a range with a lower and an upper bound, so the answer accuracy is defined by the range size ui–li. The accuracy constraint Ai of query Qi specifies the maximum acceptable width of the result. Whenever the sensor value vi changes to v0i , sensor si checks whether ri is still a valid approximation for the new value. If vi falls outside ri , a new approximation of v0i is sent to the server to update the database; this process is called source-initiated update. Otherwise, there is no need to transmit the update to the server, hence reducing communication overhead. Queries are executed over the cached ranges at the server. If the error tolerance of the query is larger than the data error, i.e., Ai  ui  li , it is processed without communication with the sensor. Otherwise, the approximation offered by the database is insufficient; the server may request the exact value from a remote sensor. The sensor responds with current exact value and a new approximation to be used by subsequent queries. This process is called consumer-initiated request and update. Figure 5.20 illustrates the data collection process. • Problem statement. Given m user queries, the objective is to minimize the sensor energy consumption in the process of answering all m queries. Since a sensor consumes energy even when it is not transmitting or receiving data, it is required besides reducing the communication overhead between a sensor and the server, to minimize the time a sensor is either active and/or listening even when not transmitting updates to the server. Assuming that the probability of source- and consumer-initiated updates at each time instant is Psu and Pcu, it is required to:

5.1 Data-Driven Approach Taxonomy

319

Queries Q1, …, Qm (accuracy constraints A1, …, Am) … Source-initiated update

Sensor si

Server Consumer-initiated request

Database (ri=[li,ui])

Consumer-initiated update Fig. 5.20 Data collection process (Han et al. 2004)


¼ Esu  Psu þ Ecu  Pcu þ Eextra minimize E subject to ai Ai ; 0 i m; ti D; 0 i m where, ai ti Esu Ecu

is the answer accuracy for query i, is the query response time, is the energy required to send an update to the server, and is the energy required to both receive the request for the data and for transmitting the sensor value to the server.

Note that Esu and Ecu are not constant, they depend on the state where the sensor was when the source-initiated update and consumer-initiated update occurred. For example, for a sensor that is in the sleeping state, if its value changes exceeding the range associated at the server, it will first have to transition to the active state followed by transmitting the value to the server. Thus, the total energy consumed would be the sum of energy spent to transition from the sleeping state to the active state, and the energy spent to transmit the update to the server. Contrarily, if the value divergence occurs when the sensor is in the active state, the energy consumption would be only for transmitting the update to the server. Eextra is the amount of energy consumed while not receiving or transmitting any data; such energy is not constant and depends on the state of a sensor while being free. To achieve the objective of minimizing the energy consumption at the sensor, two issues must be addressed: – How to maintain the database. An optimal range needs to be maintained and adjusted for each sensor so that it reduces sensor energy consumption while still being able to satisfy query accuracy constraints. If the range is high, the accuracy constraints of many queries will be violated resulting in expensive probes; likewise, if the range is small, sensor update would needlessly be transmitted frequently to the server. Both cases will consume a large amount of energy.

320

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

– How to manage sensor state. It is needed to determine sensor state transition strategies. Sensors consume power not only when sending and receiving data, but also when idling at the active and listening states. To save energy, a sensor has to power down into a lower energy state. Powering down a sensor requires additional cost to power up when a request that needs to be processed arrives. Furthermore, it could result in increased latency for queries. Data precision adjustment is to be straightaway detailed, in the sense of how the approximation range for the sensor can be set at the server to minimize the energy consumption due to communication between sensors and the server. The energy cost due to communication depends on the number of source- and consumer-initiated updates, which in turn depend on the range size adaptation, the patterns of the changes in sensor values, and the query workload characteristics. The data collection strategies for sensor environments encompass a series of sensor models based on power-saving sensor states. For each model, there is a study of how to determine the ranges ri such that the overall energy consumption is minimized while meeting the quality constraint of the query. Table 5.10 summarizes the symbols used throughout. Four models are proposed: specifically, always active (AA), active–listening (AL), active–sleeping (AS), and active–listening–sleeping (ALS): • Always active (AA) model. In this model, sensors are always active. The total normalized energy consumption is:
aa ¼ PCa  ðPsu  Ttx Þ þ PCa  ðPcu  ðTrx þ Ttx ÞÞ E þ PCa  ½1  Psu  Ttx  Pcu  ðTrx þ Ttx Þ ¼ PCa

ð5:28Þ

Obviously, the normalized energy consumption is equal to the power consumption at the active state. Therefore, irrespective of how the range is set, the energy consumption is constant. This model serves as a baseline to study the energy savings that lead to utilizing sensor states that consume less energy.

Table 5.10 Symbols used (Han et al. 2004) Symbol

Meaning

r Psu Pcu Pi T rx T tx T ij PCi E ij

Interval size Probability of source-initiated update Probability of consumer-initiated update Probability of being in state i (i = a, l, s) Time needed for receiving a consumer-initiated request Time needed for sending an update Transition time from state i to state j (i, j = a, l, s) Power consumption when a sensor is in state i Energy consumed in switching from state i to state j

5.1 Data-Driven Approach Taxonomy

321

Upon first source- or consumer-initiated update

Active

Listening Waiting Ta after processing last source- or consumer-initiated update Fig. 5.21 Active–listening model (AL) (Han et al. 2004)

• Active–listening (AL) model. The sensor has two states, active and listening; initially, it is in the listening mode. The sensor moves to the active state if either the sensor value diverges from the range that represents it at the server, or if the sensor receives from the server a request for its current value (Fig. 5.21). When a sensor is in the active state, it processes all its pending requests and waits for Ta units of time before switching to the listening mode. Switching from a lower energy state to a higher energy state is associated with a significant energy cost, that is why waiting for Ta time units in the active (higher energy) state is opted over powering down to the listening (lower energy) state immediately. From an energy perspective, it might be advantageous to wait in the higher energy state, instead of powering down, if the sensor is to transition back quickly to higher energy state. Obviously, the optimal value of Ta that minimizes energy consumption depends on the application workload and sensor value change patterns. It is assumed that Ta has been optimally set; the problem is to optimally determine the range r for the sensor that minimizes the energy consumption. The sensor energy consumption under this model is as follows:
al ¼ ½Pl  ðEla þ PCa  Ttx Þ þ Pa  ðPCa  Ttx Þ  Psu E þ ½Pl  ðPCl  Trx þ Ela þ PCa  Ttx Þ þ Pa  ðPCa  ðTrx þ Ttx Þ  Pcu þ ðPCa  Pa þ PCl  Pl Þ

ð5:29Þ

 ½1  ðPl  ðTla þ Ttx Þ þ Pa  Ttx Þ  Psu  ðPl  ðTrx þ Tla þ Ttx Þ þ Pa  ðTrx þ Ttx ÞÞ  Pcu  The energy consumption depends on the probabilities Pa and Pl of the sensor being in the active and listening state. These probabilities can be expressed in terms of the probability of source- and consumer-initiated updates. If, Ta = 0, the sensor state transition matrix capturing the state transition probabilities is as follows:

322

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

2

Listening 1  Pla 1  Paa

P ¼ 4 Listening Active

3 Active Pla ¼ Psu þ Pcu 5 Paa ¼ Psu þ Pcu

ð5:30Þ

The long-term probability that the system will be in each state can be obtained by computing the steady-state vector of the Markov chain. Therefore: Pl ¼ 1  Psu  Pcu

ð5:31Þ

Pa ¼ Psu þ Pcu

ð5:32Þ

where from Olston et al. (2001), Psu ¼ K1 =r 2 Pcu ¼ K2  r; r is the range size; K1 and K2 are the model parameters that depend on the characteristics of source updates and queries.
al , the root of the derivative d E
al =dr can be obtained, To find the minimum E then: r ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p 3 2  K1 =K2

ð5:33Þ

At this optimal point, we have: Pcu K2 3 ¼ r Psu K1 Thus, energy consumption is minimized when the ratio Pcu =Psu ¼ 2; which is a constant. Following similar derivation for the case when Ta > 0, the same conclusion can be observed. • Active–sleeping (AS) model. In this model, the sensor toggles between the sleeping and active modes. Initially, the sensor is in the sleeping state. Similar to the AL model, a sensor moves to the active state if its value diverges from the range that represents it at the server (Fig. 5.22). Since a sensor in the sleeping state cannot receive a request from the server, it periodically wakes up on a timeout if it has been sleeping uninterrupted for Ts time units. Such a timeout-based transition is necessary in order to meet the

5.1 Data-Driven Approach Taxonomy

323

Upon first source-initiated update

Sleeping

After Ts without traffic

Active

Waiting Ta after processing last source- or consumer-initiated update Fig. 5.22 Active–sleeping model (AS) (Han et al. 2004)

quality requirements of queries that might occur in the consumer-initiated update at the sensor. The sensor, on switching to the active state, ends its current value at the server. Such update can be used by the server to answer queries issued by consumer-initiated requests, while the sensor was in the sleeping state. The sensor remains in the active state while there are requests for its value; it switches to the sleeping state after waiting for Ta time units without handling any requests. In the AS model, the total energy consumption consists of (Eq. 5.34): – Energy consumed by source-initiated updates. Besides the energy spent in transmitting source-initiated updates, energy is spent in transitioning from sleeping to active if the sensor was sleeping when a source-initiated update is due. – Energy consumed by transitioning from the sleeping state to the active state and the associated value updates, when a sensor wakes up on timeout. – Energy consumed by consumer-initiated updates. – Energy consumed while sleeping or being active without receiving or transmitting.
as ¼ ½ðEsa þ PCa  Ttx Þ  Ps þ PCa  Ttx  Pa   Psu E þ ðPsa  Psu  Ps Þ  ðEsa þ PCa  Ttx Þ þ ½PCa  ðTrx þ Ttx Þ  Pa  Pcu þ ðPCa  Pa þ PCs  Ps Þ

ð5:34Þ

 ½1  ððTsa þ Ttx Þ  Ps þ Ttx  Pa Þ  Psu þ ðPsa  Psu  Ps Þ  ðTsa þ Ttx Þ þ ðTrx  Ttx Þ  Pcu  Pa  Under the assumption that Ta has already been set, the optimal setting of the sensor range ri that minimizes the energy consumption is to be derived. If Ta = 0, the sensor switches to the sleeping state once there are no requests waiting. Hence:

324

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Psa ¼ Psu  d1=Ts  Psu e ¼ 1=Ts þ a  Psu ; 0 a\1

ð5:35Þ

Paa ¼ Psu þ Pcu

ð5:36Þ

The sensor state transition matrix is as follows: 2

3 Active Psa 5 Paa

Listening 1  Psa 1  Paa

P ¼ 4 Listening Active

ð5:37Þ

The long-term probabilities of the sensor being in the sleeping and active states are as follows: Ps ¼

1  Psu  Pcu 1=Ts þ a  Psu þ 1  Psu  Pcu

ð5:38Þ

Pa ¼

1=Ts þ a  Psu 1=Ts þ a  Psu þ 1  Psu  Pcu

ð5:39Þ

If Ta > 0, the sensor stays active for a period of time so that bursty update requests can be processed without state switching. The state transition probabilities are to be derived: – The probability of switching from sleeping to active is given by Eq. (5.35). – The probability of switching from active to sleeping is:  Pas ¼ ðPsu þ Pcu Þ 



1 Ta  ðPsu þ Pcu Þ

 1

1 ¼  b  ðPsu þ Pcu Þ; 0\b 1 Ta

ð5:40Þ

– The sensor state transition matrix is as follows: 2 P ¼ 4 Listening Active

Listening 1  Psa Pas

3 Active Psa 5 1  Pas

ð5:41Þ

– The long-term state probabilities are as follows: Ps ¼

1 Ta 1 Ta

 b  ðPsu þ Pcu Þ

 b  ðPsu þ Pcu Þ þ 1=Ts þ a  Psu

ð5:42Þ

5.1 Data-Driven Approach Taxonomy

Pa ¼

1 Ta

1=Ts þ a  Psu  b  ðPsu þ Pcu Þ þ 1=Ts þ a  Psu

325

ð5:43Þ

For both Ta = 0 and Ta > 0, by applying the probability formulae into Eq. (5.34), the total energy consumption can be expressed as a ratio of two complex polynomials of size r (Han et al. 2004). Since it is not possible to express the ratio Pcu to Psu in terms of other parameters, it is needed to monitor parameters K1, K2, a and b at runtime. For this purpose, the following information of the sliding window of last k updates is maintained: – The number of sensor state transitions Nsa and Na of the last k updates. – The number of source- or consumer-initiated updates Nsu or Ncu of the last k updates. Using this information, the values of K1, K2, a and b are estimated. For example, K1 is set to be Psu  r 2 where Psu is estimated to be the number of source-initiated updates, Nsu, divided by T, where T is the time period of the current window. The parameter K2 can be estimated similarly. Given these
as =dr are calculated and the energy values at parameter values, the roots of d E the roots are compared to determine the value of r that minimizes the energy
as . Since the computation is too complex to be performed at consumption E the resource-constrained sensors, it is done at the server. As the values of a and b depend on the number of sensor state transitions during the window, the sensors determine the values of a and b and piggyback them for the last k updates with the kth update. At the same time, the server monitors K1 and K2; upon receiving the kth update, it computes the new optimal range to be transmitted to the sensor. • Active–listening–sleeping model (ALS). The sensor is initially in the sleeping state; it toggles to the active state when a source-initiated update occurs or when it has been sleeping for Ts time units without interruption (Fig. 5.23). In the active state, the sensor processes all the waiting requests; after being free in the active state for Ta time units, it goes to the listening state. Once in the listening state, any source-initiated update or consumer-initiated update triggers the sensor to go to the active state; otherwise, it goes to sleep, if idle for Tl time units. The sensor energy consumption is calculated to be:

326

5 Energy Management Techniques for WSNs (2): Data-Driven Approach Upon first source-initiated update or after Ts

Sleeping

Active

Upon first source- or consumerinitiated update After Tl without transition Waiting Ta after processing last source- or consumer-initiated update

Listening

Fig. 5.23 Active–listening–sleeping model (ALS) (Han et al. 2004)


als ¼ðPs  Esa þ Pl  Ela þ PCa  Ttx Þ  Psu E þ ðPsa  Psu  Ps Þ  ðEsa þ PCa  Ttx Þ þ ½Pl  ðPCl  Trx þ Ela þ PCa  Ttx Þ þ PCa  Pa  ðTrx þ Ttx Þ  Pcu þ ðPCa  Pa þ PCl  Pl þ PCs  Ps Þ

ð5:44Þ

 ½1  ðPs  Tsa þ Pl  Tla þ Ttx Þ  Psu  ðPsa  Psu  Ps Þ  ðTsa þ Ttx Þ  ððPl þ Pa Þ  ðTrx þ Ttx Þ þ Pl  Tla Þ  Pcu  Similar to the AS model, the optimal range size can be set based on the parameters monitored at runtime (Han et al. 2004). Also, the same data collection approach is applicable. Adaptive sensor state transition is the concern of the coming paragraphs, specifically how to obtain optimal Ta. In the various sensor models presented above, transitions among states besides being triggered by the sensor value diverging from its representation at the server also occur due to timeouts. In AL, AS, and ALS models, Ta is to be determined. The AS model is used to display how an optimal value of Ta can be derived. Waking up a sensor at the sleeping state necessitates additional energy and latency, so putting the sensor to sleep immediately after it accomplishes the ongoing requests, is not necessarily the most energy-efficient choice. Depending on the request arrival rate, the sleeping period could be so short such that the powering-up

5.1 Data-Driven Approach Taxonomy

327

costs are greater than the energy saved in that state. On the other hand, waiting too long to power down may not achieve the best energy reductions possible. Thus, a careful selection of Ta is essential. Intuitively, if updates, either initiated by source or consumer, are not bursty, it is worthy to set Ta to zero; otherwise, the sensor should remain active for a while before going to sleep, so that more requests can be answered timely, and frequent state switching avoided. Hence, good understanding of source- and consumer-initiated update patterns helps determining the optimal active time. Assuming that f(t) is the probability of receiving any type of requests at any time instant t. Also, let pðtÞ be the probability of being silent for t time units; i.e., there are no requests before t until a request arrives at time t. Since any incoming request means the end of the silent period, then pðtÞ ¼ f ðtÞ. If either source- or consumer-initiated update requests are uniformly distributed in the interval (0, Ta + Ts], knowing that at the end of Ts there must be a timeout update request, then: pðtÞ ¼ f ðtÞ ¼

1 Ta þ Ts

ð5:45Þ

In this case, the expected energy consumption for a single silent period will be: ZTa pðtÞ  PCa  tdt

E¼ 0

TZ a þ Ts

pðtÞ  ½PCa  Ta þ PCs  ðt  Ta Þ þ Esa dt

þ

ð5:46Þ

Ta

  PCa  Ta2 þ 2  PCa  Ts  Ta þ PCs  Ts2 þ 2  Esa  Ts ¼ 2  ðTa þ Ts Þ Since Ta  0; and E > 0, E is non-decreasing and is minimal when Ta = 0; i.e., the sensor should go to sleep immediately after it fulfills all ongoing requests. While Ta = 0 is optimal if requests inter arrival pattern follows the uniform distribution, this assumption is yet unrealistic, the problem of finding pðtÞ thus persists. Hence, the approach is learning pðtÞ at runtime and adaptively selecting Ta accordingly. The basic idea is choosing a window size w in advance. The algorithm keeps track of the last w idle period lengths and summarizes this information in a histogram. Periodically, the histogram is used to generate a new Ta.

328

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

The set of all possible interarrival period lengths (0, Ta + Ts] is partitioned into n intervals, where n is the number of bins in the histogram. Let ti be the left endpoint of the ith interval. The ith bin has a counter that indicates the number of idle periods among the last w idle periods whose length falls in the range [ti, ti+1). The bins are numbered from 0 to n–1 and ti = 0, tn = Ta + Ts. The counter for bin i is denoted by ci. The threshold for changing states is selected among n possibilities, t0, …, tn–1. The distribution p generates an idle P period of length ti with probability ci =w for 8 if0; . . .; n  1g; n1 i¼0 ci ¼ w. Thus, Ta is chosen to be the value tm that minimizes the energy consumption: ( mintm

m1 X cj j¼1

þ

w

 PCa  tj

n X cj j¼m

w

    PCa  tm þ PCs  tj  tm þ Esa

)

ð5:47Þ

A similar derivation obtains Ta for AL and ALS models and Tl for the ALS model (Han et al. 2004). The performance of the proposed models was checked through simulation with the goal of comparing the performance of sensor state models AA, AL, AS, and ALS for quality-aware data collection in terms of energy consumption and average query response time. The built C simulator consists of a server, a database and a number of sensors. User queries are sent to the server which returns their replies. User queries’ arrival times at the server are Poisson distributed with mean interarrival time which is set at 2 s. Each query is accompanied by an accuracy constraint specifying the maximum acceptable width of the result. The accuracy constraints are sampled uniformly for the range [0,40]. The performance analysis findings were meaningful: • Energy consumption and query response time (Fig. 5.24). Several outcomes were acquired: – The energy consumption of the AA model is the highest, and its query response time is the lowest, as to be expected. In the AA model, no energy is saved; sensors are always active; thus, any consumer-initiated requests can be detected and processed immediately. – The AL model does not decrease energy consumption to a great extent, since the listening state consumes an amount of power similar to the active state. Most of the time, the sensor is in the listening state when requests arrive; it switches to the active state to send out updates. This power-up process consumes time, which explains why query response time under the AL model is higher than in the AA model. – Models that incorporate sleeping state reduce energy consumption significantly. This comes through at the price of higher query response time, since

Normalized sensor energy consumption ( J)

5.1 Data-Driven Approach Taxonomy

AA

329

AL

AS

ALS

Average query response time (μsec)

(a) Sensor energy comparison

AA

AL

AS

ALS

(b) Query response time comparison Fig. 5.24 Energy consumption and query response time comparison (Han et al. 2004)

it takes more time for a sensor to switch from the sleeping state to the active state than from the listening state to the active state. – Given that decreasing sensor energy consumption is the target, the AS model outperforms the other models significantly due to its low-energy cost. • Impact of Ta adaptation on system performance (Fig. 5.25) Since Ta = 0 was shown to be optimal when requests arrival follows a uniform distribution, the adaptive approach is compared to the approach that fixes Ta to 0. It was revealed that: – Adaptive Ta saves energy by half and also decreases query response time. – When requests are bursty, adaptive Ta saves energy and shortens query waiting time by remaining active for a certain period of time. – Adapting Ta to user query patterns and sensor value changes performs better than fixing its value. • Impact of range size adaptation on system performance (Fig. 5.26). The impact of range size r was tested for four cases; it was grasped that: – At r = 0, the database stores single instantaneous values instead of intervals. All queries can be answered by just retrieving values from the database, so query response time is minimized, but every change in sensor values is to be reported to the server, thus consuming considerable energy. – For r set to be the average accuracy constraint, queries can on the average be satisfied by stored values.

5 Energy Management Techniques for WSNs (2): Data-Driven Approach Normalized sensor energy consumption ( J)

330

Static Ta (0)

Adaptive Ta

Average query response time (μsec)

(a) Impact of Ta selection on sensor energy consumption

Static Ta (0)

Adaptive Ta

(b) Impact of Ta selection on query response time

Normalized sensor energy consumption ( J)

Fig. 5.25 Impact of Ta adaptation on system performance (Han et al. 2004)

Fixed (0)

Average accuracy constant

Adaptive adjustment

Fixed (large)

Average query response time ( sec)

(a) Impact of range size adjustment on sensor energy consumption

Fixed (0)

Average accuracy constant

Adaptive adjustment

Fixed (large)

(b) Impact of range size adjustment on query reponse time Fig. 5.26 Impact of range size adaptation on system performance (Han et al. 2004)

5.1 Data-Driven Approach Taxonomy

331

– For adaptive r, it was found that optimal r periodically minimizes the energy consumption. – At large r, most source value changes will not exceed current range, so the likelihood of source-initiated updates is low. However, the coarse data representation is insufficient for most of the queries; hence, a number of consumer-initiated updates will occur. As a result, the average query response time is significantly high. To prolong system lifetime of WSNs, various approaches have been devised to exploit low duty-cycle operation or the cooperation among sensor nodes. Differently, EEDC saves energy by taking into consideration application-level information to optimize sensor energy consumption. Since many real-word applications can tolerate data imprecision at varying levels, the error tolerance of applications can be exploited to reduce energy consumption during sensor data collection. EEDC explores the tradeoff between sensor data accuracy and energy consumption in a distributed sensor environment. Buddy Buddy protocol acknowledges that radio is the most energy-consuming resource in a sensor node. The energy cost of transmitting/receiving one byte is at least an order of magnitude higher than that for sensing and computation. Also, idle listening, when radio is ON and idle waiting for communication from its neighbors, is the dominant factor in energy consumption. This suggests that a sensor node should not only transmit as few bytes as possible, and it should also turn OFF its radio as much as possible. Buddy achieves this objective by exploiting the temporal correlation in the readings of sensor nodes (Goel et al. 2006). The key idea is that two neighboring nodes can help reduce each other energy consumption by entering into a collaborative buddy8 (Merriam-Webster 2018b) relationship. These buddies take turns in keeping their radio ON. At any point of time, the node that has its radio ON also acts as a representative for its buddies, answering queries and participating in monitoring operation on their behalf. Buddies exchange information in order to make sure that the representative node response on behalf of its buddies meets the accuracy constraint. Such a collaborative arrangement allows both nodes to cut down their energy consumption by a factor of 2, in the ideal case. This basic idea can be extended to a group of N collaborating nodes; these nodes take turns to keep their radio ON. The representative node responds on behalf of its N − 1 buddies. In the ideal case, this cuts down the energy consumption by a factor of N. At a conceptual level, when all sensor nodes act individually, they can be seen as burning their batteries in parallel. Buddy protocol may be viewed as an attempt to serialize the burning of batteries of

8

– Companion, partner. – Friend.

332

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

these sensor nodes, as much as possible, thereby extending their lifetime. While saving energy, buddy protocol also ensures that three key characteristics of the network are maintained: • Any degradation in the quality of monitoring operation is bounded by networkor application-specified constraints. • Nodes meet the constraints on per-hop communication delay. • Turning OFF node radio does not adversely impact the reachability of the network. These three characteristics differentiate buddy from other distributed clustering protocols, e.g., GAF (Xu et al. 2001), detailed in the preceding chapter, and CEC (Xu et al. 2003). Such protocols divide the nodes in the network into clusters, such that even if any one node from each cluster is ON, the reachability in the network is preserved. They save energy by turning OFF radios of nodes that are redundant with respect to connectivity. However, such solutions do not consider the fact that while nodes may be redundant with respect to connectivity, their readings may not be redundant. These solutions save energy at the cost of unbounded reduction in quality of sensing. Buddy protocol attempts to save energy while at the same time meeting the constraint on quality of monitoring. The system model is a multihop WSN that supports the monitoring operation. A monitoring operation requires sensor nodes to report their readings every Ta units of time. It is assumed that with each query or monitoring operation, the application specifies an error threshold, c. This defines the maximum acceptable difference between the readings received by the application and the true readings of the sensor. Also, all nodes must guarantee a maximum per-hop communication delay of Td. Buddy is built upon the prediction-based monitoring (PREMON) paradigm introduced by the same authors (Goel and Imielinski 2001). PREMON exploits correlation in sensor readings in order to save energy. Instead of reporting readings periodically, a sensor node in PREMON generates a prediction of its readings, encodes it concisely as a prediction model, and sends it to the monitoring entity. The sensor node also specifies the duration for which the prediction model is valid; during this lifetime, it transmits only those readings that differ from the predicted readings by more than a certain prespecified error threshold, c. Such readings are termed as “violations.” In the absence of transmissions from the sensor node, the monitoring entity uses the sensor prediction model to determine the readings of the sensor node. At the end of the prediction model lifetime, the sensor generates a new prediction model and sends it to the monitoring entity. PREMON paradigm trades increased computation, due to deriving prediction models, against saving the number of transmissions. For Berkeley mote, the energy cost of transmitting one packet of 30 Bytes is 240 µJ, given that the transmission cost is 1 µJ/bit (Hill et al. 2000). With this energy, one can perform 30,000 machine instructions, knowing that the cost of performing 100 machine instructions is approximately 0.8 µJ. Given these relative costs, saving a few transmissions would fairly compensate for the cost of computing a prediction model.

5.1 Data-Driven Approach Taxonomy

333

Based on the characteristics of the application, the network designer may prespecify a set of basis functions, F ¼ ff1 ðtÞ; f2 ðtÞ; . . .; fn ðtÞg. The prediction model, M(t), may then be concisely expressed as a vector of coefficients, W ¼ fw1 ; w2 ; . . .; wn g, such that: M ðtÞ ¼ w1  f1 ðtÞ þ w2  f2 ðtÞ þ . . . þ wn  fn ðtÞ

ð5:48Þ

where, t ¼ t1 ; t1 þ Ta ; t1 þ 2  Ta ; . . .; t2 ; Ta is the reporting interval requested by the application. M(t) is defined only at these discrete instants of time; at all other points, its value is undefined. At points where M(t) is defined, the sensor node makes sure that the prediction model is accurate within c, by sending violation, if required. Computing a prediction model may involve simple linear regression, or it may use sophisticated regression techniques (Guestrin et al. 2004). As a test of how efficient is simply predicting data in the real world, a simplistic linear regression model is adopted for generating predictions. Specifically, K readings of the sensor are used to estimate the coefficients of the linear model, M ðtÞ ¼ w1 þ w2  t, which is a best fit (minimum mean square error). The resulting model is the predictor for the next D readings. After D readings, the values of coefficients w1 and w2 of the model are estimated once again, using the last K readings. The parameter values chosen for the test were K = 5, Δ = 10, and the error threshold c = 3 units. The results indicated that this simple prediction model correctly predicted 91% of the readings of a temperature sensor and 57% of the readings for a humidity sensor. This test illustrated how a simple prediction model can save more than 50% transmissions from the sensor node. In buddy, each node attempts to establish a buddy relationship with its neighbors, in order to exploit correlation in its readings. This causes the formation of a number of buddy groups in the network. Each buddy group has one representative node responsible for answering queries and participating in the monitoring operation on behalf of all the other nodes in the group. In the ideal case, all other nodes may turn their radio OFF, which saves significant energy. The members of the buddy group collaborate to rotate the responsibility of being the representative, so as to spread the consumption of energy uniformly over the group members. Two key characteristics must be preserved in order to make this energy-saving mechanism transparent to the application: specifically, the quality of results returned by the network and the reachability in the network: • Quality of results returned by the network. The formation of buddy groups should not result in unbearable level of degradation in the quality of results returned by the network as response to queries or monitoring operation. In order

334

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

for the representative node to respond on behalf of other nodes in the buddy group and still meet constraint on quality, it must know that their readings are accurate to within c. A straightforward approach to meet such requirement would be to have all nodes in the buddy group treat the representative node as a cache (or proxy) and keep its state “consistent” by sending updates. A more general and promising approach is to use PREMON paradigm, where the nodes may send their prediction model to the representative node. Subsequently, they only report violations to the representative node. When the readings of the node are highly correlated, substantial reduction in transmissions might be obtained. As long as the prediction model at the representative node is close to the actual readings of the sensor, i.e., sensor readings are predictable, it can switch OFF its radio. Hence, idle listening is trimmed, as a dominant factor in energy consumption at a sensor node (Ye et al. 2002). Using PREMON mode does not always result in energy savings. During the lifetime of a sensor node, there are intervals where the readings are highly unpredictable, e.g., the readings of magnetometer sensor in battlefield as a tank passes by. These unpredictable readings represent “interesting” events and, in many cases, are precisely the ones that are of attention to the network users. During such intervals, it may be more energy-efficient to operate in DEFAULT mode, wherein the node reports readings at regular intervals. Buddy protocol allows a node to intelligently choose its mode of operation so as to save energy. • Reachability in the network. Energy savings increase when fewer nodes keep their radio ON. However, when many nodes turn their radio OFF, the reachability in the network may be affected. For example, in a chain network, if any of the intermediate nodes turns its radio OFF, it would partition the network. The goal of preserving reachability influences the way buddy groups may be formed. Furthermore, a node impact on network reachability is dependent on the decision of its neighbors. A simple distributed mechanism is needed to guide the formation of buddy relationships while preserving reachability in the network. One possible solution is to make use of one of the distributed clustering algorithms, e.g., GAF (Xu et al. 2001), CEC (Xu et al. 2003). The goal of such algorithms is to divide network nodes into clusters, so that even when any one node from each cluster has its radio ON, the reachability of the network is still preserved. In buddy, GAF is used for clustering. With GAF, all nodes within a cluster are equivalent with respect to connectivity. Only one of them, the cluster head, needs to keep its radio ON; it is elected by the nodes using a cluster-head election algorithm. Cluster-head election algorithm is run periodically in order to rotate this responsibility among nodes in the cluster, thereby spreading the energy consumption uniformly over all the nodes. In buddy, each cluster is a buddy group, and the cluster head is the representative node for the group.

5.1 Data-Driven Approach Taxonomy

335

Buddy protocol works through several basics: • A distributed clustering protocol is used to group the nodes in clusters. The set of cluster heads form the routing backbone of the network. Within a cluster, nodes follow a style of operation similar to power saving mode (PSM) of IEEE 802.11 (Tseng et al. 2003). • The cluster head performs the role of access point, buffering data for all the nodes in the cluster. The cluster head indicates the presence of buffered data by using MAC-level beacons that are sent out periodically. • In order for nodes to meet the per-hop communication delay requirements, the nodes wake up every Td units of time to receive a beacon from the cluster head. It is assumed that Td is an integer multiple of the beacon period. • Within each cluster, every node (except the cluster head) estimates the cost of operating in PREMON and DEFAULT modes. Every TΔ units of time (TΔ = Δ*Ta), the node decides the mode of operation that is going to be more energy-efficient and switches to that mode, where Δ is the number of readings to be predicted. The node stays in the chosen mode for TΔ units of time. • If a node decides to use PREMON mode, it sends a prediction model of its readings to the cluster head. The node may then turn OFF its radio and turning it ON only for sending violations. It is assumed that all messages between cluster nodes and the cluster head are sent with 100% reliability. • If a node decides to operate in DEFAULT mode, it needs not sending any message to the cluster head. • The cluster head knows the mode of operation of all nodes in the cluster; it responds on behalf of the nodes operating in PREMON mode and passes queries/monitoring operation request to any node operating in DEFAULT mode. These nodes report their readings directly to the querying/monitoring entity. In buddy performance analysis, expressions for the cost of operations in the DEFAULT and PREMON modes are defined. Factors that influence these costs are identified and “feasible regions” are determined in the parameter space where using PREMON mode is more cost-effective than DEFAULT mode. The energy cost is expressed in terms of the duration where the radio was ON. This is a good approximation for two reasons: • Energy consumed by radio is an order of magnitude higher than that of computation. • The costs of transmitting and receiving data are roughly the same (Raghunathan et al. 2002; Anastasi et al. 2004). The performance analysis yielded interesting outcomes regarding costs in DEFAULT and PREMON modes, how are the feasible regions when a node is not monitored and when it is monitored, and to what extent is the energy saving in PREMON mode: • Estimating cost in DEFAULT mode. When a node is not being monitored, it only needs to listen to beacons from the cluster head every Td units of time. The cost of doing this over the time interval TD is given by:

336

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

TDNot

monitored

¼

TD  Tbeac Td

ð5:49Þ

where, Tbeac is the duration of a beacon packet. A node that is being monitored needs to also send its readings every Ta units of time; the due cost over the time interval TD is given by: TDMonitored ¼

TD TD  Tbeac þ  T ðd Þ Td Ta

ð5:50Þ

where T ðd Þ is the average time required to send a packet of size d bits, reliably to the cluster head; it is a function of channel bit error rate, BER, and packet size. Noticeably, the parameters required to compute the cost of operating in DEFAULT mode, for TD units of time, are available locally at the sensor node. • Estimating cost in PREMON mode. As in DEFAULT mode, the node listens to the beacon from the cluster head every Td units of time. Every TD units of time, the node generates a prediction model; this model is sent to the cluster head. Every Ta units of time, the node compares its reading with the one given by the prediction model. If they differ by more than c, the error threshold, the node declares its current reading as “violation” and sends it to the cluster head. Let pv represent the probability of occurrence of this event. Then, during time interval TD , on the average, the sensor node needs to transmit pv*Δ readings to the cluster-head. The total cost of operating in PREMON mode is given by:
 TB ¼ fT ðmÞ þ Tbeac þ pv  D  ðTbeac þ Tðd ÞÞg þ

  TD  D  Tbeac Td ð5:51Þ

The first part of the above equation represents the cost of sending a prediction model of size m bits and sending violations. The second part represents the cost of listening to the beacons from the cluster head. Every Ta units of time, the node can skip listening to the beacon because for this epoch the cluster head knows the reading of the sensor node. Unlike DEFAULT mode, the energy cost in PREMON mode is the same irrespective of whether the sensor node is being monitored or not. To compute the energy cost of PREMON mode, all parameters except pv are known locally at the sensor node. A sensor node keeps a running estimate of pv, by locally generating a prediction model and comparing its actual readings against the predicted ones; this is done irrespective of its mode of operation. • Feasible regions. The energy costs are analyzed in the two modes in order to identify “feasible regions” in the parameter space where PREMON mode is more energy-efficient than the DEFAULT mode. Let Ips be the ratio of energy

5.1 Data-Driven Approach Taxonomy

337

cost in PREMON mode to that in DEFAULT mode. Define ^ pv to be the value of pv for which Ips = 1. Thus, ^pv defines range of pv ¼ ½0; ^ pv Þ for which PREMON mode is more energy-efficient, i.e., the boundary of the feasible region. The outcomes when a node is not monitored and when it is monitored are highlighted: – For a monitored node (Fig. 5.27a). The feasible region in the parameter space for pv and Δ is large. It is shown that ^ pv increases with increase in Δ, the increase is sharp for low values of Δ, and it flattens out for larger values of Δ. This is expected because with the increase in Δ, the extra overhead of transmitting the prediction model gets amortized over larger lifetime of the prediction model. The flattening out of the graph for larger values of Δ has important implications, and a prediction model need not predict correctly too far into the future to be effective. This affirms for simple prediction models with good prediction accuracy in the short term. Also noted, as Δ increases, PREMON mode becomes more energy-efficient even for nodes with less predictable readings; e.g., for Δ > 4, pv = 0.6 also falls in feasible region. The feasible region is largely insensitive to the value of Td. This is due to the small cost of listening to the beacons against that for transmitting readings. – For a non-monitored node (Fig. 5.27b). The feasible region is small and is very sensitive to the value of Td. This is because the cost in PREMON mode remains the same, whereas the cost in DEFAULT mode is substantially smaller. The only operation involved is listening to beacons from the cluster head every Td units of time. Thus, in general, when a node is not being monitored, it is more energy-efficient when operating in DEFAULT mode. • Energy saving in PREMON mode. A focus is on nodes that are being monitored and on the analysis of the impact of various parameters on the cost of operating

BER=10-3 m/d=2

Range of

Range of

BER=10-3 m/d=2

Number of readings to be predicted (∆)

(a) Monitored node

Number of readings to be predicted (∆)

(b) Non-monitored node

Fig. 5.27 Range of pv ð^pv Þ versus the number of readings to be predicted (Δ) (Goel et al. 2006)

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Fig. 5.28 Energy saving versus model accuracy (Goel et al. 2006)

Ratio of energy in PREMON mode to that in DEFAULT mode (Ips)

338

BER=10-3 m/d=2 Td=Ta

Probability of occurrence of a violation

Ratio of energy in PREMON mode to that in DEFAULT mode (Ips)

Fig. 5.29 Ips versus BER and m/d (Goel et al. 2006)

pv=0 ∆=10 Td=Ta

1*10-4

2*10-4

5*10-4

1*10-3

2*10-3

5*10-3

Bit error rate (BER)

on PREMON mode. The effect of polling on the cost is eliminated by assuming that Td = Ta. – Effect of prediction model accuracy. Figure 5.28 shows Ips as a function of pv for different values of Δ. The linear increase in Ips indicates that the energy consumption in PREMON mode increases linearly with pv. As expected, the increase in Δ reduces the energy consumption in PREMON mode. As shown, in the best case, for the parameters chosen, the energy cost in PREMON mode is only 10% of that in DEFAULT mode. – Effect of prediction model size and BER. A consideration is accorded to the effect of the relative size by the ratio m=d, the ratio of the size of the packet

5.1 Data-Driven Approach Taxonomy

339

containing the prediction model to that containing the reading. Figure 5.29 illustrates the dependence of Ips on m=d and BER on the wireless link. The results were generated for perfectly predictable sensor readings (pv = 0): As the ratio increases, the energy required to deliver the prediction model increases. As BER increases, the difference in energy cost for delivering the two types of packets becomes more pronounced. Even for perfectly predictable sensor readings, the PREMON mode may cost more energy than DEFAULT mode for certain range of values of BER and m/d. As pv increases, although the shape of curves remains the same, the set of curves moves up. Thus, it is important to use smaller prediction models and to estimate BER before deciding on the mode of operation. As a last mention, most characteristics of natural environments such as chemical concentration, temperature, humidity, pollution do not change abruptly in space and time. From a sensor perspective, the environmental phenomena are quite predictable and, hence, “boring”. Buddy introduces the idea of buddies to exploit temporal correlation to save energy.

Appraisal of Data Prediction Protocols In this section, a detailed description of data prediction protocols has been provided. The sub-classification of techniques into stochastic, time-series forecasting, and algorithmic techniques identified the pros and cons of each approach and determined when to use or avoid (Anastasi et al. 2009): • The stochastic techniques are all-purpose and provide means to perform high-level operations such as aggregation (Section “Stochastic Approaches”). The main drawback of this class of techniques is their high computational cost, which may be too unbearable for current off-the-shelf sensor devices. Stochastic approaches might be more convenient for powerful sensors, like Stargate nodes in a heterogeneous WSN (Raghunathan et al. 2002). Possible enhancements on stochastic techniques would focus on deriving simplified distributed models to obtain the desired tradeoff between computation and fidelity. • Time-series forecasting techniques can provide satisfactory accuracy even when simple models, i.e., low-order AR or ARMA, are used (section “Time-Series Forecasting Approaches”). Their implementation in sensor devices is simple and lightweight. Moreover, most advanced techniques like PAQ do not require the exchange of all sensed data until a model is available (section “Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ)”). Furthermore, the ability to detect outliers and model inconsistencies is made possible. However, using a specific type of model imposes its suitability to represent the phenomenon of interest. This would require an a priori validation

340

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

phase, which may not be always feasible. The adoption of a multi-model, such as AMS, might be an interesting approach (section “Adaptive Model Selection for Time-Series Prediction in WSNs (AMS)”). The literature is rich with time-series techniques. • Algorithmic techniques are to be considered case by case, as they tend to be more application-specific (section “Algorithmic Approaches”). It is essential to focus on assessing if a specific solution is efficient for a certain class of applications in real scenarios, so that it can be taken as a reference for further study and possible improvements.

5.1.2

Energy-Efficient Data Acquisition

Some applications are sensing-constrained, in contrast with the general assumption that sensing is not energy consuming. In fact, the energy consumption of the sensing subsystem can be greater than the energy consumption of the radio and the rest of the sensor node. This can be due to many factors (Raghunathan et al. 2006): • Power hungry transducers. Some sensors intrinsically require high power resources to perform their sampling task. For example, sensing arrays such as CCDs or CMOS image sensors, multimedia sensors (Akyildiz et al. 2007) and chemical or biological sensors (Diamond et al. 2008) generally require significant power. • Power hungry A/D converters. Acoustic sensors (Simon et al. 2004) and seismic transducers (Werner-Allen et al. 2006) generally require high-rate and high-resolution A/D converters. The power consumption of the converters can account for the most significant power consumption of the sensing subsystem (Schott et al. 2005). • Active sensors. A class of sensors can get data about the sensed phenomenon by using active transducers, e.g., sonar, radar, or laser rangers. In this case, sensors have to send out a probing signal in order to acquire information about the observed quantity (Ditzel and Elferink 2006).

Energy-efficient data acquisition

Adaptive sampling

Multi-level and cooperative sampling

Model-based active sampling

Fig. 5.30 Energy-efficient data acquisition taxonomy (based on Anastasi et al. 2009)

5.1 Data-Driven Approach Taxonomy

341

• Long acquisition time. The acquisition time may be in the order of hundreds of milliseconds or even seconds; hence, the energy consumed by the sensing subsystem may be high, even if the sensor power consumption is moderate. In this case, reducing communications may be insufficient, but energy conservation schemes have to actually reduce the number of acquisitions, i.e., data samples. Moreover, energy-efficient data acquisition techniques do not target reducing the energy consumption of the sensing subsystem. By reducing the data sampled by source nodes, they decrease the number of communications as well. Worth noticing, several energy-efficient data acquisition techniques have been conceived for minimizing the radio energy consumption, under the assumption that the sensor consumption is negligible. A categorization of the approaches for energy-efficient data acquisition is presented in Raghunathan et al. (2006) and illustrated in Fig. 5.30: • Adaptive sampling techniques exploit the correlation between measured samples to reduce the amount of data to be acquired from the transducer (Sect. 5.1.2.1). For example, data of interest may change slowly with time. In this case, temporal correlations where subsequent samples do not differ considerably may be exploited to reduce the number of acquisitions. A similar approach can be applied when the investigated phenomenon does not change sharply between areas covered by neighboring nodes. In this case, energy due to sampling (and communication) can be reduced by benefiting from spatial correlations between sensed data. Clearly, both temporal and spatial correlations may be jointly exploited to further reduce the amount of data to be acquired. • The multi-level and cooperative sampling approach assumes that nodes are equipped with different types of sensors (Sect. 5.1.2.2). As each sensor is characterized by a given resolution and its associated energy consumption, this technique dynamically selects which class of sensors to activate, in order to get a tradeoff between accuracy and energy conservation. Basically, accuracy can be traded off for energy efficiency by using the low-power sensors to get coarse-grained information about the sensing field. Then, when an event is detected or a region has to be observed with greater detail, the accurate power hungry sensors can be activated. Low-power and power hungry sensors work in coordination. • Model-based active sampling takes an approach similar to data prediction (Sect. 5.1.1.3). A model of the sensed phenomenon is built upon sampled data, so that future values can be forecasted with certain accuracy (Sect. 5.1.2.3). Model-based active sampling exploits the obtained model to reduce the number of data samples and also the amount of data to be transmitted to the sink, even though this is not their main goal.

342

5.1.2.1

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Adaptive Sampling

The concept of adaptive sampling has been tackled several times; however, the main goal was optimizing the communication between sensor nodes and the basestation (Willett et al. 2004; Jain and Chang 2004; Zhou and De Roure 2007). In Jain and Chang (2004) the sampling rate is adapted to the characteristics of the data stream to allocate more bandwidth to the sensor nodes with larger activity. In Willett et al. (2004), adaptive sampling is used to support routing without proposing a specific sampling algorithm. In Zhou and De Roure (2007), adaptive sampling basically activates the appropriate number of sensor nodes to achieve a target error level, depending on spatial correlation and activity. Sections “Adaptive Sampling for Energy Conservation in WSNs for Snow Monitoring Applications” and “Event-Sensitive Autonomous Adaptive Sensing and Low-Cost Monitoring in Networked Sensing Systems (e-Sampling)” elaborate on two protocols that pertain to the adaptive sampling approach.

Adaptive Sampling for Energy Conservation in WSNs for Snow Monitoring Applications To forecast avalanches, embedded sensors that monitor snow composition in mountain slopes are proposed in Alippi et al. (2007). Targeting lower sensor energy consumption, an adaptive sampling algorithm is devised to dynamically estimate the optimal sampling frequency of the signal to be monitored. This minimizes the activity of both the sensor and the radio, hence saving energy, while maintaining an acceptable accuracy on the acquired data. Energy-saving mechanisms for the radio have been extensively studied (Anastasi et al. 2006), and this works though approaches the problem at the sensor board level. The methodology applied to snow sensors is quite general and can be tuned to any unit characterized by sensors with non-negligible energy consumption. Energy conservation at the sensor level can be achieved through adopting an adaptive duty-cycling approach that includes switching OFF the sensor board between two consecutive samples, and using the optimal sampling frequency for the physical quantity to be monitored. Specifically, an adaptive sampling algorithm is proposed for estimating the optimal sampling frequency. The basic idea is to find dynamically the minimum sampling rate compatible with the monitored signal. By reducing the sampling rate, the algorithm also reduces the amount of data to be transmitted, and consequently the energy consumed by the radio. Differently from (Willett et al. 2004; Jain and Chang 2004; Zhou and De Roure 2007), the approach presented in this section relies on the cumulative sum control chart (CUSUM) change detection test9 (Basseville and Nikiforov 1993) and focuses on reducing the energy consumption for both sensing and communication. Analysis

9

CUSUM charts have shown to be efficient in detecting small shifts in the mean of a process.

5.1 Data-Driven Approach Taxonomy

343

revealed that it is practical, and actually, it reduced up to 97% the energy consumption of the snow sensor. Monitoring the status of snow coverage allows experts to forecast avalanches. Moreover, information regarding the snow coverage helps quantifying the potential presence of water to be subsequently released and used to optimally plan hydropower generation. To gather information related to snow instabilities on mountains slopes, it is crucial to identify the composition of layers of snow at different heights from the ground. Snow is a mix of ice, water, and air, and its dielectric constant, measured at different frequencies in the range 0.1–100 kHz, varies with the percentage of water and ice in the mix. Therefore, by estimating the snow dielectric constant over time, it is possible to achieve information about the composition of different snow layers. The snow sensor is an ad hoc engineered probe fixed on a pole in the mountain to measure snow capacitance; an electronic injection board drives the probe and determines the capacity at different frequencies of excitation. For each sample cycle, the sensor provides measurements of snow capacitance at 100 Hz (low frequency) and 100 kHz (high frequency). At the same time, a second sensor provides a measurement of the ambient temperature. The three readings are then passed to a node, packed in a single message, and sent over the wireless channel. For each measurement, the injection board electronics of the snow sensor makes several procedures (calibration, electrode precharging, charge sharing) in a cyclic way to obtain a reasonably stable and reliable measure. This activity makes the sensor very energy consuming; for instance, by sampling data every 15 s, the average energy consumed is 880 mJ/sample. Such a high value is due to the fact that the sensor is not optimized for energy consumption, and it is always active (no energy management is available on the sensor). A suitable duty-cycling for the sensor is found to be around 2 s for a 150 mJ/sample energy consumption. By integrating the duty-cycling concept, the energy consumption decreased by 83%. To introduce the algorithm for estimating the optimal sampling rate of snow capacitance, some foundations are to be laid out. The Nyquist frequency is the minimum sampling frequency FN that guarantees the reconstruction of the sampled signal (Jerri 1977): FN [ 2  Fmax

ð5:52Þ

where Fmax is the maximum frequency in the power spectrum of the signal. Unfortunately, Fmax is generally not a priori available. Moreover, in a non-stationary process, the frequency spectrum of the signal, and as a consequence the maximum frequency, may change over time. The aim of the proposed adaptive sampling algorithm is thus to track the dynamics of the physical process under monitoring by adapting the sampling frequency of the sensor to the process dynamics. The proposed solution is based on the non-stationarity change detection CUSUM test. Change detection techniques are statistical tests that assess the stationarity hypothesis for a process under investigation. The traditional CUSUM test is modified to assess the non-stationarity, and hence the change, of the maximum

344

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

frequency of the power spectrum of the signal, and not the signal itself. As computations are substantial, a centralized approach is embraced; i.e., the algorithm is executed at the sink for each sensor node. The estimated sampling rates obtained by the sink are then reported to the sensor nodes. The algorithm initially estimates, through a fast Fourier transform (FFT), the
max of the signal by relying on the first W samples coming maximum frequency F
max becomes the starting reference value to be from the process. The estimated F contrasted for new estimates, possibly associated with changes in the maximum frequency of the signal spectrum. Directly from Nyquist theorem, a suitable sam
max is given by: pling frequency for F
max ; c [ 2 Fc ¼ c  F

ð5:53Þ

where c is a confidence parameter. To allow the algorithm detect frequencies higher
max , c is set to 2.1. than 2  F The algorithm defines two alternative hypotheses for the maximum frequency, in order to track the process dynamics: 

 c2
max Fup ¼ 1 þ F 4   c2
max Fdown ¼ 1  F 4

ð5:54Þ

ð5:55Þ

During the operational life, the algorithm computes the current maximum frequency Fcurr of the signal on sequences of W samples and the CUSUM test can then
max for h consecutive samples; a be applied if Fcurr is closer to Fup or Fdown than F change is detected in the maximum frequency of the signal and a new sampling frequency is defined. Specifically, the change detection test is constituted by the following detection rule:

Fig. 5.31 Frequency change recognition (Alippi et al. 2007)

5.1 Data-Driven Approach Taxonomy

345

 
max jÞ for h consecutive samples ifðFcurr  Fup \jFcurr  F or
max jÞ ifðjFcurr  Fdown j\jFcurr  F

for h consecutive samples

update the sampling frequency Fc ¼ c  Fcurr An example of frequency change detection is presented in Fig. 5.31, where it is
max and Fdown. A change is detected when the maximum possible to observe Fup, F frequency of the signal Fcurr overcomes one of the two thresholds (the horizontal dotted lines):
max þ Fup Þ=2 thup ¼ ðF
max  Fdown Þ=2 thdown ¼ ðF for h consecutive samples. The choice of h is critical to the robustness of the algorithm. With low values of h, e.g., 1 or 2, the algorithm quickly detects a variation in the maximum frequency of the signal, but it might suffer from false detections, which might cause a continuous change of the sampling frequency. On the contrary, very high values of h, e.g., 1000 or 2000, decrease the false alarm rate, but the algorithm might be less prompt in detecting changes. Setting h = 40 is suggested in this application to trade off between robustness and quickness in the frequency change detection. If available, a priori information about the process could provide the designer with a suitable parameter h. The proposed algorithm was tested using four different datasets; each dataset consisted of approximately 6000 samples acquired at fixed 15 s periods. To evaluate the performance of the algorithm, two performance metrics were chosen: • Sampling fraction, the number of samples acquired by the adaptive sampling algorithm divided by the number of samples acquired using fixed-rate oversampling, i.e., sampling every 15 s. It provides an indication of the energy saved by the adaptive sampling algorithm. • Mean relative error (MRE) gives a measure of the relative error introduced in the data sequence reconstructed at the basestation; it is defined to be:

MRE ¼

N 1X j
xi  xi j N i¼1 jxi j

ð5:56Þ

346

5 Energy Management Techniques for WSNs (2): Data-Driven Approach h=40, W=400, c=2.1

Sampling rate (%)

Fig. 5.32 Sampling rate versus message loss rate (Alippi et al. 2007)

Packet loss rate (%)

Mean relative error

h=40, W=400, c=2.1

Packet loss rate (%) Fig. 5.33 MRE for low-frequency capacitance versus message loss rate (Alippi et al. 2007)

where xi denotes the ith sample in the original data sequence,
xi is the ith data sample in the data sequence reconstructed at the basestation, and N is the total number of data in the original data sequence. When the sampling rate estimated by the algorithm is larger than the fixed oversampling rate, some samples in the original sequence are skipped and thus are not transmitted to the basestation. Also, when the message loss rate is greater than zero, some samples transmitted by the sensor node are missed by the basestation.

5.1 Data-Driven Approach Taxonomy

347

Mean relative error

h=40, W=400, c=2.1

Packet loss rate (%) Fig. 5.34 MRE for high-frequency capacitance versus message loss rate (Alippi et al. 2007)

Mean relative error

h=40, W=400, c=2.1

Packet loss rate (%) Fig. 5.35 MRE for temperature versus message loss rate (Alippi et al. 2007)

Lost samples are replaced by the previous ones (lost compensation). To compute MRE, it is assumed that
xi ¼ xi if the ith sample is correctly received by the basestation, and
xi ¼
xi1 otherwise. Messages loss was generated according to a Bernoulli distribution; to increase the accuracy of simulation, a replication method with 90% confidence level is

348

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

High frequency capacitance (ADC units)

Scenario 4, Packet loss= 40%, h=40, W=400, c=2.1

Number of samples Fig. 5.36 Original and reconstructed high-frequency capacitance (Alippi et al. 2007)

adopted (Banks 1998). The parameter settings were W = 400, c = 2.1, h = 40. Out of simulation, several findings were acquired: • Sampling rate versus the message loss rate. Figure 5.32 shows the percentage of samples acquired by the snow sensor with adaptive sampling, versus fixed oversampling (1 sample every 15 s), for increasing values of the message loss rate (messages are expressed in packets). The adaptive sampling algorithm is shown to reduce significantly the number of samples the snow sensor has to acquire. The percentage varies in the range 18–27% depending on the scenario taken into consideration. As a result, the proposed adaptive sampling algorithm can save approximately 73–82% of the energy consumed when operating at fixed oversampling. Moreover, this percentage does not depend significantly on the message loss rate of the wireless link between the sensor node and the basestation. • MRE versus the message loss rate. Figures 5.33 and 5.34 illustrate the MRE for snow capacitance at low and high frequencies, respectively, while Fig. 5.35 reports the same metric for the temperature. The MRE for low-frequency capacitance is under 3% even when the message loss rate is high. For high-frequency capacitance, MRE remains always under 4% except in Scenario 4. This is because the corresponding data sequence exhibits several spikes that are filtered out by the algorithm. Nevertheless, the data sequence reconstructed at the basestation is always very close to the original data sequence, as Fig. 5.36 displays. • The MRE for ambient temperature is high in all scenarios as Fig. 5.35 depicts. This is because temperature values range from −3 °C to 23 °C. When the absolute value is small, small deviations cause high error too. However, the

5.1 Data-Driven Approach Taxonomy

349

temperature data sequence, which is reconstructed at the basestation, is close to the original one even when the message loss rate is high. • The adaptive sampling algorithm proved its ability to reduce the number of samples by 73–82% with respect to fixed oversampling. When used in combination with a simple duty-cycling technique that switches OFF the sensor between consecutive readings, there is a saving of 95–97% of the energy consumed if the sensor was always ON. • The MRE remains at acceptable values even when the message loss is significantly high. Also, by decreasing the number of samples, the adaptive sampling algorithm reduces accordingly the amount of data to be transmitted by the sensor node, hence reducing the energy for communication by the same percentage (73–82%).

Event-Sensitive Autonomous Adaptive Sensing and Low-Cost Monitoring in Networked Sensing Systems (e-Sampling) WSNs are typically deployed to detect events, such as objects or physical changes, at a high- or low-frequency sampling rate that is usually adapted by a central unit (or a sink), thus requiring additional resource usage. However, the topic of autonomous adaptive sampling regarding the detection of events needs further probing. The event-sensitive adaptive sampling and low-cost monitoring (e-Sampling) are proposed to address this concern over two stages, leading to reduced resource usage in WSNs, such as energy and radio bandwidth (Bhuiyan et al. 2017). Due to severe resource constraints, in particular, energy and bandwidth, a lot of research has centered on extending the lifetime of WSNs from different perspectives, e.g., aggressively reducing the spatial sampling rate, rate assignment, utility-based sensing and communication (Shu et al. 2008; Alippi et al. 2010; Wang et al. 2010; Chen et al. 2012). Although these schemes attempt to reduce the energy cost of sensors, they may not be feasible in practice, due to the following serious limitations: • Sensors cannot take actions independently to adapt their rates and intervals. They all rely on a central unit (or a sink) that knows everything (signal complexities at all of the sensor positions) and periodically transmits suitable sampling rates to sensors. Such adaptation inflicts extra communication overhead on traffic-sensitive WSNs. • It is impractical to assign sampling rates or allocate bandwidth to sensors in a specific region where an interesting event occurs, especially in the case of emergency alarming applications. Also, due to unreliable wireless communications, the sink may obtain incomplete or sometimes distrustful information, leading to inaccurate judgments on rate adaptation and timely event detection.

350

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

WSNs are being increasingly suggested for many high-rate data collection applications, such as physical activity monitoring, structural health monitoring (SHM), fire event monitoring; they are required to monitor events in these applications on a long-term basis. However, sensors generate too much data for their radios in such applications, especially when transmission involves audio, imaging, and vibration (Chen et al. 2012). Frequent transmission of such raw data results in significant data loss, due to channel contention and congestion. Moreover, off-line computation at the sink is unviable in many applications. Thus, sensors should be able to reduce data autonomously before transmission. On the other hand, natural environments are often characterized by the occurrence of dynamic events. However, some events may not appear once in hours, days, months, even years, e.g., demolition, fire, snow level. Thus, sensors are required to continuously adjust their activities to dynamic systems, such as cyber-physical systems (Hackmann et al. 2014). The challenge is to provide an accurate depiction of changes in events and environmental variables. This can only be achieved if events are sensed or sampled from the environment at accurate rates. Hence, a sampling rate should be regarded as a function of both the dynamic phenomena and the application. Motivated by the limitations and requirements, e-Sampling is designed to reduce resource usage in WSNs over two stages: • In the first stage, each sensor has “short” and recurrent “bursts” of high-rate sampling; while at any other time, samples are at a much lower rate. Whenever one of the short intervals of high-rate sampling is long enough, possibly due to the presence of an event, the frequency content of signals becomes important. Depending on the frequency content, each sensor automatically switches (takes actions) its rates. Previously debated serious limitations are overcome, as e-Sampling enables reliable analysis to estimate appropriate future sampling rates and net reduction in acquired samples. • In the second stage, e-Sampling enables sensors to compute a lightweight indication of the presence of an event by analyzing, in a decentralized manner, only the important frequency content. A significant change in the content (event-sensitive or interesting data) indicates that a possible event occurred in a given monitoring application. If the event has truly occurred, the sink that receives the indications may query about detailed information from the sensors located around the event; otherwise, in the absence of event, sensors reduce data transmission to the sink (non-interesting data). The key aspect that differentiates e-Sampling from prior efforts lies in both data acquisition and decision making. e-Sampling allows an ongoing estimate of frequency content. Also, sensors adjust sampling rates independently and do not wait for the sink or neighbors’ interruption. Evaluation via simulation and experimentation on realistic datasets validated the advantages of e-Sampling in low-cost event monitoring and in expanding the

5.1 Data-Driven Approach Taxonomy

351

capacity of WSNs for high data rate applications. It was shown that e-Sampling saved up to 87% of the energy consumed by Imote2 sensors (Crossbow 2005). The proposed e-Sampling is compared further against FloodNet, an application-specific approach to adaptive sampling that provides a flood warning system (Zhou and De Roure 2007). The system includes a grid-based flood predictor to adjust the reporting rate of individual nodes. FloodNet adaptive routing (FAR) optimizes the power consumption of nodes by jointly applying adaptive sampling and energy-aware routing based on interest diffusion. The routing algorithm uses two metrics for forwarder nodes selection: specifically, priority and data importance. Priority is strictly related to energy consumption, i.e., the residual battery power and the energy needed for transmissions. Data importance is concerned with the data reporting frequency, i.e., data with high sampling rates are more important because they are associated to critical zones where variation in the phenomenon dynamics is high. The routing algorithm selects forwarder nodes among those with higher priority and lower data importance. The focus is on using nodes with higher energy resources and nodes that are less loaded with sampling tasks. Experimentation has shown that e-Sampling ensured lower energy consumption and higher network lifetime against FloodNet.

5.1.2.2

Multi-level and Cooperative Sampling

The multi-level and cooperative sampling approach promotes using a multiplicity of sensor nodes equipped with different types of sensors to acquire the event of interest; specific performance characterizes each sensor, such as accuracy and power consumption. Simple sensors are energy-efficient but have a remarkably limited resolution; meanwhile, the more sophisticated sensors can give a more detailed characterization of the sensed data at the expense of higher energy consumption. Functionally, accuracy might be traded off for energy efficiency by using the low-power sensors to get coarse-grained information about the sensing field. Then, when an event is detected or a region has to be observed with improved detail, the accurate power hungry sensors can be activated instead of keeping them always ON. Interests on multiple cameras began by focusing on improving the accuracy of computer vision algorithms through fusing data from multiple cameras and sensors. Later, due to the reduced cost, the usage of cameras and sensors expanded significantly for wide-area coverage, object tracking, and occlusion handling. Most of the proposed schemes have fused, at a central server, the data from multiple cameras and attempted to resolve the identity of targets across multiple camera views through feature matching for the sake of consistent tracking. The approach was to obtain redundant data from the environment to increase the accuracy and scope of surveillance tasks. Most of the implementations on object detection/ tracking were done using fixed wide-view cameras. Such cameras often cannot

352

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

capture high-resolution images of the targets, especially when the targets are far away from the cameras. Subsequently, with the advent of active pan-tilt-zoom cameras (PTZ), capturing high-resolution images of the targets and their activities became easier. Initially, PTZ cameras were controlled manually to focus with a high resolution on a particular region of interest (ROI). Later, PTZ cameras were coupled with the static wide-view and omnidirectional cameras in a master–slave setup. Based on the observations from static cameras, PTZ cameras were controlled automatically to focus on ROIs such as people’s faces or vehicle license plates. Recently, researchers also explored multi-agent architectural designs for surveillance systems and extended master–slave camera coordination to cooperative sensing and tracking for surveillance. In sections “Multi-Camera Coordination and Control in Surveillance Systems” and “Multiscale Approach for Structural Health Monitoring”, the multi-level and cooperative approach will be detailed from two perspectives, the surveillance systems and structural health monitoring, respectively.

Multi-Camera Coordination and Control in Surveillance Systems The use of multiple heterogeneous cameras is becoming more common in today surveillance systems. In order to perform surveillance tasks, effective coordination and control in multi-camera systems is crucial and is catching significant research attention. These systems help to enhance security and safety and are being increasingly used in public places such as banks, airports, and shopping malls as well as in prohibited locations like government and military premises. The types of cameras in these systems usually include pan-tilt-zoom (PTZ) (Micheloni et al. 2010), omnidirectional (Khoshabeh et al. 2007) and smart cameras (Rinner et al. 2008). The heterogeneity and complexity of such systems make it infeasible for human operators to manually coordinate and control the cameras to find desired targets, especially when the number of cameras and targets increases (Costello and Wang 2005). Therefore, researchers have begun focusing on automatic methods for multi-camera coordination and control (Natarajan et al. 2014). Multi-camera coordination and control (MC3) is a mechanism by which multiple heterogeneous cameras capture and analyze their videos, collect and fuse the shared knowledge about the surveillance environment from their neighboring camera nodes through communication, and finally compute and execute optimal control actions in order to perform a desired surveillance task in a collaborative manner (Natarajan et al. 2015). There are complex situations in multi-camera systems, when one or more cameras need to be alerted to perform a surveillance task seamlessly. For example, if an intruder is suspected, the cameras are required to track him/her continuously within the observed area. The system should be able to predict possible trajectories of the intruder and alert the appropriate cameras about the incoming intruder. When using PTZ cameras, they should be controlled to track the intruder at a high

5.1 Data-Driven Approach Taxonomy

353

Hey Cam2, an intruder is about to enter your OK Cam1, I will FOV. focus on it.

Cam1

Cam2

Cam2, can you leave this object to me? Cam5, you can focus on the door.

Cam3

Cam4

I am free. What shall I do now?

Cam5

294

Fig. 5.37 A scenario of coordination and control of multiple PTZ cameras (Natarajan et al. 2015)

resolution so as to gain additional information. Also, the cameras currently focusing on the intruder should transmit their data to the concerned neighboring cameras. The cameras receiving the data should be set and ready to track the potential intruder. All these operations can only be carried out if there is proper coordination and control among the camera nodes. Surveillance cameras are required to communicate and coordinate analogous to human security teams, where personnel communicate with each other during policing operation. An illustrative example of complex coordination among multiple PTZ cameras in a multi-camera system is portrayed in Fig. 5.37. To achieve this kind of active coordination and control in real time, there are many challenges that need to be addressed, making this area avid for research. The perspective of MC3 reveals several issues of importance (Natarajan et al. 2015): • The use of multiple heterogeneous cameras with complementary modalities is increasing. • There have been several attempts to fuse the raw data, features, and decisions from multiple cameras and sensors. • There have been rudimentary attempts at coordination and control of cameras in surveillance settings. • Distributed architectures for surveillance systems are gaining attention because of the growth of camera networks and the usage of autonomous smart cameras. • Developing a rigorous and principled MC3 approach for surveillance is an open problem. The fundamental architectures of MC3 that have been adopted in the literature are classified as centralized, distributed, hybrid, and multitier (Fig. 5.38):

354

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Fig. 5.38 Multi-camera systems architectures (Natarajan et al. 2015)

(a) Centralized

(c) Hybrid

(b) Distributed

(d) Multitier

• Centralized architecture. An architecture based on a manager/supervisor node that controls a set of subordinate camera nodes to perform the surveillance task (Singh et al. 2008; Sommerlade and Reid 2010; Wang et al. 2013). Usually, the manager/supervisor node is a workstation or a camera node that is responsible for fusing data from multiple subordinate camera nodes, for information assimilation and making control decisions. All information exchanged among the cameras goes through the manager/supervisor node. • Distributed architecture. This architecture consists of a set of independent camera nodes that operate autonomously by exchanging information with their neighboring camera nodes to perform a desired surveillance task (Li and Bhanu 2011). • Hybrid architecture. It combines both centralized and distributed architectures (Prati et al. 2005). In the hybrid architecture, camera nodes perform low-level functions such as object detection, tracking, and classification, and report results to the manager/supervisor node. The manager/supervisor node collects data from these cameras and fuses them to obtain useful information or to make high-level decisions. • The multitier architecture. It is a variant of the hybrid architecture; camera nodes in the top layers control cameras at the bottom layers in a hierarchical manner (Matsuyama and Ukita 2002; Bramberger et al. 2005). The internal architecture of camera nodes in centralized and distributed architectures for a typical surveillance application is shown in Fig. 5.39: • In a centralized architecture, a camera node performs object detection, tracking, and classification and passes the results to its supervisor node (Fig. 5.39a). The supervisor node then fuses the tracking results from different camera nodes and decides which targets are to be tracked by which camera nodes. • In a distributed architecture, each camera node performs both application-specific tasks and coordination tasks. For instance, in Fig. 5.39b, each camera node has

5.1 Data-Driven Approach Taxonomy Camera node

355 Manager/Supervisor node Coordination tasks

Application tasks

Image

Object detection

Data fusion Information assimilation

Object tracking

Decision making Object classification Control signals

(a) In centralized architecture Camera node Application tasks

Image

Data fusion Information assimilation

Object tracking

Decision making

Object classification

Control signals

Neighboring camera nodes

Object detection

Coordination tasks

(b) In distributed architecture

Fig. 5.39 Functionalities of camera nodes (Natarajan et al. 2015)

object detection, tracking and classification as its application-specific tasks, and data fusion, information assimilation, and decision making as its coordination tasks. All camera nodes share their own knowledge about the surveillance environment with their neighboring nodes and develop the global knowledge of the surveillance environment. With this global knowledge, each camera node tries to make local decisions that improve the overall surveillance goal. The surveillance tasks that have been accomplished by MC3 include low-level tasks, mid-level tasks, and high-level tasks (Natarajan et al. 2015): • Low-level tasks are tasks done locally within each camera. These tasks include background subtraction and foreground detection (Kim et al. 2006), blob detection and analysis, feature extraction for object detection (Tu et al. 2007) and classification (Kerhet et al. 2007), camera calibration (Pflugfelder and Bischof 2007), within a single camera view. Such tasks are performed either locally by each of the individual cameras or at the central processing server. The output of these low-level

356

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

tasks is fused to obtain more useful information for mid-level and high-level tasks. The low-level tasks are mainly restricted to individual cameras/sensors. • Mid-level tasks are performed one level above the low-level tasks. They involve object detection (Snidaro et al. 2009), tracking (Tu et al. 2007) and recognition/ classification (Ngoc et al. 2005) in multi-camera setups, camera selection and handoff (Esterle et al. 2014), view correspondence (Hu et al. 2006), etc. These tasks are accomplished mainly by fusing data from multiple cameras and/or sensors. Data from multiple cameras are fused either to increase the surveillance coverage area or to improve the accuracy of the object detection, recognition, and classification algorithms. In multi-camera tracking, occlusion is handled by combining data from multiple camera views (Santos and Morimoto 2008). • High-level tasks are accomplished based on the output from the mid-level and low-level tasks. The tasks include behavior analysis of targets (Wu et al. 2010), intrusion/anomaly detection (Arsic et al. 2008), event detection and analysis (Schreiber et al. 2010), human activity summarization (Petrushin et al. 2006). They also include computing optimal configurations of PTZ cameras and view selection from static cameras to achieve a desired surveillance goal (Ding et al. 2012). In some of the schemes, in order to achieve high-level tasks, data fusion is combined with the decision-making process; that is, after combining data from multiple cameras and sensors, decisions are made for intrusion detection, activity summarization, and people localization. The multi-camera coordination and control for surveillance are incessantly gaining significance at the research, military, and industry agendas and have become an indispensible necessity for today life.

Multiscale Approach for Structural Health Monitoring The multi-scale network concept helps to improve power efficiency, minimize packet loss and latency, and eliminate synchronization issues through the use of a decentralized analysis scheme and the activation of subnetworks only in the vicinity of suspected damage. In the meantime, the size of reference databases and requisite model orders is reduced to relieve computational burden and extend WSN lifetime (Kijewski-Correa et al. 2005). This approach contributes to the four stages of the structural health monitoring (SHM) process, namely data acquisition, data reduction, assessment, and decision making, as portrayed in Fig. 5.40 and further detailed throughout. Given the burdens associated with inspection and maintenance of civil infrastructure, the development of effective, automated damage diagnosis techniques, including the sensor technologies that support them, has become a major research focus. While developments in WSNs have demonstrated their potential to provide continuous structural response data to quantitatively assess structural health, many important issues including network lifetime and stability, damage detection reliability, and tradeoffs in model order to balance computational capabilities are

5.1 Data-Driven Approach Taxonomy

357

Stage

Approach

Benefit

Decision

Data fusion within subnetwork

Enhanced reliability

Assessment

Data driven DSF with statistical signal

More reliable. Computational demand reduced

Data reduction

Bivariate autoregressive reference database

More sensitive to damage. Easily embedded

Data acquisition

Heterogeneous. Multiscale WSN. Restricted event triggering

Event synchronized. Minimized reference pool. Low power. Scalable

Fig. 5.40 Key features of WSNs for SHM (Kijewski-Correa and Su 2009)

becoming of growing interest. Therefore, wireless embedded sensor networks are becoming a practical tool for structural health monitoring (SHM) for large, complex civil structures. In response to these growing needs, the concept of network architecture is recast to a multi-scale format with data fusion and spatially distributed heterogeneous sensing, and a restricted input network activation scheme (RINAS), along with the integration of data from a heterogeneous sensor array, to improve damage detection for low-order models. This multi-scale approach realizes several benefits: • The role of WSNs is extended from just being a means for communication, to achieving a technology that can be engineered to advance structural assessment through the interaction of subnetworks that fuse data from distributed, heterogeneous sensor arrays. Thus, superior damage detection (reduction of false positives and negatives) is attained, as well as enhanced localization capabilities through the fusion of data from different types of spatially distributed sensors. • Power efficiency is improved, packet loss and latency are minimized, and synchronization issues are eliminated, through the use of a decentralized analysis scheme and the activation of subnetworks only in the vicinity of suspected damage. • The limitations of low-cost sensors are compensated through the aggregation of sensor outputs at various scales and locations. The multi-scale WSN, built on MICAz (Crossbow 2006) and MICA2 motes (Crossbow 2002), divides the structure into a series of MESO nets (m-nets), as shown in Fig. 5.41. Within a MESO (m) net, there are wireless motes with onboard accelerometers tethered to multiple distributed strain gauges to monitor behavior at critical locations. Each accelerometer and their supporting strain gauges form a

358

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

MICRO net (l-net), where the initial diagnosis of damage is conducted. This decentralized approach not only has power conservation benefits, but also escapes the need for strict synchronization and provides resistance to latency that a centralized approach to system identification would require. The only information shared outside the MICRO (l) net is the binary damage diagnosis and estimated damage-sensitive feature (DSF), which is a customized metric for rating damage presence, severity, and location.

(a)

MACRO (M) node System controller User interface

MESO (m) node Accelerometers Local ID of system

MICRO ( ) net

MICRO ( ) node Strain gages Local stress levels

MESO (m) net

(b) MACRO (M) node Report End-user

RINAS beacon

Data fusion Damage report

DSF

Data fusion Damage report

DSF

DSF

DSF

Legend:

Wireless mote

Accelerometer

Strain gage

MICRO ( ) net

MESO (m) net

Fig. 5.41 Multiscale approach for SHM. a Multi-scale network applied to a beam (KijewskiCorrea et al. 2005). b Picturization of multiscale WSN [based on (Kijewski-Correa and Su 2009)]

5.1 Data-Driven Approach Taxonomy

Alert

359 MACRO (M) node initiates test

MESO (m) net Activated. Acquires data. Performs local ID

MESO (m) net hibernates

Alerted MESO (m) nodes hibernate Alerted MESO (m) nodes hibernate after transmission

MICRO ( ) net hibernates

MICRO ( ) net Activated in damage zone. Acquires data. Perform ID

MICRO ( ) net hibernates after transmission

Fig. 5.42 Operation of the multi-scale network [based on (Kijewski-Correa et al. 2005)]

Unlike many networks that rely on sentinels for triggering the network, this system remains dormant, conserving power until the signal to collect data is initiated by a central MACRO node (M node). Thus, this system is cycled to perform regular inspections when approaching traffic and environmental conditions meets specified criteria. Traffic classification can be accomplished through the use of camera or weigh-in-motion systems, and environmental conditions can be established through any class of meteorological station, all operating at the MACRO (M) node. Since ambient vibration monitoring is being employed, to minimize disruption, the input to the bridge is never explicitly known or controlled. However, the use of a restricted input network activation scheme (RINAS), acquiring data only when a target loading condition is satisfied, does not allow the input to be explicitly measured or controlled, but does allow the operational and environmental states to be restricted to a specific subset for which a reliable reference pool has been generated, e.g., the passage of a semi-trailer at night under a particular weather condition. This reduces the size of the reference pool, thereby easing computational burden and memory demands. The MACRO (M) node also serves as the network gateway, receiving information on structural condition and potential damage locations and severity, wirelessly from the MESO (m) nets through multihop wireless communication, and then, it interfaces with the end user to report the findings. As shown in Fig. 5.42, the detection process begins with the activation of all nodes in the MESO (m) net.

360

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

The multi-scale approach contributes to the four stages of SHM: specifically, data acquisition, data reduction, assessment, and decision making (Fig. 5.40). As elaborated throughout this section, the data acquisition aspect includes a scalable WSN acquiring acceleration and strain data, triggered using RINAS that extends network lifetime and reduces the size of the necessary undamaged reference pool. Another study focuses on the concept of multi-scale WSNs for damage detection in civil infrastructure systems (Kijewski-Correa and Su 2009). Major emphasis is given to data reduction and assessment, as a stage of SHM, through supporting a decentralized approach that operates within the hardware and power constraints of WSNs, to avoid issues associated with packet loss, synchronization and latency. The concept of a data-driven bivariate regressive adaptive index (BRAIN) for damage detection is introduced. Experimentation and simulation are used to verify two major hypotheses related to the BRAIN concept; specifically, data-driven damage metrics are becoming more robust and reliable, and the use of heterogeneous sensing enhances overall detection capability of such data-driven damage metrics. Later, a compact wireless impedance sensor node (the WID3, wireless impedance device) is suggested for use in high-frequency impedance-based SHM, sensor diagnostics and validation, and low-frequency (< *1 kHz) vibration data acquisition (Taylor et al. 2010). The compact sensor node collects relatively low-frequency acceleration measurements to estimate natural frequencies and operational deflection shapes, as well as relatively high-frequency impedance measurements to detect structural damage. The WID3 combines onboard processing using a microcontroller, data storage using flash memory, wireless communications capabilities, and a series of internal and external triggering options into a single package to realize a truly comprehensive, self-contained wireless active-sensor node for SHM applications.

5.1.2.3

Model-Based Active Sampling

Wireless sensor networks provide the flexibility of untethered sensing, but pose the challenge of achieving extended lifetime with a limited energy budget, often provided by batteries. In this respect, it is has been iterated throughout that communication causes the biggest energy drain. This is unfortunate, given that the ability to report sensed data is the main motive of resorting to WSNs in several pervasive computing applications. An approach to reduce communication without compromising data quality is to predict the trend followed by the data being sensed. This technique is referred to as model-driven data acquisition and is applicable when data are reported periodically, the common case in many pervasive computing applications. In these cases, a model of the data trend is computed locally at a node and then builds the information being reported to the data collection sink, in place of several raw samples. As long as the locally sensed data are compatible with the model prediction, no further communication is needed; only when the sensed data deviate from the model, they must be updated latter and sent to the sink.

5.1 Data-Driven Approach Taxonomy

361

Sections “Model-Driven Data Acquisition in Sensor Networks (BBQ)” and “Derivative- Based Prediction (DBP)” present BBQ and DBP, respectively, to illustrate in details the model-based active sampling approach.

Model-Driven Data Acquisition in Sensor Networks (BBQ) Declarative querying is powerful in allowing programmers to “task” an entire network of sensor nodes, rather than worrying about programming individual nodes. However, the metaphor that “WSN is a database” has proven misleading (Deshpande et al. 2004). Databases are typically treated as complete, authoritative sources of information; the job of a database query engine has traditionally been to answer a query “correctly” based upon all the available data. Applying this mindset to WSNs results in two problems: • Misrepresentations of data. In WSNs, it is impossible to gather all the relevant data. The physically observable world consists of a set of continuous phenomena in both time and space, so the set of relevant data is in principle infinite. Sensing technologies acquire samples of physical phenomena at discrete points in time and space. Yet, the data acquired by WSNs are unlikely to be random independent and identically distributed (iid) samples of physical processes, due to the non-uniform placement of sensors in space, the faulty sensors, and the high packet loss rates. So a straightforward interpretation of WSN readings as a database may not be a reliable representation of the real world. • Inefficient approximate queries. Since a WSN cannot acquire all possible data, readings are “approximate,” in the sense that they only represent the true state of the world at the discrete instants and locations where samples were acquired. However, query processing in WSNs tends to acquire as much data as possible from the environment at a given point in time, even when most of that data provide little benefit in approximate answer quality (Yao and Gehrke 2002; Madden et al. 2003). Worth noting, query execution cost, in both time and power consumption, can be orders of magnitude more than required for a reasonably reliable answer. A prototype called BBQ10 that uses a model based on time-varying multivariate Gaussians is proposed (Deshpande et al. 2004). This prototype compensates for both of these deficiencies by incorporating statistical models of real-world processes into a WSN query processing architecture. Models can help provide more robust interpretations of sensor readings; they can account for biases in spatial sampling, can help identify sensors that are providing faulty data, and can extrapolate the values of missing sensors or sensor readings at geographic locations where sensors are no longer operational. Furthermore, models provide a framework for optimizing the acquisition of sensor readings; sensors should be used to acquire

10

BBQ is short for Barbie-Q: A Tiny-Model Query System.

362

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

30 28 26 24 22 20

(a) Architecture

1

2

3

4

(b) Graph representation

Fig. 5.43 Model-based querying (Deshpande et al. 2004)

data only when the model itself is not sufficiently rich to answer the query with acceptable confidence. Underneath this architectural shift in WSN querying, a key optimization problem is defined. Given a query and a model, choose a data acquisition plan for the WSN to best refine the query answer. This optimization problem is characterized by two forms of dependencies: one in the statistical benefits of acquiring a reading and the other in the system costs associated with wireless sensor systems. With more detailing: • Any non-trivial statistical model will capture correlations among sensors; for instance, the temperatures of geographically proximate sensors are likely to be correlated. Given such a model, the benefit of a single sensor reading can be used to improve estimates of other readings; the temperature at one sensor node is likely to improve the confidence of model-driven estimates for nearby nodes. • The second form of dependency hinges on the connectivity of the WSN. If a sensor node is not within radio range of the query source, then one cannot acquire a reading from that sensor without forwarding the request/result pair through another near node. This presents not only a non-uniform cost model for acquiring readings, but one with dependencies due to multihop networking; the acquisition cost for the near node will be much lower if one has already chosen to acquire data from the far node by routing through the near. The proposed architecture is described through the example illustrated in Fig. 5.43. Users submit SQL queries to the database, which are translated into probabilistic computations over the model, as later described. The queries include

5.1 Data-Driven Approach Taxonomy

363

error tolerances and target confidence bounds that specify how much uncertainty the user is willing to accept. Such bounds will be intuitive to many scientific and technical users, as they are similar to the confidence bounds used for reporting results in most scientific fields. Based on the model, the system decides that the most efficient way to answer the query with the requested confidence is to read battery voltage from sensors 1 and 2 and temperature from sensor 4. Based on the knowledge of the WSN topology, it generates an observation plan that acquires samples in this order and sends the plan into the network, where the appropriate readings are collected. These readings are used to update the model, which can then be used to generate query answers with specified confidence intervals. Noticeably, this model elects to observe the voltage at some nodes despite the fact that the user query was over temperature. This goes for two reasons: • Temperature and voltage are highly correlated. The relationship between temperature and voltage is due to the fact that, for many types of batteries, as they heat or cool, their voltages vary significantly (by as much as 1% per °C). The voltages may also decrease as the sensor nodes consume energy from the batteries, but the timescale at which that happens is much larger than the timescale of temperature variations, and so the model can use voltage changes to infer temperature changes. • Cost differential. Depending on the specific type of temperature sensor used, it may be much cheaper to sample the voltage than to read the temperature. For example, on sensorboards of Berkeley motes (Crossbow 2002), the temperature sensor requires several orders of magnitude more energy to sample as simply reading battery voltage. One of the important properties of many probabilistic models, including the one used in BBQ, is that they can capture correlations between different attributes. How to exploit such correlations during optimization to generate efficient query plans is clarified in Eqs. (5.62)–(5.68). The proposed model-driven data acquisition tackles several issues: • Confidence intervals and correlation models. The user in Fig. 5.43 could have requested 100% confidence and no error tolerance, in which case the model would be required to interrogate every sensor. The returned result could still include some uncertainty, as the model may not have readings from particular sensors or locations at some points in time, due to sensor or communications failures, or lack of sensor instrumentation at a particular location. These confidence intervals computed from the proposed probabilistic model provide considerably more information than what traditional sensor network systems like TinyDB (Madden et al. 2005) and Cougar (Section “Cluster-Based Data Aggregation Protocols”) provide in this setting. With such systems, the user would simply get no data regarding missing times and locations. Conversely, the user could have requested very wide confidence bounds, in which case the model might have been able to answer the query without acquiring any additional data from the network. Through experimentation with BBQ on several real-world datasets, strong correlations between sensors during certain times of the

364

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

day mean that even queries with relatively tight confidence bounds can be answered with a very small number of sensor observations. In many cases, tight confidences can be provided despite significant changes in sensor readings, as known correlations between sensors make possible the prediction of such changes. • In BBQ, there is a specific probabilistic model based on time-varying multivariate Gaussians. A multivariate Gaussian is the extension of the unidimensional normal probability density function (PDF), known as the “bell curve.” Just as with its one-dimensional counterpart, a multivariate Gaussian PDF over d attributes, X1, …, Xd can be expressed as a function of two parameters: explicitly, a length-d vector of means, µ, and a d d matrix of covariances, R. Historical data are used to construct the initial representation of this PDF using TinyDB, the traditional WSN querying system. Once the initial p is constructed, queries can be answered using this model, which can be updated as new observations are obtained from the WSN, and as time passes. • Supported queries. Answering queries probabilistically based on a distribution, Gaussian for instance, is straightforward. For example, a query asks for an 2 approximation to the value of a set of attributes, with confidence at least 1 − d. The PDF can be used to compute the expected value, µi, of each attribute in the query; these will be the reported values. The PDF can be used to compute the probability that Xi is within 2 from the mean, PðXi 2 ½li  2; li þ 2Þ. If all of these probabilities meet or exceed user-specified confidence threshold, then the requested readings can be directly reported as the mean li . If the model confidence is too low, then additional readings are required before answering the query. Choosing which readings to observe at this point is an optimization problem with the goal to pick the best set of attributes to observe, i.e., minimizing the cost of observation required to bring the model confidence up to the user-specified threshold for all of the query predicates. The query and optimization engine are used in BBQ to answer a number of SQL queries, including: – Simple selection queries requesting the value of one or more sensors, or the value of all sensors in a given geographic region. – Whether or not a predicate over one or more sensor readings is true. – Grouped aggregates such as AVERAGE. However, BBQ is not designed for outlier detection; that is, it will not immediately detect when a single sensor reading is extremely far from its expected value or from the value of neighbors it has been correlated with in the past. • Networking model and observation plan format. The initial implementation of BBQ focused on static WSNs, such as those deployed for building and habitat monitoring; hence, network topologies change relatively slowly. Network

5.1 Data-Driven Approach Taxonomy

365

topology information is captured when collecting data by including, for each sensor, a vector of link quality estimates for neighboring sensor nodes. This topology information is used when constructing query plans by assuming that the nodes previously connected will still be in the near future. When executing a plan, if a particular link is found unavailable (e.g., because one of the sensors has failed), the topology model is updated accordingly. Querying the network collects new topology information, so that new links become available. This approach is found effective if the topology is relatively stable; highly dynamic topologies will need more sophisticated techniques. In BBQ, observation plans consist of a list of sensor nodes to visit and, at each of these nodes, a possibly empty list of attributes that need to be observed at that node. The possibility of visiting a node, but observing nothing is included to allow plans to observe portions of the network that are separated by multiple radio hops. • Cost model. During plan generation and optimization, it is required to compare the relative costs of executing different plans in the network. As energy is the primary concern in battery-powered WSNs (Pottie and Kaiser 2000; Intanagonwiwat et al. 2003), the goal is to pick plans of minimum energy cost. The primary contributors to energy cost are communication and data acquisition from sensors. CPU overheads beyond what is required when acquiring and sending data are small, as there is no significant processing done on the nodes. The cost model uses numbers obtained from the datasheets of sensors and radio used on MICA2 motes with the MTS400 environmental sensorboard (Crossbow 2002). Model-based querying is centered on resorting to a probabilistic model to answer queries about the attributes in the WSN. To start with, a probability density function (PDF), or prior density, pðX1 ; :::; Xn Þ assigns a probability for each joint value x1 ; . . .; xn for the attributes X1 ; :::; Xn . Specific queries are range predicates, attribute value estimates, and standard aggregates: • Range queries. Range queries ask if an attribute Xi is in the range ½ai ; bi . Typically, a WSN is queried to obtain the value of the attribute and then to test whether the query is true or false. Using a probabilistic model, the probability PðXi 2 ½ai ; bi Þ can be computed. If this probability is very high, the predicate Xi 2 ½ai ; bi  is true. Analogously, if the probability is very low, the predicate is false. Otherwise, there is no enough information to answer this query with sufficient confidence and more data are required from the WSN. The probability PðXi 2 ½ai ; bi Þ can be computed in two steps: First, marginalize, or project, the PDF pðX1 ; :::; Xn Þ to a density over only attribute Xi :

366

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Z pð x i Þ ¼

pðx1 ; . . . ; xn Þdx1 . . . dxi1 dxi þ 1 . . . dxn :

ð5:57Þ

Marginalization gives the PDF over only Xi . Then, compute PðXi 2 ½ai ; bi Þ: Zbi PðXi 2 ½ai ; bi Þ ¼

pðxi Þdxi

ð5:58Þ

ai

Range queries over multiple attributes can be answered by marginalizing the joint PDF to that set of attributes. Thus, the joint probability density pðX1 ; :::; Xn Þ provides probabilistic answers to any range query. If the user specifies a confidence level 1  d, for d 2 ½0; 1; the query can be answered if this confidence is either PðXi 2 ½ai ; bi Þ > 1 − d or PðXi 2 ½ai ; bi Þ\d. However, in some cases, the computed confidences may be low compared to the ones required by the query, and new observations might be required, that is, acquiring new sensor readings. • Value queries. A probability density function can be used to answer many other query types. For example, if the user is interested in the value of a particular attribute Xi , this query can be answered by using the posterior PDF to compute the mean
xi value of Xi , given the observations o: Z


xi ¼ xi  pðxi joÞdxi

ð5:59Þ

Confidence intervals on this estimate of the value of the attribute are possible. For a given error bound e [ 0, the confidence is simply given by PðXi 2 ½xi  e;
xi þ ejoÞ, which can be computed as in the range queries in Eq. (5.58). If this confidence is greater than the user-specified value 1 − d, then an approximately correct value might be provided for the attribute, without observing it. • AVERAGE aggregates. Average queries can be answered similarly, by defining an appropriate PDF. For an interest on the average value of a set of attributes A, such as the average temperature in a spatial region, A can be defined to be the set of sensors in thisPregion. Define a random variable Y to represent this average, where Y ¼ ð i2A Xi Þ=jAj: The PDF for Y is simply given by appropriate marginalization of the joint PDF over the attributes in A: " Z

pðY ¼ yjoÞ ¼ pðx1 ; . . .; xn joÞI

X i2A

! xi Þ=jAj

# ¼ y dx1 . . . dxn

ð5:60Þ

5.1 Data-Driven Approach Taxonomy

367

where I½: is the indicator function.11 Once pðY ¼ yjoÞ is defined, an average query can be answered by simply defining a value query for the new random variable Y as described. Probabilistic answers to more complex aggregation queries can also be computed, such as the average value of the attributes in A that have value greater than z. Thus far, the focus was on a single static probability density function over the attributes. This distribution represents spatial correlation in WSN deployment. However, many real-world systems include attributes that evolve over time; typically, temperatures have both temporal and spatial correlations. Thus, the temperature values observed earlier in time should help estimating the temperature later in time. A dynamic probabilistic model can represent such temporal correlations. In particular, for  each discrete time index t, there is a need to estimate a PDF pðX1t ; . . .; Xnt o1...t Þ that assigns a probability for each joint assignment to the attributes at time t; given o1...t , all observations made up to time t. A dynamic model describes the evolution of this system over time, illustrating how to compute   tþ1 t þ 1  1...t t t  1...t pðX1 ; . . .; Xn o Þ from pðX1 ; . . .; Xn o Þ. Thus, all measurements made up to time t can be used to improve the estimate of the PDF at time t + 1. Markovian models are effective and simple to represent such a stochastic and dynamic system; given the value of all attributes at time t, the value of the attributes at time t + 1 is independent of those for any time earlier than t. A conditional  density pðX1t þ 1 ; . . .; Xnt þ 1 X1t ; . . .; Xnt Þ called the transition model summarizes the dynamics.  Using this transition model, pðX1t þ 1 ; . . .; Xnt þ 1 o1...t Þ can be computed using a simple marginalization operation:      Z ðxt1þ 1 ; . . .; xtnþ 1 o1...t Þ ¼ pðxt1þ 1 ; . . .; xtnþ 1 xt1 ; . . .; xtn Þ  p xt1 ; . . .; xtn o1...t dxt1 . . . dxtn

ð5:61Þ   This formula assumes that the transition model p Xt þ 1 jXt is the same for all times t. As illustrated, PDFs can be conditioned on the value o of the set of observed attributes to obtain a more confident answer to a query; clearly, the choice of the observed attributes has a crucial effect on the resulting posterior density. Also, there should be a focus on selecting the attributes that are expected to increase the

11

The indicator function translates a Boolean predicate into the arithmetic value 1 if the predicate is true and 0 if false.

368

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

confidences in the answer to a particular query at minimal cost. The notions of the cost of observing a particular set of attributes, the expected improvement in the query answer from observing this set, and optimizing the choice of attributes are detailed in the following bullets: • Cost of observations. Denote a set of observations by O  f1; . . .; ng. The expected cost CðOÞ of observing attributes O is divided additively into the data acquisition cost Ca ðOÞ, representing the cost of sensing these attributes, and the expected data transmission cost Ct ðOÞ, measuring the communication cost required to download these data. The acquisition cost Ca ðOÞ is deterministically given by the sum of the energy required to observe the attributes O: Ca ðOÞ ¼

X

Ca ðiÞ

ð5:62Þ

i2O

where Ca ðiÞ is the cost of observing attribute Xi . The definition of the transmission cost Ct ðOÞ is somewhat trickier, as it depends on the particular data collection mechanism used to collect these observations from the network and on the network topology. Furthermore, if the topology is unknown or changes over time, or if the communication links between nodes are unreliable, as in most WSNs, this cost function becomes stochastic. For simplicity, the focus was on networks with known topologies, but with unreliable communication. A network graph is defined by a set of edges e, where each edge eij is associated with two link quality estimates, pij and pji , indicating the probability that a packet from i will reach j and vice versa. With the simplifying assumption that these probabilities are independent, the expected number of transmission and acknowledgment messages required to guarantee a successful transmission between i and j is 1=pij  pji . These simple values can be used to estimate the expected transmission cost. There are many possible mechanisms for traversing the network and collecting these data. The focus was on simply choosing a single path through the network that visits all sensors that observe attributes in O and returns to the basestation. Clearly, choosing the best such path is an instance of the traveling salesman problem, where the graph is given by the edges e with weights 1=pij  pji : Although this problem is NP-complete, heuristics, such as k-OPT (Lin and Kernighan 1973), known to perform well might be used. Define Ct ðOÞ to be the expected cost of this suboptimal path. The expected total cost for observing O can thus be obtained by: CðOÞ ¼ Ca ðOÞ þ Ct ðOÞ

ð5:63Þ

5.1 Data-Driven Approach Taxonomy

369

• Improvement in confidence. Observing attributes O should improve the confidence of the posterior density. That is, after observing these attributes, it is possible to answer a query with more certainty.12 For a particular value o of the observations O, the posterior density pðX1 ; . . .; Xn joÞ can be computed and the confidence estimated as described in Eqs. (5.58), (5.59), and (5.60). More specifically, for a range query Xi 2 ½ai ; bi , the benefit Ri ðoÞ of observing the specific value o can be computed as: Ri ðoÞ ¼ max½PðXi 2 ½ai ; bi joÞ; 1  PðXi 2 ½ai ; bi joÞ

ð5:64Þ

That is for a range query, Ri ðoÞ simply measures the confidence after observing o. For value and average queries, define the benefit by: Ri ðoÞ ¼ PðXi 2 ½
xi  e;
xi þ ejoÞ

ð5:65Þ

where
xi is the posterior mean of Xi given the observations o. However, the specific value o of the attributes O is not known a priori; the expected benefit Ri ðOÞ must thus be computed: Z Ri ðOÞ ¼

pðoÞ  Ri ðoÞdo

ð5:66Þ

This integral may be difficult to compute in closed form, and an estimate of Ri ðOÞ using numerical integration is required. • Optimization. The cost CðOÞ and expected benefit RðOÞ of observing attributes O were described in Eqs. (5.62)–(5.66). Different sets of observed attributes will lead to different benefit and cost levels. The user will be able to define a desired confidence level 1  d, and the set of attributes O that meet this confidence at a minimum cost must be picked up: minimizeOf1;...;ng CðOÞ; such that RðOÞ  1  d

ð5:67Þ

This optimization problem is known to be NP-hard. In the proposed approach, two algorithms were suggested to solve this optimization problem: – The first exhaustively searches over the possible subsets of possible observations, Of1; . . .; ng. The algorithm can thus find the optimal subset of attributes to observe, but at an exponential running time. – The second algorithm uses a greedy incremental heuristic. The search was initialized with an empty set of attributes, O ¼ ;. At each iteration, for each attribute Xi that is not in the set, that is i 62 O, the new cost C ðO [ iÞand expected benefit RðO [ iÞare computed; hence:

12

This is not true in all cases; for range predicates, the confidence in the answer may decrease after an observation, depending on the observed value.

370

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

if some set of attributes G reach the desired confidence, i.e., for j 2 G, RðO [ jÞ  1  d, then among the attributes in G, the one with lowest total cost C ðO [ jÞ is selected, and the search is terminated returning O [ j. else if G ¼ ; then the desired confidence is not attained, and the attribute with the highest benefit over cost ratio is added to the set of attributes:   RðO [ jÞ O ¼ O [ arg maxj62O C ðO [ jÞ

ð5:68Þ

This process is repeated until the desired confidence is reached. With the goal set to demonstrate that BBQ provides the ability to efficiently execute approximate queries with user-specifiable confidences, experimentation with TinyDB was implemented on several real-world datasets.

Derivative-Based Prediction (DBP) This proposal investigates model-driven data acquisition approaches on a real application and checks if the energy savings they enable in theory are still worthwhile once the network stack is taken into account (Raza et al. 2012). At each node, the proposed derivative-based prediction (DBP) predicts the sampled data; when these data deviate from the current model, a new model is generated and sent to the data sink. While several model-driven data acquisition approaches have been applied in real-world pervasive applications, their practical applicability remains unascertained. Moreover, evaluating the gains was only in terms of messages suppressed compared to a standard approach sending all samples. This data-centric view, however, is quite optimistic. WSN network protocols consume energy not only when transmitting and receiving data, but also in several continuous control operations, e.g., when maintaining a routing tree for data collection, or probing for ongoing communication at the MAC layer. Therefore, the true question, yet to be fully answered, is to what extent the theoretical savings shown in model-driven data acquisition are actually observable in practice when the application and network stack are combined? The experimental setup for the proposed approach was based on a WSN deployed in a road tunnel to acquire light readings (Ceriotti et al. 2011). Readings are relayed in multihop to a gateway, which feeds a programmable logic controller (PLC) that closes the control loop by setting the intensity of the lamps inside the tunnel. Such closed-loop adaptive lighting system maintains optimal light levels by considering the actual conditions inside the tunnel; this increases safety and enables considerable energy savings. WSNs are an asset in this scenario, as the nodes can be placed at arbitrary points along the tunnel, not only where power and networking cables are available. Installation and maintenance costs are thus considerably reduced, and WSNs become particularly appealing for already existing tunnels,

5.1 Data-Driven Approach Taxonomy

371

where changes to the infrastructure should be minimized. The downside to such flexibility is the reliance on an autonomous energy source; nevertheless, battery costs are minimal and the replacement process can be easily combined with regularly planned tunnel maintenance. The proposed approach helps attaining essential goals: • Investigating the benefits of model-driven data acquisition in an existing deployment that provides closed-loop adaptive lighting in an operational road tunnel. The WSN reporting periodically light samples is representative of several pervasive computing applications, e.g., smart environments, building management, home automation. • Assessing the interplay of data modeling and the underlying network protocols, by evaluating qualitatively and quantitatively the relationship between them. These goals are achieved in more than a step: • Proposing the derivative-based prediction (DBP) to locally predict the trend of data sensed by a WSN node. DBP is simple, an advantage on resource-scarce WSN platforms. • Analyzing to what extent the realized improvement is affected by the interaction with network protocols, by running the implemented application on top of popular WSN protocols such as CTP (Gnawali et al. 2009) and Box-MAC (Moss and Levis 2008). The application is fed with the same light data (replayed) from the tunnel deployment, to accurately compare the theoretical gains against the practical ones. Two settings are introduced: specifically, an operational tunnel, representative of the target application, and a 40-node indoor testbed, representative of alternate application scenarios. Figure 5.44 shows the placement of WSN nodes inside the 260-m-long, two-way, two-lane tunnel. Overall, 40 nodes are split evenly between the tunnel walls and placed at a 1.70 m height, compatible with legal regulations. Their data reports are collected by a gateway, installed 2 m from the entrance. Each node is a TelosB mote (Moteiv 2004), augmented with a sensor board equipped with four ISL29004 digital light (luminance) sensors (Intersil 2011). The light readings, collected at a sampling rate of 5 s, are locally aggregated and filtered. Every 30 s, the outcome of this aggregation is reported to the sink. The WSN nodes are not time synchronized; a node reports its light value whenever its 30 s timer expires.

Fig. 5.44 Placement of WSN nodes in the tunnel (Raza et al. 2012)

372

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Value

Time tolerance

Value tolerance

Time Last value fitting the model

Transmission of value and new model to sink

Fig. 5.45 Value and time tolerances (Raza et al. 2012)

The implemented application is an instance of a general class of WSN applications where nodes periodically take sensor measurements and report the corresponding samples to a data sink. Moreover, it is assumed that the application running at the sink allows for a small tolerance in the accuracy of the reported data. Deviations from the exact reports are acceptable, as long as their extent in terms of difference in value and time interval during which the deviation occurs are small enough. This is in contrast with the ideal requirements of the sink obtaining exact values in all data reports; the correctness of these applications is unaffected providing the reported values match closely the exact ones, and the inaccurate values occur only occasionally. Several assumptions, common to many applications, are made as illustrated in Fig. 5.45: • Let Vi be an exact measurement taken at time ti . The value tolerance is defined  rel abs by the maximum relative and absolute errors acceptable, eV ¼ r ; r . From the application perspective, reading a value Vi becomes equivalent to reading ^i in the range RV defined by the maximum error: any value V ^i 2 RV ¼ ½Vi  c; Vi þ c V

ð5:69Þ

  where c ¼ max Vi =100  rrel ; rabs . ^i 2 RV as correct. In other words, the application considers a value V     ^j ; . . .; V ^k be the set of values ^T ¼ V • Let T ¼ tj  tk  be a time interval, and V reported to the application during T. The time tolerance eT is the maximum acceptable value of T such that all the values reported in this interval are ^i 2 V ^T : ^i 62 RV , 8V incorrect, i.e., V A main performance metric is chosen for DBP, the transmission ratio TR defined to be:

5.1 Data-Driven Approach Taxonomy

TR ¼

373

Number of messages generated with DBP Number of messages generated without DBP

ð5:70Þ

Like other model-driven data acquisition techniques, DBP aims at suppressing as many data reports from the WSN nodes as possible, while ensuring that the data used by the application at the sink is within the value and time tolerances eV and eT specified as part of the requirements. The combined use of absolute and relative errors in the value tolerance has its justification. When light levels are low, e.g., at night, even small absolute variations are large in terms of percentage. With only an absolute error, these minimally perceivable changes would trigger model changes. Instead, by considering the maximum between the relative and absolute error, the proposed control algorithm is able to both adjust to the meaningful changes and avoid unnecessary communication. Getting to DBP functionalities, it is based on the observation that in the considered application, the trends of the sensed values in short- and medium-time intervals can be accurately approximated using a linear model. This idea is not new, as previous studies compute models that aim to reduce the approximation error to the data points in the recent past; however, DBP targets producing models that are consistent with the trends in the recently observed data. Interestingly, DBP does not require expert domain knowledge or lengthy training as required in BBQ (section “Model-Driven Data Acquisition in Sensor Networks (BBQ)”), but provides hard accuracy guarantees on the collected data. DBP is performed over two phases, initialization then prediction, as Fig. 5.46 illustrates: • Initialization consists of a learning phase, gathering enough data to produce the first model. The learning phase involves m data points; the first and the last l are called edge points. The model is linear and is computed as the slope d of the segment that connects the average values over the l edge points at the beginning Edge points

Value

Edge points

Time Learning phase Fig. 5.46 Derivative-based prediction (Raza et al. 2012)

Prediction phase

374

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

and end of the learning phase. This computation resembles the calculation of the derivative, hence the name derivative-based prediction. It is interesting to note that the computation of this prediction is remarkably simple and also mitigates the problem of noise and outliers. • The first DBP model generated is then sent to the sink, along with its last data point. From that point on, each node buffers a sliding window of the last m data points sampled from its sensor. Upon sampling a point, the “true” value sensed is compared to the “predicted” one as computed by DBP according to the current model, i.e., following the slope d: If the sensor reading is within a value tolerance eV at the model, no action is required, then the sink will automatically generate a new value that is an acceptable approximation of the real one. else if the readings continuously deviate from the model for more than eT time units, a new model must be recomputed. This is accomplished by using the last m data points in the buffer; the resulting model is transmitted to the sink along with the last data point. DBP performance was compared against three techniques: • Piecewise linear approximation (PLA). A popular technique that uses least-squares error linear segments to approximate a set of values (Palpanas et al. 2008). In this case, each node uses a single segment to model sensed values. • Similarity-based adaptable framework (SAF). SAF as described in Section “Time-Series Forecasting for Approximate Query Answering in SensorNetworks (PAQ)” relies on an autoregressive moving-average model of order 3 with moving-average parameter of order 0 (Tulone and Madden 2006a). In SAF, a value Vi . is predicted by a linear combination of the last three: Vi ¼ li þ a1  ðVi1  li1 Þ þ a2  ðVi2  li2 Þ þ a3  ðVi3  li3 Þ

ð5:71Þ

where a1 , a2 , a3 are constants to be estimated by the model, and li models the linear trend of data over time. Fig. 5.47 Comparison between DBP and PLA, SAF, POR (Raza et al. 2012)

DBP PLA SAF POR

0.00830 0.00817 0.00907 0.01900

(a) Average error

0.00259 0.00328 0.00312 0.00712

(b) Average TR

5.1 Data-Driven Approach Taxonomy

375

• Polynomial regression (POR) method was implemented as an additional theme of comparison. In contrast to DBP, POR allows the use of nonlinear models for prediction. Intuitively, this may yield better performance through a better fit to the data. Like PLA, POR uses the least-squares measure for selecting the most appropriate coefficients for the polynomials, which have the form P y ¼ pk¼0 ai  xi . In this study, polynomials of order p = 2, 3, 4 were evaluated. A setup of p = 2 is selected as it provides the best results for POR. Adopting eV ¼ ð5; 25Þ and eT ¼ 2 as the requirements of the tunnel application, Fig. 5.47a shows the errors in predicting the actual sensor readings when using DBP, PLA, SAF, and POR. The values shown are the average error per point, over the entire 47-day dataset and over all nodes, computed as the Euclidean distance between the real sensed value and the value predicted by the corresponding model. It is noted that PLA achieves a lower error than DBP. This is because DBP inherently permits some amount of error in the model, while PLA employs an objective function that explicitly chooses the model that minimizes the error. In terms of communication performance, DBP achieves a higher reduction in TR, as shown in Fig. 5.47b, because it better models the data trends. DBP suppresses 99% of the message reports when following the requirements of the implemented tunnel application. It is worth noting that finding the derivative of the sensed data, at the core of DBP, is significantly less complex than solving linear equations with 2 or 3 unknowns, as required by PLA, SAF, and POR. Moreover, the impact of the network stack on DBP performance was evaluated based on experimentation on the tunnel and on indoor test bed. Typically chosen, the common network stack composed of CTP, BoX-MAC, and TinyOS v2.1.1 (TinyOS 2012). How metrics such as data delivery to the application, network lifetime, and routing costs are found affected by DBP: • All of these metrics, and particularly data delivery to the application and network lifetime, are deeply affected by the operation of the MAC layer, specifically the rate at which the radio duty-cycles, which therefore becomes a key parameter in the experiments. At low sleep intervals, nodes frequently check the channel but find no activity, hence increasing idle listening costs. At large sleep intervals, the cost to transmit a packet increases. In BoX-MAC, transmission to a non-sink node takes on average half the sleep interval, due to the fact that the sender must transmit until the receiver wakes up, receives the packet, and then acknowledges its reception (Moss and Levis 2008). This long transmission interval also increases the probability of packet collisions among hidden terminals, further decreasing the delivery ratio and increasing energy consumption. The ideal sleep interval balances idle listening and active transmission costs.

376

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

• It was interestingly determined that further reductions in data traffic would have little practical impact on the system lifetime, as routing costs are dominated by topology maintenance rather than data forwarding. Further, applying alternate data modeling techniques, e.g., PLA, SAF, and POR, will not have a significant effect on system lifetime, as they cannot reduce these fixed routing costs. Therefore, improvements are more likely to come from radical changes at the routing and MAC layers, taking into account the traffic patterns of model-driven data acquisition. DBP does not require expert domain knowledge or lengthy training as required in BBQ, but provides hard accuracy guarantees on the collected data. Acquired results confirm the expectation that the gains attained in practice when considering the network stack are dramatically lower than those derived in theory by taking into account only the application messages. In tangible terms, DBP triples the WSN lifetime as compared to a standard solution with periodic reporting. Also, model-driven data acquisition can yield substantial lifetime improvements in practical settings. However, the results also suggest that, to fully exploit the energy savings obtained by model-driven data acquisition, coordination between the data and network layers is necessary. Out of the presented study on what does model-driven data acquisition really achieve in WSNs, a few conclusions are drawn: • A large fraction of energy costs arise from the continuous maintenance of the data collection tree. These costs are negligible when reporting is frequent, but become dominant with model-driven data acquisition as it greatly reduces data generation. To improve lifetime further, network design choices must be revisited, as well as addressing the extremely low data rates resulting from data modeling techniques. • Although a certain amount of loss is usually tolerable, the loss of a single data model may significantly increase the error in data used by the application. Therefore, reliable mechanisms, beyond those of most routing protocols, should be considered. • Based on experimentation, new network solutions expressly targeting model-driven data acquisition are needed to achieve significant lifetime improvements.

5.1.2.4

Appraisal of Energy-Efficient Data Acquisition

Section 5.1.2 has its focus on energy-efficient data acquisition techniques for WSNs. The techniques are categorized into adaptive sampling, muti-level and cooperative sampling, and model-based active sampling:

5.1 Data-Driven Approach Taxonomy

377

• Adaptive sampling techniques are rather general and efficient (Sect. 5.1.2.1). However, most of the proposed solutions are limited to a single characterization, i.e., adaptive in time or in space. A more energy-efficient approach would combine both time and space in a single solution, so that multiple directions of information redundancy could be exploited at the same time (Rahimi et al. 2004). Furthermore, adaptive sampling techniques are often centralized, mostly because of their rather expensive computations requirements. More work has to be done for reducing the complexity of these solutions, so that viable distributed approaches can be affordable. This trend has been addressed in the context of the FloodNet system (Zhou and De Roure 2007) as cited in Section “Event-Sensitive Autonomous Adaptive Sensing and Low-Cost Monitoring in Networked Sensing Systems (e-Sampling)”. • The multi-level and cooperative sampling approach promotes using a multiplicity of sensor nodes equipped with different types of sensors to acquire the event of interest; specific performance characterizes each sensor, such as accuracy and power consumption (Sect. 5.1.2.2). Simple sensors are energy-efficient but have a remarkably limited resolution; meanwhile, the more sophisticated sensors can give a more detailed characterization of the sensed data at the expense of higher energy consumption. Functionally, accuracy might be traded off for energy efficiency by using the low-power sensors to get coarse-grained information about the sensing field. Then, when an event is detected or a region has to be observed with improved detail, the accurate power hungry sensors can be activated instead of keeping them always ON. This approach is energy-efficient and is applicable to multi-camera coordination and control for surveillance systems and in structural health monitoring. • Model-based active sampling solutions share almost the same strengths and weaknesses of the data prediction techniques (Sect. 5.1.1.3), although they embrace prediction to save energy due to data acquisition (Sect. 5.1.2.3). These models reduce communication without compromising data quality by predicting the trend followed by the data being sensed; they are applicable when data are reported periodically, a usual case in many pervasive computing applications. A model of the data trend is computed locally at a node and then builds the information being reported to the data collection sink, in place of several raw samples. A comparative assembly of the techniques offered throughout this chapter is made available in Table 5.11.

Data-driven techniques (Sect. 5.1)

(continued)

Energy-efficient data acquisition protocols (Sect. 5.1.2) • Adaptive sampling (Sect. 5.1.2.1) Adaptive sampling algorithm for snow monitoring (The temporal analysis of sensed data is used in the algorithm. The periodically sampled snow equivalent capacity leads to deriving the actual signal. Choosing the Fmax Nyquist frequency is not trivial because it cannot be known a priori, thus leading to choosing an unnecessary high sampling frequency (oversampling), and because it may vary over time as the process may be non-stationary. As a way out, the proposed adaptive algorithm dynamically estimates the current maximum frequency Fmax , according to the trend of measured data. The algorithm relies on a modified CUSUM test to set the sampling rate. As computations are excessive, a centralized approach is chosen where the algorithm is executed at the sink, and the estimated sampling rates are then reported to the sensor nodes) e-Sampling (tackles the concerns over autonomous adaptive sampling in events detection in two stages, which leads to reduced resource usage. First, e-Sampling automatically switches between high-and low-frequency intervals to reduce the resource usage while minimizing false negative detections. Second, by analyzing the frequency content, e-Sampling presents an event identification algorithm suitable for decentralized computing in resource-constrained WSNs. In the absence of an event, “uninteresting” data are not transmitted to the sink. e-Sampling is applied to structural health monitoring, a typical application of high-frequency events) • Multi-level and cooperative sampling (Sect. 5.1.2.2) Multi-camera coordination and control in surveillance systems (The MC3 systems help enhancing security and safety; they are thus being increasingly used. Their fundamental architectures have been classified as centralized, distributed, hybrid, and multitier. The internal architecture of the camera nodes is either centralized

Data reduction protocols (Sect. 5.1.1)

• In-network processing (Sect. 5.1.1.1) – Tree-based data aggregation protocols GIT (a data-centric routing protocol that allows data aggregationusing directed diffusion) TAG (a data-centric data aggregation framework built upon SPT routing. It is specifically designed for monitoring applications and allows an adjustable sleep schedule for sensor nodes; by permitting parent nodes, let their children know about the waiting time for transmission. Also, parent nodes cache their children data to prevent from data loss. TAG performs data aggregation in two phases, distribution and collection) Directed diffusion (a reactive data-centric protocol that encompasses three phases; typically, interest dissemination, gradient setup, and path reinforcement and forwarding) EDD (integrates directed diffusion with a cluster-based architecture so that the efficiency of the local interactions during gradient setup phase increases) PEGASIS (organizes sensor nodes in a chain. Each data aggregationchain has a leader responsible for transmitting aggregated data to the basestation. In order to evenly distribute the energy expenditure in the network, sensor nodes take turns acting as the chain leader. The chain forming can be achieved either in centralized manner by the basestation or decentralized by using a greedy algorithm at each sensor node) EADAT (relies on an energy-aware distributed heuristic. The basestation is the root of the aggregation tree. By considering the energy level of sensor nodes, the data forwarding task is performed by the sensor nodes that have high energy levels) DBMAC (integrates routing and MAC protocols. The main objective is both to minimize the latency for delay-bounded applications and to increase energy efficiency by taking advantage of data aggregation mechanisms. It employs a CSMA/CA medium access scheme. DBMAC is an instance of how routing and data aggregation may influence each other)

Taxonomy

Table 5.11 Data-driven techniques classified

378 5 Energy Management Techniques for WSNs (2): Data-Driven Approach

– Cluster-based data aggregation protocols LEACH (a self-organizing and adaptive clustering protocol; it takes advantage of randomization to evenly distribute the energy expenditure among the sensor nodes. Cluster heads act as data aggregation points. The protocol consists of two phases; in the first phase, cluster heads are elected and cluster structures are formed, and then, in the second phase, cluster heads aggregate and transmit data to the basestation) Cougar (a clustering scheme that performs periodic per-hop data aggregation; it is suitable for applications where sensor nodes continuously generate correlated data. Cougar selects the cluster heads based on more than one metric and allows sensor nodes to be more than one hop away from their cluster heads. Synchronization is used to correctly aggregate data) HEED (benefits from the availability of multiple power levels at sensor nodes for cluster head selection. A combined metric composed of the node residual energy and the node proximity to its neighbors is introduced. HEED defines the average of the minimum power level required by all sensor nodes within the cluster to reach the cluster head, to estimate the communication cost in each cluster) PANEL (supports asynchronous sensor network applications where the sensor readings are fetched by the basestation after some delay. PANEL ensures load balancing and supports intra and intercluster routing, allowing sensor-to-aggregator, aggregator-to-aggregator, basestation-to-aggregator, and aggregator-to-basestation communications) − Hybrid tree/cluster-based data aggregation protocols CLUDDA (combines clustering with diffusion mechanisms. It includes query definitions inside interest messages initiated by the basestation. Using clustering mechanism, it is ensured that only cluster heads that perform intercluster communication are involved in the transmission of interest messages. As the regular sensor nodes do not transmit any data unless they are capable of servicing a request, CLUDDA conserves energy)

Data reduction protocols (Sect. 5.1.1)

Taxonomy

Table 5.11 (continued)

(continued)

or distributed. The surveillance tasks accomplished through MC3 systems include low–level, mid-level, and high-level tasks) Multi-scale approach for structural health monitoring (the multi-scale network concept helps to improve power efficiency, minimize packet loss and latency, and eliminate synchronization issues through the use of a decentralized analysis scheme and the activation of subnetworks only in the vicinity of suspected damage. In the meantime, the size of reference databases and requisite model orders is reduced to relieve computational burden and extend WSN lifetime. This approach contributes to the four stages of the structural health monitoring process (SHM), namely data acquisition, data reduction, assessment and decision making) • Model-based active sampling (Sect. 5.1.2.3) BBQ (uses a model based on time-varying multivariate Gaussians. Model-based querying is centered on resorting to a probabilistic model to answer queries about the attributes in the WSN. Specific queries are range predicates, attribute value estimates, and standard aggregates) DBP (based on the observation that in the considered application, the trends of the sensed values in short- and medium-time intervals can be accurately approximated using a linear model. As opposed to previous studies, DBP targets producing models that are consistent with the trends in the recently observed data. DBP does not require expert domain knowledge or lengthy training)

Energy-efficient data acquisition protocols (Sect. 5.1.2)

5.1 Data-Driven Approach Taxonomy 379

LCS (ensures that nodes close to each other form clusters. The clusters so formed remain static for the lifetime of the network. Within each cluster, the data from each of the nodes are routed along a shortest-path tree to a cluster head node. Data aggregation takes place at each of the intermediate nodes along the SPT. The cluster head then sends the aggregated data from its cluster to the basestation along a multihop path with no intermediate aggregation) – Multi-path-based data aggregation protocols Synopsis diffusion protocol (data aggregation is performed through a multi-path approach. The underlying topology for data dissemination is organized in concentric rings around the sink. The protocol consists of two phases: explicitly, the distribution of the queries and the data retrieval. The ring topology is formed when a node sends a query over the network. As dataflow over multiple paths, a node may receive duplicates of the same information. This may affect the aggregation result when aggregation functions are duplicate sensitive. Multipath schemes are suitable for networks with frequent packet losses, as the extra duplicates pay off in terms of robustness) – Hybrid tree/multipath-based data aggregation protocols The tributaries and deltas (combines the best features of both the treeand multi-path-based structures, both data aggregation structures may simultaneously run in different regions of the network. Hence, nodes are divided into two categories: nodes using a tree-based approach to forward packets (T nodes) and nodes using a multi-path scheme (M nodes). The protocol considers low packet loss rates and high packet loss rates. The user can set a threshold to specify the minimum percentage of nodes that should contribute to the aggregation operation) • Data compression (Sect. 5.1.1.2) − LEC (exploits the natural correlation that exists in data collected by WSNs and the principles of entropy compression. Due to the low complexity and the small amount of memory required for its execution, LEC is particularly suitable for use on commercial tiny sensor nodes. Further, it is able to compute a compressed version of each value on the fly, thus

Data reduction protocols (Sect. 5.1.1)

Taxonomy

Table 5.11 (continued) Energy-efficient data acquisition protocols (Sect. 5.1.2)

(continued)

380 5 Energy Management Techniques for WSNs (2): Data-Driven Approach

reducing storage occupation. LEC exploits a very short fixed dictionary, whose size depends on the precision of ADC) • Data prediction (Sect. 5.1.1.3) − Stochastic approaches Ken (there are a number of probabilistic models; each one is replicated at the source and at the basestation. After a training phase, a PDF referring to a set of attributes is obtained. When the model is not valid any more, the source node updates it and transmits a number of samples to the basestation, so that the corresponding instance can be updated there as well. Ken is flexible to use models tailored to a specific application and exploiting spatial or temporal correlations) − Time-series forecasting PAQ (relies on AR models built at each sensor to predict local readings. Nodes transmit these local models to a sink node, which uses them to predict sensor values without directly communicating with sensors. When needed, nodes send information about outlier readings and model updates to the sink. Such an approach can drastically reduce the amount of communication required to monitor the readings of all sensors in a network, and it also provides provably correct, user-controllable error bounds on the predicted values of each sensor) AMS (extends the time-series forecasting scheme with an adaptive multi-model selection mechanism. As an a priori knowledge of the phenomenon could be not available, it would be better to let the system itself choose the right model automatically. All nodes keep a set of models, but at a given instant only one of them is used for data prediction. At every sampling instant, all models are updated, but only the current one is used for prediction. If the error between sensed data and the current model is higher than the allowed threshold, then the current model is switched to the one satisfying the requested accuracy and minimizing the cost of the update. Then, an update procedure is performed to ensure that both source and sink nodes are synchronized to the newly selected model. To save nodes resources, poorly performing models are discarded over time by using a racing mechanism)

Data reduction protocols (Sect. 5.1.1)

Taxonomy

Table 5.11 (continued) Energy-efficient data acquisition protocols (Sect. 5.1.2)

(continued)

5.1 Data-Driven Approach Taxonomy 381

− Algorithmic approaches EEDC (each node associates an upper and a lower bound, whose difference represents the accuracy of readings, to the actual value of the sensed data. These bounds are sent to the sink, which stores them for each sensor in the network. While acquiring the data, the sensors check the samples against the current bounds. If they fall outside the expected accuracy, the nodes send a source-initiated update to the sink. On the other hand, the sink receives queries from users with an associated requested accuracy. When the requested accuracy is lower than the actual accuracy provided by the value bounds, the sink can respond using the cached range; otherwise, it may request the real value and its new approximation to be used for subsequent queries directed to the sensor. This interaction is called consumer-initiated request and update. The updates impact the power consumption of nodes; they are related to two distinct aspects: the method to select ranges and the way sensors manage their state) Buddy (idle listening, when radio is ON and the node is idle waiting for communication from neighbors, is the dominant factor in energy consumption. Thus, a sensor node should turn OFF its radio as much as possible. Buddy achieves this objective by exploiting the temporal correlation in the readings of sensor nodes. Two neighboring nodes can help reduce each other energy consumption by entering into a collaborative buddy relationship. These buddies take turns in keeping their radio ON. At any point of time, the node that has its radio ON also acts as a representative for its buddies)

Data reduction protocols (Sect. 5.1.1)

Taxonomy

Table 5.11 (continued) Energy-efficient data acquisition protocols (Sect. 5.1.2)

382 5 Energy Management Techniques for WSNs (2): Data-Driven Approach

5.2 Conclusion for Well-Managed Lifestyle

5.2

383

Conclusion for Well-Managed Lifestyle

This chapter witnessed the Syrian chemical weapons crisis that raged in April 2018, and the royal wedding of Prince Harry and Meghan Markel that flamboyantly took place in Saturday, May 19, 2018. No way to forget Russia 2018 FIFA World Cup that lasted admirably from June 14 till July 15 and witnessed surprises that started by early elimination of Germany the 2014 champion, then Argentina and Brazil, till reaching the captivating match between Belgium who climbed the third place over England the fourth. The deserved coronation of France followed a seven steps’ streak that ended by setting Croatia second after an epic game in the twenty-one finals’ memory. Then, on October 2nd, 2018, came the disappearance of Saudi journalist Jamal Khashoggi in the Saudi Consulate in Istanbul, followed by an international uproar. In Tuesday, November 6, 2018, the US midterm elections revealed the Democrats taking control of the House, while Republicans enhanced their Senate majority. In Saturday, November 17th, 2018, erupted in France the anger of the Yellow Jackets (Gilets Jaunes) against economic anguish. Back to the USA, in Friday, November 30, 2018, George H. W. Bush, the 41st president and the father of the 43rd President George W. Bush, has died at 94. Again to sports, history will say that Luka Modrić the Croatian player representing Croatia the small European country won the Ballon d’Or in the ceremony held in Paris on Monday, December 3, 2018. Also, Egypt striker Mohamed Salah has been named African Footballer of the Year for the second successive year at the Confederation of African Football 2018 awards ceremony in Dakar, Senegal, on Tuesday, January 8, 2019. From memoir to technicalities, data-driven approaches are generally focused on reducing the amount of sampled data while keeping sensing accuracy within the acceptable level. Such approaches can be classified as data reduction and energy-efficient data acquisition schemes: • The data reduction schemes address the case of unneeded samples (Sect. 5.1.1). Subcategories of this scheme involve in-network processing, data compression, and data prediction: – In-network processing techniques through data aggregation aim to combine and condense data packets of several sensor nodes to reduce data transmission (Sect. 5.1.1.1). While increasing network lifetime, data aggregation protocols may degrade important QoS metrics in WSNs, such as data accuracy, latency, fault tolerance, and security. Therefore, the design of an efficient data aggregation protocol is an inherently challenging task because the protocol designer must weight energy efficiency, data accuracy, latency, and fault tolerance, against security. In order to achieve this tradeoff, data aggregation techniques are tightly coupled with how packets are routed through the network. Hence, the architecture of the WSN plays a vital role in the performance of different data aggregation protocols. Several protocols allow routing and aggregation of data packets simultaneously. These

384

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

protocols can be categorized as tree-based, cluster-based, hybrid tree/ cluster-based, multipath-based, and hybrid tree/multipath-based: Tree-based data aggregation where a shortest-path tree (SPT), a spanning tree, rooted at the sink is first constructed (Section “Tree-Based Data Aggregation Protocols”). Subsequently, such a structure is exploited in answering queries generated by the sink through performing in-network aggregation along the aggregation tree by proceeding level by level from its leaves to its root. Thus, as two or more messages get to a given node, their aggregate can be computed exactly. Cluster-based data aggregation in which sensor nodes are subdivided into clusters (Section “Cluster-Based Data Aggregation Protocols”). In each cluster, a cluster head is elected in order to aggregate data locally and transmit the aggregation result to the basestation. Cluster heads can communicate with the sink directly via long-range radio transmission. However, this is quite inefficient for energy-constrained sensor nodes. Thus, cluster heads usually form a tree structure to transmit aggregated data by multi-hopping through other cluster heads, which results in significant energy savings. Hybrid tree/cluster-based data aggregation, where tree-based techniques are comprised in each cluster, to have the application requirements fulfilled (Section “Hybrid Tree/Cluster-Based Data Aggregation Protocols”). Multipath-based data aggregation with the main idea of having each node send the data to its, possibly, multiple neighbors by exploiting the broadcast characteristics of the wireless medium (Section “Multipath-Based Data Aggregation Protocols”). Hence, data may flow from the sources to the sinks along multiple paths and aggregation may be performed at each node. In contrast to the tree-based schemes (Section “Tree-Based Data Aggregation Protocols”), multipath approaches allow to propagate duplicates of the same information. Clearly, such schemes trade a higher robustness, as multiple copies of the same data can be sent along multiple paths, for some extra overhead due to sending duplicates. Hybrid tree/multipath-based data aggregation benefits from the advantages of tree-based and multipath schemes by adaptively tuning their data aggregation structure for optimal performance (Section “Hybrid Tree/ Multipath-Based Data Aggregation Protocols”). – Data compression techniques fall into two broad classes, lossless and lossy algorithms (Sect. 5.1.1.2). Lossless and lossy compression are terms that describe whether or not, in the compression of a file, all original data can be recovered when the file is uncompressed. Due to the limited resources available in tiny sensor nodes, to apply data compression in WSNs requires specifically designed algorithms. Two approaches have been followed in this regard, distributing the computational cost on the overall network or exploiting the statistical features of the data under monitoring. The second approach can be a valuable help in power saving only if the execution of

5.2 Conclusion for Well-Managed Lifestyle

385

compression algorithms does not require an amount of energy larger than the one saved in reducing transmission. After exploring several classic compression algorithms, it was settled that compression prior to transmission in wireless battery-powered devices might actually cause an overall increase of power consumption, if no energy awareness is introduced. On the other hand, standard compression algorithms are aimed at saving storage and not energy. Consequently, appropriate strategies have to be embraced. – Data prediction techniques are focused on building an abstraction of the sensed data, or in other words, a model for future data prediction (Sect. 5.1.1.3). The data prediction schemes can be further divided into stochastic approaches, time-series forecasting and algorithmic approaches: The stochastic approaches are built on the principle of stochastic characterization of the phenomena, as proposed in Ken (Section “Approximate Data Collection in Sensor Networks using Probabilistic Models (Ken)”). Such protocols are comprised in high-level computations such as data aggregation, with remarkably expensive computational costs. They are feasible in situations where powerful sensor nodes are available in the network, equipped with higher size batteries. In time-series forecasting, sets of historical values are obtained by periodical sampling and then used to predict a future value in the same time-series (Section “Time-Series Forecasting Approaches”). Typical such approaches are PAQ as well as SAF built upon AR and ARMA (Section “Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ)”), and AMS (Section “Adaptive Model Selection for Time-Series Prediction in WSNs (AMS)”). These schemes are simple and lightweight when implemented and provide pleasing results in terms of accuracy. In algorithmic approaches, heuristic or state transition models describing sensed phenomena are used (Section “Algorithmic Approaches”). Typical examples of such more application-specific approaches are EEDC (Section “Time-Series Forecasting for Approximate Query Answering in Sensor Networks (PAQ)”) and the buddy approach based on PREMON (Section “Buddy”). • Energy-efficient data acquisition protocols are more focused toward reducing energy consumption of the node sensing subsystem (Sect. 5.1.2). Such protocols assume that greater amount of energy is consumed by the sensing subsystem of the node than the communication subsystem. These schemes are further divided as adaptive sampling, multi-level and cooperative sampling and model-based active sampling: – In adaptive sampling, the main focus is to reduce the amount of data to be acquired from the transducer based on either spatial or temporal correlation between data (Sect. 5.1.2.1). These schemes are more general and efficient

386

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

and mostly implemented in a centralized fashion, thus requiring high computations. – In multi-level and cooperative sampling, different types of sensors are installed on nodes (Sect. 5.1.2.2). These schemes are more energy-efficient but are more application-specific. However, the cost associated with the extra transceiver can be considered as a drawback of such schemes. – Model-based active sampling approaches are similar to data prediction schemes (Sect. 5.1.2.3). The goal is to reduce the number of data samples by using computed models and hence saving energy consumed through data acquisition. Protocols such as BBQ (Section. “Model-Driven Data Acquisition in Sensor Networks (BBQ)”) and DBP (Section “Derivative-Based Prediction (DBP)”) depend on such models. Being done with categorization, it is required to recall that most monitoring applications based on WSNs rely on a synchronous philosophy where readings are carried out with a specified sampling frequency. In such a case, two main approaches can be considered to reduce the energy consumed by a sensor: explicitly, duty-cycling and adaptive sensing. Duty-cycling, as fully described in the preceding chapter, is waking up the sensorial system only for the time needed to acquire a new set of samples and powering it OFF immediately afterward. This strategy allows for optimally managing energy on condition that the dynamics of the phenomenon to be monitored are time-invariant and known in advance. Since such hypotheses only partly hold in many applications, periodic sensing is typically considered; such that the “fixed” sampling rate is computed a priori, based on partial available information about the process to be monitored and assuming that the process dynamics are stationary. As a consequence, the sampling rate becomes larger than necessary (oversampling), e.g., three to five times, hence inducing energy waste. A better approach, as detailed in Sect. 5.1.2.1, would require an adaptive sensing strategy able to dynamically adapt the sensor activity to the real dynamics of the process. It is obvious that an efficient sensing strategy that reduces the number of samples also reduces the amount of data to be processed and possibly transmitted to clusters and/or the basestation. Duty-cycling and adaptive sensing are complementary approaches that can be used in combination as shown in Fig. 5.48. To accomplish the combination between the two approaches, some considerations must be ensured (Alippi et al. 2009): • The operating system has to provide a set of primitives for powering ON/OFF the sensors to support duty-cycle mechanisms. Subsequently, the application uses such primitives to acquire data according to the adaptive sensing strategy it implements. • While designing the sensor drivers for the operating system, some aspects must be considered to grant an effective handling of the duty-cycle concern. Otherwise, a not-valid acquired data will result, and/or a larger energy dissipation than that associated with the always-ON mode (Kim et al. 2008).

5.2 Conclusion for Well-Managed Lifestyle

387

In fact, each sensor is characterized by a set of functional characteristics, e.g., wakeup latency and break-even cycle, impacting the energy management of the sensor. The wakeup latency is the time required by the sensor to generate a correct value once activated. Clearly, if the sensor reading is performed before the wakeup latency has elapsed, the acquired data are consequently not valid. The break-even cycle is defined as the rate at which the power consumption of a node with a power management policy is equal to that of not power-managed node. Such value is inversely proportional with the power consumption overhead introduced by the non-ideal ON/OFF sensor transition and represents the highest sampling rate for which applying a power management is worth. Moreover, the break-even cycle is not fixed since the energy consumed by the sensor during normal operations and in ON/OFF transitions depends on the supply voltage, which changes over time (Kim et al. 2008). • The drivers should be designed by using, at least, information about wakeup latency and break-even cycle for the sensors to provide an effective sensor-specific energy management (Kim et al. 2008). Unfortunately, most currently available operating systems for sensor nodes do not follow this philosophy and let the application programmer decide when to power ON/OFF the sensor (manual management). Future operating systems would have to adopt the automated and sensor-specific approach for both relieving the application programmer from manual handling and improving the effectiveness of the duty-cycling mechanism. The general framework of Fig. 5.48 allows the WSN designer to focus on the selection of the best adaptive sensing strategy leaving low-level duty-cycling aspects to the operating system. The use of prediction algorithms and dynamic duty-cycling is proposed in Vigorito et al. (2007). After the exhaustive categorization of the data-driven techniques all over this chapter, it is important to note that clear-cut bounds are not always apparent between the different approaches; meaning that even if differently categorized, or Fig. 5.48 General framework for sensor energy management (Alippi et al. 2009)

Adaptive sensing strategy

Duty-cycling

Application

Operating system

Sensors

388

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

described over more than a chapter, energy management techniques might overlap in the concept and basics. Moreover, different techniques may be used jointly. There are many more techniques in the literature; the most innovative is presented, leading the readership for comprehensive searching, categorization, and comparisons.

5.3

Exercises

1. Write a technical report on the data reduction protocols (Sect. 5.1.1). 2. Write a technical report on energy-efficient data acquisition protocols (Sect. 5.1.2). 3. In Sect. 5.1.1.1, it is shown that data aggregation protocols are application-dependent, and the literature is stuffed with such protocols. From the literature, categorize the data aggregation protocols based on the applications they use. 4. From Sect. 5.1.1.2, compression algorithms for WSNs are based on general compression algorithms. Discuss this argument comprehensively with a focus on the compression algorithms’ design approaches. 5. Identify and compare the stochastic approaches (Section “Stochastic Approaches”). 6. Identify and compare the time-series forecasting approaches (Section “Time-Series Forecasting Approaches”). 7. Identify and compare the algorithmic approaches (Section “Algorithmic Approaches”). 8. The adaptive sampling algorithm presented in Section “Adaptive Sampling for Energy Conservation in WSNs for Snow Monitoring Applications” relies on a modified CUSUM test to set the sampling rate. Write a technical report on the CUSUM test and its modification. 9. The multi-level and cooperative sampling approach elaborated in Sect. 5.1.2.2 can add more applications to those presented in Sections “Multi-Camera Coordination and Control in Surveillance Systems” and “Multiscale Approach for Structural Health Monitoring”. Write a technical report on such applications. 10. Differentiate between the data prediction protocols presented in Sect. 5.1.1.3 and the model-based active sampling protocols as illustrated in Sect. 5.1.2.3. 11. Dig the literature for more protocols that may fit in the energy management taxonomy elaborated in this chapter. 12. Write a technical report on the data-driven approach taxonomy (Sect. 5.1).

References

389

References Akkaya, K., and M. Younis. 2005. “A Survey on Routing Protocols for Wireless Sensor Networks.” Ad Hoc Networks 3(3): 325–349. Akkaya, K., M. Demirbas, and R. S. Aygun. 2008. “The Impact of Data Aggregation on the Performance of Wireless Sensor Networks.” Wireless Communications and Mobile Computing 8: 171–193. Akyildiz, I. F., T. Melodia, and K. R. Chowdhury. 2007. “Wireless Multimedia Sensor Networks: A Survey.” IEEE Wireless Communication (IEEE) 14(6): 32–39. Alexander, T. S. 2012. Adaptive Signal Processing: Theory and Applications. New York: Springer Science & Business Media. Alippi, C., G. Anastasi, C. Galperti, F. Mancini, and M. Roveri. 2007. “Adaptive Sampling for Energy Conservation in Wireless Sensor Networks for Snow Monitoring Applications.” In International Conference on Mobile Adhoc and Sensor Systems (MASS), 1–6. Pisa, Italy: IEEE. Alippi, C., G. Anastasi, M. Di Francesco, and M. Roveri. 2009. “Energy Management in Wireless Sensor Networks with Energy-Hungry Sensors.” Instrumentation and Measurement Magazine (IEEE) 12(2): 16–23. Alippi, C., G. Anastasi, M. Di Francesco, and M. Roveri. 2010. “An Adaptive Sampling Algorithm for Effective Energy Management in Wireless Sensor Networks With Energy-Hungry Sensors.” Transactions on Instrumentation and Measurement (IEEE) 59(2): 335–344. Al-Karaki, J. N., and A. E. Kamal. 2004. “Routing Techniques in Wireless Sensor Networks: A Survey.” Wireless Communications (IEEE) 11(6): 6–28. Al-Karaki, J. N., R. Ul-Mustafa, and A. E. Kamal. 2009. “Data Aggregation and Routing in Wireless Sensor Networks: Optimal and Heuristic Algorithms.” Computer Networks 53(7): 945–960. Anastasi, G., A. Falchi, A. Passarella, M. Conti, and E. Gregori. 2004. “Performance Measurements of Motes Sensor Networks.” In The 7th International Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM), 174–181. Venice, Italy: ACM. Anastasi, G., Di Francesco, M., M. Conti, and M. Passarella. 2006. How to Prolong the Lifetime of Wireless Sensor Networks, 26. Technical, Italian Ministry for Education and Scientific Research (MIUR), Pisa, Italy: MIUR. Anastasi, G., M. Conti, M. Di Francesco, and A. Passarella. 2009. “Energy Conservation in Wireless Sensor Networks: A Survey.” Ad Hoc Networks 7(3): 537–568. Arsic, D., E. Hristov, N. Lehment, B. Hornler, B. Schuller, and G. Rigoll. 2008. “Applying Multi Layer Homography for Multi Camera Person Tracking.” In The 2nd International Conference on Distributed Smart Cameras (ICDSC), 1–9. Stanford, CA: ACM/IEEE. AZoSensors. An Introduction to Silicon Bandgap Temperature Sensors. January 1, 2018. https:// www.azosensors.com/article.aspx?ArticleID=369. Accessed 17 Aug 2018. Banks, J. 1998. Handbook of Simulation: Principles, Methodology, Advances, Applications, and Practice. New York: Wiley. Barr, K. C., and K. Asanović. 2006. “Energy-Aware Lossless Data Compression.” Transactions on Computer Systems (TOCS) (ACM) 24(3): 250–291. Basseville, M., and I. V. Nikiforov. 1993. Detection of Abrupt Changes: Theory and Application. Prentice Hall, Inc. Bhuiyan, M. Z. A., J. Wu, G. Wang, T. Wang, and M. H. Hassan. 2007. “e-Sampling: Event-Sensitive Autonomous Adaptive Sensing and Low-Cost Monitoring in Networked Sensing Systems.” Transactions on Autonomous and Adaptive Systems (TAAS) (ACM) 12(1): Article 1. Box, G. E., G. M. Jenkins, and G. C. Reinsel. 2008. Time Series Analysis Forecasting and Contro. New York: Wiley.

390

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Bramberger, M., M. Quaritsch, T. Winkler, B. Rinner, and H. Schwabach. 2005. “Integrating Multi-Camera Tracking into a Dynamic Task Allocation System for Smart Cameras.” International Conference on Advanced Video and Signal Based Surveillance (AVSS), 474–479. Como, Italy: IEEE. Brockwell, P. J., and R. A. Davis. 2016. Introduction to Time Series and Forecasting. Springer. Buonadonna, P., D. Gay, J. M. Hellerstein, W. Hong, and S. Madden. 2005. “TASK: Sensor Network in a Box.” In The 2nd European Workshop on Wireless Sensor Networks, 133–144. Istanbul, Turkey: IEEE. Buttyán, L., and P. Schaffer. 2010. “Position-Based Aggregator Node Election in Wireless Sensor Networks.” International Journal of Distributed Sensor Networks 6(1). Capo-Chichi, E. P., H. Guyennet, and J. M. Friedt. 2009. “K-RLE: A new Data Compression Algorithm for Wireless Sensor Network.” In The 3rd International Conference on Sensor Technologies and Applications (SENSORCOMM), 502–507. Athens, Glyfada, Greece: IEEE. Castelluccia, C., A. Chan, E. Mykletun, and G. Tsudik. 2009. “Efficient and Provably Secure Aggregation of Encrypted Data in Wireless Sensor Networks.” Transactions on Sensor Networks (TOSN) (ACM) 5(3): Article 20. Ceriotti, M., et al. 2011. “Is there Light at the Ends of the Tunnel? Wireless Sensor Networks for Adaptive Lighting in Road Tunnels.” In The 10th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN), 187–198. Chicago, IL: IEEE. Chatterjea, S., and P. Havinga. 2003. “A Dynamic Data Aggregation Scheme for Wireless Sensor Networks.” The Program for Research on Integrated Systems and Circuits (ProRISC). Veldhoven, Netherlands: Technology Foundation. Chen, H., J. Li, and P. Mohapatra. 2004. “RACE: Time Series Compression with Rate Adaptivity and Error Bound for Sensor Networks.” In International Conference on Mobile Ad-hoc and Sensor Systems (MASS), 124–133 Fort Lauderdale, FL: IEEE. Chen, S., and Z. Zhang. 2006. “Localized Algorithm for Aggregate Fairness in Wireless Sensor Networks.” In The 12th Annual International Conference on Mobile Computing and Networking (MobiCom), 274–285. Los Angeles, CA: ACM. Chen, Z., G. Barrenetxea, and M. Vetterli. 2012. “Share Risk and Energy: Sampling and Communication Strategies for Multi-Camera Wireless Monitoring Networks.” In The 31st International Conference on Computer Communications (INFOCOM), 1862–1870. Orlando, FL: IEEE. Cheng, R., D. V. Kalashnikov, and S. Prabhakar. 2003. “Evaluating Probabilistic Queries over Imprecise Data.” In The International Conference on Management of Data (SIGMOD), 551– 562. San Diego, CA: ACM SIGMOD. Chu, D., A. Deshpande, J. M. Hellerstein, and W. Hong. 2006. “Approximate Data Collection in Sensor Networks using Probabilistic Models.” In: The 22nd International Conference on Data Engineering (ICDE), 48–59 Atlanta, GA: IEEE. Cohen, R., and B. Kapchits. 2009. “An Optimal Wake-up Scheduling Algorithm for Minimizing Energy Consumption while Limiting Maximum Delay in a Mesh Sensor Network.” Transactions on Networking (TON) (IEEE/ACM) 17(2): 570–581. Considine, J., F. Li, G. Kollios, and J. Byers. 2004. “Approximate Aggregation Techniques for Sensor Databases.” In The 20th International Conference on Data Engineering, 449–460. Boston, MA: IEEE. Costello, C. J., and I.-J. Wang. 2005. “Surveillance Camera Coordination through Distributed Scheduling (CDC).” In The 44th IEEE Conference on Decision and Control, 1485–1490. Seville, Spain: IEEE. Cowell, R.G., P. Dawid, S. L. Lauritzen, and D. J. Spiegelhalter. 2006. Probabilistic Networks and Expert Systems: Exact Computational Methods for Bayesian Networks. New York: Springer-Verlag. Croce, S., F. Marcelloni, and M. Vecchio. 2008. “Reducing Power Consumption in Wireless Sensor Networks Using a Novel Approach to Data Aggregation.” The Computer Journal 51 (2): 227–239.

References

391

Crossbow. Imote2: High-performance Wireless Sensor Network Node. Crossbow. January 1, 2005. http://web.univ-pau.fr/*cpham/ENSEIGNEMENT/PAU-UPPA/RESA-M2/DOC/Imote2_ Datasheet.pdf. Accessed 19 March 19 2014. Crossbow. MICA2. January 1, 2002. http://www.eol.ucar.edu/isf/facilities/isa/internal/CrossBow/ DataSheets/mica2.pdf. Accessed 3 Feb 2014. Crossbow. MICAz. January 1, 2006. http://www.openautomation.net/uploadsproductos/micaz_ datasheet.pdf. Accessed 5 Feb 2014. Deshpande, A., C. Guestrin, and S. R. Madden. 2005. “Using Probabilistic Models for Data Management in Acquisitional Environments.” In The 2nd Biennial Conference on Innovative Data Systems Research (CIDR). Asilomar, CA: VLDB Endowment. Deshpande, A., C. Guestrin, S. R. Madden, J. M. Hellerstein, and W. Hong. 2004. “Model-Driven Data Acquisition in Sensor Networks.” In The 30th International Conference on Very Large Data Bases (VLDB), 588–599. Toronto, Canada: VLDB Endowment. Di Bacco, G., T. Melodia, and F. Cuomo. 2004. “A MAC Protocol for Delay-Bounded Applications in Wireless Sensor Networks.” In The 3rd Annual Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net), 208–220. Bodrum, Turkey: Ericsson/Turkcell. Diamond, D., S. Coyle, S. Scarmagnani, and J. Hayes. 2008. “Wireless Sensor Networks and Chemo-/Biosensing.” Chemical Reviews 108(2): 652–679. Ding, C., A. A. Morye, J. A. Farrell, and A. K. Roy-Chowdhury. 2012. “Opportunistic Sensing in a Distributed PTZ Camera Network.” In The 6th International Conference on Distributed Smart Cameras (ICDSC), 1–6. Hong Kong, China: IEEE. Ding, M., X. Cheng, and G. Xue. 2003. “Aggregation Tree Construction in Sensor Networks.” In The 58th Vehicular Technology Conference (VTC), 2168–2172. Orlando, FL: IEEE. Ditzel, M., and F. H. Elferink. 2006. “Low-power Radar for Wireless Sensor Networks.” In European Radar Conference, 139–141. Manchester, UK: IEEE. Elias, P. 1975. “Universal Codeword Sets and Representations of the Integers.” Transactions on Information Theory (IEEE) 21(2): 194–203. Erramilli, V., I. Matta, and A. Bestavros. 2004. On the Interaction between Data Aggregation and Topology Control in Wireless Sensor Networks. Technical Report, Computer Science Department, Boston University, Boston, USA: Boston University Computer—Science Department. Esterle, L., P. R. Lewis, X. Yao, and B. Rinner. 2014. “Socio-Economic Vision Graph Generation and Handover in Distributed Smart Camera Networks.” Transactions on Sensor Networks (TOSN) (ACM) 10(2): Article 20. Fan, K.-W., S. Liu, and P. Sinha. 2007. “Structure-Free Data Aggregation in Sensor Networks.” Transactions on Mobile Computing (IEEE) 6(8): 929–942. Fasolo, E., M. Rossi, J. Widmer, and M. Zorzi. 2007. “In-Network Aggregation Techniques for Wireless Sensor Networks: A Survey.” Wireless Communication (IEEE) 14(2): 70–87. Ganesan, D., D. Estrin, and J. Heidemann. 2003. “Dimensions: Why Do We Need a New Data Handling Architecture for Sensor Networks?” Computer Communication Review 33(1): 143– 148. Girao, J., D. Westhoff, E. Mykletun, and T. Arakic. 2007. “TinyPEDS: Tiny Persistent Encrypted Data Storage in Asynchronous Wireless Sensor Networks.” Ad Hoc Networks 5(7): 1073– 1089. Girod, B., A.M. Aaron, S. Rane, and D. Rebollo-Monedero. 2005. “Distributed Video Coding.” Proceedings of the IEEE (IEEE) 93: 71–83. Gnawali, O., R. Fonseca, K. Jamieson, D. Moss, and P. Levis. 2009. “Collection Tree Protocol.” 7th ACM Conference on Embedded Networked Sensor Systems (SenSys), 1–14. Berkeley, CA: ACM. Goel, S., A. Passarella, and T. Imielinski. 2006. “Using Buddies to Live Longer in a Boring World (Sensor Network Protocol).” In The 4th Annual IEEE International Conference on Pervasive Computing and Communications Workshops (PERCOMW), 342–346. Pisa, Italy: IEEE.

392

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Goel, S., and T. Imielinski. 2001. “Prediction-Based Monitoring in Sensor Networks: Taking Lessons from MPEG.” Computer Communication Review—Special Issue on Wireless Extensions to the Internet 31(5): 82–98. Golomb, S. W. 1996. “Run-Length Encodings.” Transactions on Infromation Theory (IEEE) 399– 401. Golub, G.H., and C.F. Van Loan. 2012. Matrix Computations. Baltimore, MD: The John Hopkins University Press. Guestrin, C., P. Bodik, R. Thibaux, M. Paski, and S. Madden. 2004. “Distributed Regression: An Efficient Framework for Modeling Sensor Network Data.” In The 3rd International Symposium on Information Processing in Sensor Networks (IPSN), 1–10. Berkeley, CA: ACM. Gupta, H., V. Navda, S. Das, and V. Chowdhary. 2008. “Efficient Gathering of Correlated Data in Sensor Networks.” Transactions on Sensor Networks (TOSN) (ACM) 4(1): Article 4. Hackmann, G., W. Guo, Z. Sun, C. Lu, and S. Dyke. 2014. “Cyber-Physical Codesign of Distributed Structural Health Monitoring with Wireless Sensor Networks.” Transactions on Parallel and Distributed Systems (IEEE) 25(1): 63–72. Hamilton, J. 1994. Time Series Analysis. Princeton, NJ: Princeton University Press. Han, Q., S. Mehrotra, and N. Venkatasubramanian. 2004. “Energy Efficient Data Collection in Distributed Sensor Environments.” In The 24th International Conference on Distributed Computing Systems (ICDCS), 590–597. Tokyo, Japan: IEEE. Harris, A. F., R. Kravets, and I. Gupta. 2007. “Building Trees based on Aggregation Efficiency in Sensor Networks.” Ad Hoc Networks 5(8): 1317–1328. Heinzelman, W .B., A .P. Chandrakasan, and H. Balakrishnan. 2002. “An Application-Specific Protocol Architecture for Wireless Microsensor Networks .” Transactions on Wireless Communications (IEEE) 1(4): 660–669. Hellerstein, J. M., and W. Wang. 2004. “Optimization of In-network Data Reduction.” In The 1st International Workshop on Data Management for Sensor Networks (DMSN), 40–47. Toronto, Canada: ACM. Hill, J., R. Szewczyk, Al. Woo, S. Hollar, D. Culler, and K. Pister. 2000. “System Architecture Directions for Networked Sensors.” In The 9th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS IX), 93–104. Cambridge, MA: ACM. Hoeffding, W. 1963. “Probability Inequalities for Sums of Bounded Random Variables.” Journal of the American Statistical Association 58(301): 13–30. Hu, L., and D. Evans. 2003. “Secure Aggregation for Wireless Networks.” Symposium on Applications and the Internet Workshops (SAINT), 384–391. Orlando, FL: IEEE. Hu, W., M. Hu, X. Zhou, T. Tan, J. Lou, and S. Maybank. 2006. “Principal Axis-Based Correspondence between Multiple Cameras for People Tracking.” Transactions on Pattern Analysis and Machine Intelligence (IEEE) 28(4): 663–671. Intanagonwiwat, C., D. Estrin, R. Govindan, and J. Heidemann. 2002. “Impact of Network Density on Data Aggregation in Wireless Sensor Networks.” In The 22nd International Conference on Distributed Computing Systems, 457–458. Vienna, Austria: IEEE. Intanagonwiwat, C., R. Govindan, and D. Estrin. 2000. “Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks.” In: The 6th Annual International Conference on Mobile Computing and Networking (MobiCom), 56–67. Boston, MA: ACM SIGMOBILE. Intanagonwiwat, C., R. Govindan, D. Estrin, J. Heidemann, and F. Silva. 2003. “Directed Diffusion for Wireless Sensor Networking.” Transactions on Networking (TON) (IEEE/ACM) 11(1): 2–16. Intersil. 2011. “ISL29004: Light-to-Digital Output Sensor with Address Selection, High Sensitivity, Gain Selection, Interrupt Function and I2C Interface.” Mouser Electronics. November 17, 2011. http://www.mouser.com/ds/2/465/fn6221–74182.pdf. Accessed 10 Jan 2019.

References

393

Jain, A., and E. Y. Chang. 2004. “Adaptive Sampling for Sensor Networks.” In The 1st International Workshop on Data Management for Sensor Networks (DMSN), 10–16. Toronto, Canada: ACM. Jain, A., E. Y. Chang, and Y.-F. Wang. 2004. “Adaptive Stream Resource Management using Kalman Filters.” In International Conference on Management of Data (SIGMOD), 11–22. Paris, France: ACM SIGMOD. Jerri, A. J. 1977. “The Shannon Sampling Theorem-Its Various Extensions and Applications: A Tutorial Review.” Proceedings of the IEEE (IEEE) 65(11): 1565–1596. Jesus, P., C. Baquero, and P. S. Almeida. 2015. “A Survey of Distributed Data Aggregation Algorithms.” Communications Surveys & Tutorials (IEEE) 17(1): 381–404. Kanagal, B., and A. Deshpande. 2008. “Online Filtering, Smoothing and Probabilistic Modeling of Streaming data.” In The 24th International Conference on Data Engineering (ICDE), 1160– 1169. Cancun, Mexico: IEEE. Kerhet, A., F. Leonardi, A. Boni, P. Lombardo, M. Magno, and L. Benini. 2007. “Distributed Video Surveillance using Hardware-Friendly Sparse Large Margin Classifiers.” In The 4th Conference on Advanced Video and Signal Based Surveillance, 87–92. London, UK: IEEE. Khoshabeh, R., T. Gandhi, and M. M. Trivedi. 2007. “Multi-Camera Based Traffic Flow Characterization & Classification.” Intelligent Transportation Systems Conference (ITS), 259– 264. Seattle, WA: IEEE. Kijewski-Correa, T., and S. Su. 2009. “BRAIN: A Bivariate Data-Driven Approach to Damage Detection in Multi-Scale Wireless Sensor Networks.” Smart Structures and Systems (Techno-Press) 5(4): 415–426. Kijewski-Correa, T., M. Haenggi, and P. Antsaklis. 2005. “Multi-Scale Wireless Sensor Networks for Structural Health Monitoring.” In The 2nd International Conference on Structural Health Monitoring of Intelligent Infrastructure (ISHM-II). Shenzhen, China: International Society for Structural Health Monitoring of Intelligent Infrastructure (ISHMII). Kim, N., I.-J. Kim, and H.-G. Kim. 2006. “Video Surveillance using Dynamic Configuration of Mutiple Active Cameras.” In The 13th International Conference on Image Processing (ICIP), 1761–1764. Atlanta, GA: IEEE. Kim, N., S. Choi, and H. Cha.2008. “Automated Sensor-Specific Power Management for Wireless Sensor Networks.” In The 5th International Conference on Mobile Ad Hoc and Sensor Systems (MASS), 305–314. Atlanta, GA: IEEE. Kotidis, Y. 2005. “Snapshot Queries: Towards Data-Centric Sensor Networks.” In The 21st International Conference on Data Engineering (ICDE), 131–142. Tokoyo, Japan: IEEE. Krishnamachari, L., D. Estrin, and S. Wicker. 2002. “The Impact of Data Aggregation in Wireless Sensor Networks.” In The 22nd International Conference on Distributed Computing Systems Workshops (ICDCS), 575–578. Vienna, Austria: IEEE. Kumar, D., T. C. Aseri, and R. B. Patel. 2009. “EEHC: Energy Efficient Heterogeneous Clustered Scheme for Wireless Sensor Networks.” Computer Communications 32(4): 662–667 . Lazaridis, I., and S. Mehrotra. 2003. “Capturing Sensor-Generated Time Series with Quality Guarantees.” In The 19th International Conference on Data Engineering (ICDE), 429–440. Bangalore, India: IEEE. LCAV. Sensorscope: Downloads. School of Computer and Communication Sciences (IC) at EPFL. May 22, 2017. https://lcav.epfl.ch/page-145180-en.html. Accessed 17 Aug 2018. Le Borgne, Y.-A., S. Santini, and G. Bontempi. 2007. “Adaptive Model Selection for Time Series Prediction in Wireless Sensor Networks.” Signal Processing 87(12): 3010–3020. Lee, M., and V. W. S. Wong. 2005. “An Energy-Aware Spanning Tree Algorithm for Data Aggregation in Wireless Sensor Networks.” In The Pacific Rim Conference on Communications, Computers and Signal Processing (PACRIM), 300–303. Victoria, B.C., Canada: IEEE. Li, Y., and B. Bhanu. 2011. “Utility-Based Camera Assignment in a Video Network: A Game Theoretic Framework.” Sensors Journal (IEEE) 11(3): 676–687. Liang, Y., and Y. Li. 2014. “An Efficient and Robust Data Compression Algorithm in Wireless Sensor Networks.” Communications Letters (IEEE) 18(3): 439–442.

394

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Lin, C.-Y., E. T. Chu, L.-W. Ku, and J. W. Liu. 2014. “Active Disaster Response System for a Smart Building.” Sensors 14(9): 17451–17470. Lin, S., and B.W. Kernighan. 1973. “An Effective Heuristic Algorithm for the Traveling-Salesman Problem.” Operations Research 21(2): 498–516. Lindsey, S., C. Raghavendra, and K. M. Sivalingam. 2002. “Data Gathering Algorithms in Sensor Networks using Energy Metrics.” Transactions on Parallel and Distributed Systems (IEEE) 13 (9): 924–935. Luo, H., J. Luo, W. Liu, and S. K. Das. 2006. “Adaptive Data Fusion for Energy Efficient Routing in Wireless Sensor Networks.” Transactions on Computers (IEEE) 55(10): 1286–1299. Lynch, J. P. 2007. “An Overview of Wireless Structural Health Monitoring for Civil Structures.” Philosophical Transactions A. Madden, S. Intel Lab Data. Intel Berkeley Research lab. June 2, 2004. http://db.csail.mit.edu/ labdata/labdata.html. Accessed 17 April 2018. Madden, S., M. J. Franklin, J. M. Hellerstein, and W. Hong. 2002. “Tag: A Tiny Aggregation Service for Ad Hoc Sensor Networks.” Operating Systems Review 131–46. Madden, S. 2003. “The Design of an Acquisitional Query Processor for Sensor Networks.” In The International Conference on Management of Data (SIGMOD), 491–502. San Diego, CA: ACM SIGMOD. Madden, S. R., M.J. Franklin, J. M. Hellerstein, and W. Hong. 2005. “TinyDB: An Acquisitional Query Processing System for Sensor Networks.” Transactions on Database Systems (TODS) (ACM) 30(1): 122–173 . Mainwaring, A., D. Culler, J. Polastre, R. Szewczyk, and J. Anderson. 2002. “Wireless Sensor Networks for Habitat Monitoring.” In The 1st ACM International Workshop on Wireless Sensor Networks and Applications (WSNA), 89–97. Atlanta, GA: ACM. Makridakis, S. G., S. C. Wheelwright, and R. J. Hyndman. 1997. Forecasting: Methods and Applications. New York: Wiley. Manjhi, A., S. Nath, and P. B. Gibbons. 2005. “Tributaries and Deltas: Efficient and Robust Aggregation in Sensor Network Streams.” In The International Conference on Management of Data, 287–298. Baltimore, MD: ACM SIGMOD. Marcelloni, F., and V. Vecchio. 2009. “An Efficient Lossless Compression Algorithm for Tiny Nodes of Monitoring Wireless Sensor Networks.” The Computer Journal 52(8): 969–987. Maron, O., and A. W. Moore. 1997. “The Racing Algorithm: Model Selection for Lazy Learners.” In Lazy Learning, ed. D. W. Aha, 193–225. Dordrecht, South Holland: Springer. Matsuyama, T., and N. Ukita. 2002. “Real-Time Multitarget Tracking by a Cooperative Distributed Vision System.” Proceedings of the IEEE (IEEE) 90(7): 1136–1150. Meriam-Webster. 2018a. Isomorphic. Meriam-Webster. January 1, 2018. https://www.merriamwebster.com/dictionary/isomorphic. Accessed 20 March 2018. Merriam-Webster. 2018b. Buddy. Merriam-Webster, Inc. June 9, 2018. https://www.merriamwebster.com/dictionary/buddy. Accessed 11 June 2018. Merriam-Webster. 2018c. Entropy. Merriam-Webster, Inc. January 1, 2018. https://www.merriamwebster.com/dictionary/entropy. Accessed 15 Aug 2018. Merriam-Webster. 2018d. Ken. Merriam-Webster. January 1, 2018. https://www.merriam-webster. com/dictionary/ken. Accessed 20 March 2018. Michalski, R. S., J. G. Carbonell, and T. M. 2013. Mitchell. Machine Learning: An Artificial Intelligence Approach. Berlin, Heidelberg: Springer-Verlag. Micheloni, C., B. Rinner, and G. L. Foresti. 2010. “Video Analysis in Pan-Tilt-Zoom Camera Networks.” Signal Processing Magazine (IEEE) 27(5): 78–90. MIDRA. IP GoodFood. MIDRA. January 1, 2006. http://www.midra.dinfo.unifi.it/CMpro-v-p-80. html. Accessed 31 May 31 2018. Moss, D., and P. Levis. 2008. “BoX-MACs: Exploiting Physical and Link Layer Boundaries in Low-Power Networking.” Technical Report: SING-08–00, Computer Systems Laboratory, Stanford University.

References

395

Moteiv. Telos: Ultra low power IEEE 802.15.4 compliant wireless sensor module. Moteiv. May 12, 2004. http://www2.ece.ohio-state.edu/*bibyk/ee582/telosMote.pdf. Accessed 27 March 2014. Natarajan, P., P. K. Atrey, and M. Kankanhalli. 2015. “Multi-Camera Coordination and Control in Surveillance Systems: A Survey.” Transactions on Multimedia Computing, Communications, and Applications (TOMM) (ACM) 11(4): Article 57. Natarajan, P., T. N. Hoang, Y. Wong, K. H. Low, and M. Kankanhalli. 2014. “Scalable Decision-Theoretic Coordination and Control for Real-time Active Multi-Camera Surveillance.” In The 8th International Conference on Distributed Smart Cameras (ICDSC). Venezia Mestre, Italy: ACM/IEEE. Article 38. Nath, S., P. B. Gibbons, S. Seshan, and Z. Anderson. 2008. “Synopsis Diffusion for Robust Aggregation in Sensor Networks.” Transactions on Sensor Networks (TOSN) (ACM) 4(2): Article 7. National Data Buoy Center. Historical NDBC Data. National Data Buoy Center. February 14, 2018. http://www.ndbc.noaa.gov/historical_data.shtml. Accessed 2 June 2018. National Geographic Society. 2018a. Delta. National Geographic Society. January 1, 2018. https:// www.nationalgeographic.org/encyclopedia/delta/. Accessed 11 Sept 2018. National Geographic Society. 2018b. Tributary. National Geographic Society. January 1, 2018. https://www.nationalgeographic.org/encyclopedia/tributary/. Accessed 11 Sept 2018. Ngoc, H.-B. B., F. Bremond, M. Thonnat, and J.-C. Faure. 2005. “Shape Recognition based on a Video and Multi-Sensor System.” In The 2nd International Conference on Advanced Video and Signal Based Surveillance (AVSS), 230–235. Como, Italy: IEEE. Olston, C., and J. Widom. 2002. “Best-Effort Cache Synchronization with Source Cooperation.” In The International conference on Management of data, 73-84. Madison, WI: ACM SIGMOD. Olston, C., B. T. Loo, and J. Widom. 2001. “Adaptive Precision Setting for Cached Approximate Values.” In The International Conference on Management of Data (SIGMOD), 355-366. Santa Barbara, CA: ACM SIGMOD. Ozdemir, S., and Y. Xiao. 2009. “Secure Data Aggregation in Wireless Sensor Networks: A Comprehensive Overview.” Computer Networks 53(12): 2022–2037. Palpanas, T., M. Vlachos, E. Keogh, and D. Gunopulos. 2008. “Streaming Time Series Summarization Using User-Defined Amnesic Functions.” Transactions on Knowledge and Data Engineering (IEEE) 20(7): 992–1006. Pattem, S., B. Krishnamachari, and R. Govindan. 2008. “The impact of Spatial Correlation on Routing with Compression in Wireless Sensor Networks.” Transactions on Sensor Networks (TOSN) (ACM) 4(4). Article 24. Pennebaker, W. P., and J. L. Mitchell. 1992. JPEG: Still Image Data Compression Standard. Norwell, MA: Kluwer Academic Publishers. Petrushin, V. A., G. Wei, and A. V. Gershman. 2006. “Multiple-Camera People Localization in an Indoor Environment.” Knowledge and Information Systems 10(2): 229–241. Pflugfelder, R., and H. Bischof. 2007. “People Tracking Across Two Distant Self-Calibrated Cameras.” In The 4th Conference on Advanced Video and Signal Based Surveillance (AVSS), 393–398. London, UK: IEEE. Polastre, J., R. Szewczyk, and D. Culler. 2005. “Telos: Enabling Ultra-Low Power Wireless Research.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN), 364–369. Los Angeles, CA: ACM/IEEE. Pottie, G.J., and W. J. Kaiser. 2000. “Wireless Integrated Network Sensors .” Communications of the ACM (ACM) 43(5): 51–58. Prati, A., R. Vezzani, L. Benini, E. Farella, and P. Zappi. 2005. “An Integrated Multi-Modal Sensor Network for Video Surveillance.” In The Third International Workshop on Video Surveillance and Sensor Networks (VSSN), 95-102. Singapore, Singapore: ACM. Qing, L., Q. Zhu, and M. Wang. 2006. “Design of a Distributed Energy-Efficient Clustering Algorithm for Heterogeneous Wireless Sensor Networks.” Computer Communications 19(12): 2230–2237.

396

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Raghunathan, V., C. Schurgers, S. Park, and M.B. Srivastava. 2002. “Energy-Aware Wireless Microsensor Networks.” Signal Processing Magazine. 19(2): 40–50. Raghunathan, V., S. Ganeriwal, and M. Srivastava. 2006. “Emerging Techniques for Long Lived Wireless Sensor Networks.” Communications Magazine (IEEE) 44(4): 108–114. Rahimi, M., R. Pon, W. J. Kaiser, G. S. Sukhatme, D. Estrin, and M. Srivastava. 2004. “Adaptive Sampling for Environmental Robotics.” In International Conference on Robotics and Automation (ICRA), 3537–3544. New Orleans, LA: IEEE. Raza, U., A. Camerra, A. L. Murphy, T. Palpanas, and G. P. Picco. 2012. “What Does Model-Driven Data Acquisition Really Achieve in Wireless Sensor Networks?” In The 10th International Conference on Pervasive Computing and Communications (PerCom), 85–94. Lugano, Switzerland: IEEE. Razzaque, M.A., C. Bleakley, and S. Dobson. 2013. “Compression in Wireless Sensor Networks: A Survey and Comparative Evaluation.” Transactions on Sensor Networks (TOSN) (ACM) 10 (1). Article 5. Rezvani, M., A. Ignjatovic, E. Bertino, and S. Jha. 2015. “Secure Data Aggregation Technique for Wireless Sensor Networks in the Presence of Collusion Attacks.” Transactions on Dependable and Secure Computing (IEEE) 12(1): 98–110. Rinner, B., T. Winkler, W. Schriebl, M. Quaritsch, and W. Wolf. 2008. “The Evolution from Single to Pervasive Smart Cameras.” In The 2nd International Conference on Distributed Smart Cameras (ICDSC), 1-10. Stanford, CA: ACM/IEEE. Sadler, C. M., and M. Martonosi. 2006. “Data Compression Algorithms for Energy-Constrained Devices in Delay Tolerant Networks.” In The 4th International Conference on Embedded Networked Sensor Systems (SenSys), 265–278. Boulder, CO: ACM. Said, A., and W. A. Pearlman. 1996. “An image Multiresolution Representation for Lossless and Lossy Compression.” Transactions on Image Processing (IEEE) 5(9): 1303–1310. Salomon, D. 2004. Data Compression: The Complete Reference. New York, NY: Springer-Verlag. Santos, T. T., and C. H. Morimoto. 2008. “People Detection under Occlusion in Multiple Camera Views.” XXI Brazilian Symposium on Computer Graphics and Image Processing (SIBGRAPI), 53–60. Campo Grande, Brazil: IEEE. Schoellhammer, T., B. Greenstein, E. Osterweil, W. Michael, and D. Estrin. Lightweight Temporal Compression of Microclimate Datasets. UCLA Center for Embedded Network Sensing. May 5, 2004. https://escholarship.org/uc/item/6zg2n1rh. Accessed 12 Aug 2018. Schott, B., M. Bajura, J. Czarnaski, J. Flidr, T. Tho, and L. Wang. 2005. “A Modular Power-Aware Microsensor with > 1000X Dynamic Power Range.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN). Los Angeles, CA: ACM, 2005. Article 66. Schreiber, D., A. Kriechbaum, and M. Rauter. 2010. “A Multisensor Surveillance System for Automated Border Control (eGate).” In The 10th International Conference on Advanced Video and Signal Based Surveillance (AVSS), 432–437. Krakow, Poland: IEEE. Semtech. 2018. XE1205: True Multi-Channel, Narrow-Band RF Transceiver. January 1, 2018. https://www.semtech.com/products/wireless-rf/fsk-transceivers/XE1205. Accessed 17 Aug 2018. Sensirion. 2018a. Digital Humidity Sensor SHT7x (RH/T). SENSIRION. January 1, 2018. https:// www.sensirion.com/en/environmental-sensors/humidity-sensors/pintype-digital-humiditysensors/. Accessed 17 Aug 2018. Sensirion. 2018b. Environmental Sensors. January 1, 2018. https://www.sensirion.com/en/ environmental-sensors/. Accessed 17 Aug 2018. Shah, R. C., S. Roy, S. Jain, and W. Brunette. 2003. “Data MULEs: Modeling and Analysis of a Three-Tier Architecture for Sparse Sensor Networks.” Ad Hoc Networks 1(2–3): 215–233. Sharp, C., et al. 2005. “Design and Implementation of a Sensor Network System for Vehicle Tracking and Autonomous Interception.” In The Second European Workshop on Wireless Sensor Networks (EWSN), 93–107. Istanbul, Turkey: IEEE.

References

397

Shu, W., X. Liu, Z. Gu, and S. Gopalakrishnan. 2008. “Optimal Sampling Rate Assignment with Dynamic Route Selection for Real-Time Wireless Sensor Networks.” In Real-Time Systems Symposium, 431-441. Barcelona, Spain: IEEE. Silberstein, A., and J. Yang. 2007. “Many-to-Many Aggregation for Sensor Networks.” In The 23rd International Conference on Data Engineering (ICDE), 986–995. Istanbul, Turkey: IEEE. SimIt-ARM. SimIt-ARM 3.0. June 26, 2007. http://simit-arm.sourceforge.net. Accessed 22 Aug 2018. Simon, G., et al. 2004. “Sensor Network-Based Countersniper System.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 1-12. Baltimore, MD: ACM. Singh, V. K., P. K. Atrey, and M. S. Kankanhalli. 2008. “Coopetitive Multi-Camera Surveillance using Model Predictive Control.” Machine Vision and Applications 19(5/6): 375–393. Sinha, A., and A. Chandrakasan. 2001. “Dynamic Power Management in Wireless Sensor Networks.” Design & Test of Computers (IEEE) 18(2): 62–74. Snidaro, L., I. Visentini, and G. L. Foresti. 2009. “Multi-sensor Multi-cue Fusion for Object Detection in Video Surveillance.” In The 6th International Conference on Advanced Video and Signal Based Surveillance (AVSS), 364–369. Genova, Italy: IEEE. Solis, I., and K. Obraczka. 2005. “Efficient Continuous Mapping in Sensor Networks using Isolines.” In The 2nd Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services (MobiQuitous), 325–332. San Diego, CA: IEEE. Somasundara, A. A., A. Kansal, D. D. Jea, D. Estrin, and M. B. Srivastava. 2006. “Controllably Mobile Infrastructure for Low Energy Embedded Networks .” Transactions on Mobile Computing (IEEE) 5(8): 958–973. Sommerlade, E., and I. Reid. 2010. “Probabilistic Surveillance with Multiple Active Cameras.” International Conference on Robotics and Automation (ICRA) (IEEE) 440–445. Szewczyk, R., A. Mainwaring, J. Polastre, J. Anderson, and D. Culler. 2004.“An Analysis of a Large Scale Habitat Monitoring Application.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 214-226. Baltimore, MD: ACM. Taylor, S. G., K. M. Farinholt, G. Park, M. D. Todd, and C. R. Farrar. 2010. “Multi-Scale Wireless Sensor Node for Health Monitoring of Civil Infrastructure and Mechanical Systems.” Smart Structures and Systems 6(5/6): 661–673. Teuhola, J. 1978. “A Compression Method for Clustered Bit-Vectors .” Information Processing Letters 7(6): 308–311. Texas Instruments. 2004. MSP430x11x2, MSP430x12x2 Mixed Signal Microcontroller. Texas Instruments, Inc. August 1, 2004. http://www.ti.com/lit/ds/symlink/msp430f1122.pdf. Accessed 8 Oct 2016. TinyOS. TinyOS. August 20, 2012. http://www.tinyos.net. Accessed 9 Oct 2013. Tseng, Y.-C., C.-S. Hsu, and C.-S. Hsu. 2003. “Power-Saving Protocols for IEEE 802.11-Based Multi-Hop Ad Hoc Networks.” Computer Networks 43(3): 317–337. Tu, P., N. Krahstoever, and J. Rittscher. 2007. “View Adaptive Detection and Distributed Site Wide Tracking.” In The 4th Conference on Advanced Video and Signal Based Surveillance (AVSS), 57–62. London, UK: IEEE. Tulone, D. 2004. “A Resource-Efficient Time Estimation for Wireless Sensor Networks.” The Joint Workshop on Foundations of Mobile Computing, 52–59. Philadelphia, PA: DIALM-POMC. Tulone, D., and S. Madden. 2006a. “An Energy-Efficient Querying Framework in Sensor Networks for Detecting Node Similarities.” In The 9th ACM International Symposium on Modeling Analysis and Simulation of Wireless and Mobile Systems (MSWiM), 191–300. Terromolinos, Spain: ACM, 2006. Tulone, D., and S. Madden. 2006b. PAQ: Time Series Forecasting for Approximate Query Answering in Sensor Networks. In Wireless Sensor Networks, vol. 3868, ed. K. Römer, H. Karl and F. Mattern, 21–37. Berlin, Heidelberg: Springer-Verlag.

398

5 Energy Management Techniques for WSNs (2): Data-Driven Approach

Vigorito, C. M., D. Ganesan, and A. G. Barto. 2007. “Adaptive Control of Duty Cycling in Energy-Harvesting Wireless Sensor Networks.” In The 4th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and Networks, 21–30. San Diego, CA: IEEE. Wang, J., Y. Liu, and S. K. Das. 2010. “Energy-Efficient Data Gathering in Wireless Sensor Networks with Asynchronous Sampling.” Transactions on Sensor Networks (TOSN) (ACM) 6 (3). Article 22. Wang, Y., P. Natarajan, and M. Kankanhalli. 2013. “Multi-camera Skype: Enhancing the Quality of Experience of Video Conferencing.” In The Era of Interactive Media, ed. J.S. Jin, C. Xu and M. Xu, 243–253. New York, NY: Springer. Welsh, M., and G. Mainland. 2004. “Programming Sensor Networks using Abstract Regions.” In The 1st Symposium on Networked System Design and Implementation (NSDI). San Fransisco, CA: USENIX. Werner-Allen, G., et al. 2006. “Deploying a Wireless Sensor Network on an Active Volcano.” Internet Computing (IEEE) 18–25. Willett, R., A. Martin, and R. Nowak. 2004. “Backcasting: Adaptive Sampling for Sensor Networks.” In The 3rd International Symposium on Information Processing in Sensor Networks (IPSN), 124-133. Berkeley, CA: ACM. Wu, C., A. M. Khalili, and H. Aghajan. 2010. “Multiview Activity Recognition in Smart Homes with Spatio-Temporal Features.” The 4th International Conference on Distributed Smart Cameras (ICDSC), 142–149. Atlanta, GA: ACM/IEEE. Xu, Y., J. Heidemann, and D. Estrin. 2001. “Geography-Informed Energy Conservation for Ad Hoc Routing.” In The 7th Annual International Conference on Mobile Computing and Networking (MobiCom), 70-84. Rome, Italy: ACM. Xu, Y., S. Bien, Y. Mori, J. Heidemann, and D. Estrin. 2003. Topology Control Protocols to Conserve Energy in Wireless Ad Hoc Networks. Technical, Center for Embedded Network Sensing (CENS), University of California, Los Angeles, CA: University of California. Yao, Y., and J. Gehrke. 2002. “The Cougar Approach to In-Network Query Processing in Sensor Networks.” Newsletter (ACM SIGMOD) 31(3): 9–18. Ye, W., J. Heidemann, and D. Estrin. 2002. “An Energy-Efficient MAC Protocol for Wireless Sensor Networks.” In The 21st Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM)., 1567–1576. New York, NY: IEEE. Younis, O., and S. Fahmy. 2004. “HEED: a hybrid, energy-Efficient, Distributed Clustering Approach for Ad Hoc Sensor Networks.” Transactions on Mobile Computing (IEEE) 3(4): 366–379. Zhou, B., L. H. Ngoh, B.S. Lee, and C. P. Fu. 2004. “A Hierarchical Scheme for Data Aggregation in Sensor Network.” In The 12th IEEE International Conference on Networks (ICON), 525-529. Singapore, Singapore: IEEE. Zhou, J., and D. De Roure. 2007. “FloodNet: Coupling Adaptive Sampling with Energy Aware Routing in a Flood Warning System.” Journal of Computer Science and Technology 22(1): 121–130.

Chapter 6

Energy Management Techniques for WSNs (3): Mobility-Based Approach

Wisdom is saving energy for the worthy …

6.1

Mobility in WSNs

Mobility is “a blessing” rather than a curse to network performance, a becoming familiar quote that will be fully investigated in this chapter. To start with, the traditional WSN architectures are based on the assumption that the network is dense, so that any two nodes can communicate with each other through multihop paths. As a consequence, in most cases the sensors are assumed to be static, and mobility is not considered as an option. Similar to the research trends in mobile ad hoc networks (MANETs) (Zhao and Ammar 2003) and delay-tolerant networking (DTNs) (Fall 2003), mobility has also been introduced to WSNs (Shah et al. 2003; Chakrabarti et al. 2003). In fact, mobility in WSNs is beneficial for several reasons (Anastasi et al. 2009b), as below deliberated: • Connectivity. As nodes are mobile, a dense WSN architecture may not be a requirement. Actually, mobile elements can cope with isolated regions, so that the constraints on network connectivity and nodes deployment/redeployment can be relaxed. A sparse WSN architecture consequently becomes a feasible option. • Cost. Since fewer nodes can be deployed, the network cost is reduced in a mobile WSN. Although adding mobility features to the nodes might be expensive, in many cases it is possible to exploit mobile elements already present in the sensing area, e.g., trains, buses, shuttles or cars, and attach sensors to them. • Reliability. Since traditional (static) WSNs are dense and the communication paradigm is often multihop, reliability is compromised by interference and collisions. In addition, the message loss increases with the number of hops, which may be rather high. Mobile elements, instead, can visit nodes in the network and collect data directly through single-hop transmissions. This reduces not only contention and collisions, but also the message loss. © Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_6

399

400

6 Energy Management Techniques for WSNs (3) …

• Energy efficiency. The traffic pattern inherent to WSNs is convergecast, i.e., messages are generated from sensor nodes and are collected by the sink. As a consequence, nodes closer to the sink are more overloaded than others and subject to premature energy depletion. This issue is known as the funnelling effect or the “energy hole problem” (Li and Mohapatra 2007), since the neighbors of the sink represent the bottleneck of traffic; it is also called the “crowded center effect” (Popa et al. 2007). Mobile elements can help reduce the funnelling effect, as they can visit different regions in the network and spread the energy consumption more uniformly, even in the case of a dense WSN architecture (Wang et al. 2005). However, mobility in WSNs also introduces significant challenges, which do not arise in static WSNs. These challenges are as itemized: • Contact detection. Since communication is possible only when the nodes are in the transmission range of each other, it is necessary to detect the presence of a mobile node correctly and efficiently. This is particularly true when the duration of contacts is short. • Mobility-aware power management. In some cases, it is possible to exploit the knowledge on the mobility pattern to further optimize the detection of mobile elements. Actually, if visiting times are known or can be predicted with certain accuracy, sensor nodes can be awake only when they expect the mobile element roaming in their transmission range. • Reliable data transfer. As available contacts might be scarce and short, there is a need to maximize the number of messages correctly transferred to the sink. In addition, since nodes move during data transfer, message exchange must be mobility-aware. • Mobility control. When the motion of mobile elements is controllable, a policy for visiting nodes in the network has to be defined. Hence, the path and the speed or sojourn time of mobile nodes have to be defined in order to improve (maximize) the network performance.

6.1.1

Architecture of WSNs with Mobile Elements

To better grasp the specific features of WSNs with mobile elements (WSN-MEs), the network architecture is introduced according to the role of the MEs. The main components of WSN-MEs are the following (Di Francesco et al. 2011): • Regular sensor nodes (just nodes, for short) are the sources of information. Such nodes perform sensing as their main task. They may also forward or relay messages in the network, depending on the adopted communication paradigm. • Sinks (basestations) are the destinations of information. They collect data sensed by sensor nodes either directly by visiting sensors and collecting data from each

6.1 Mobility in WSNs

401

of them or indirectly through intermediate nodes. They can use data coming from sensors autonomously or make them available to interested users through an Internet connection. • Special support nodes perform a specific task, such as acting as intermediate data collectors or mobile gateways. They are neither sources nor destinations of messages, but exploit mobility to support network operation or data collection. Noticeably, mobility might be involved at the different network components. For instance, nodes may be mobile while sinks are static, or vice versa. In any case, WSN-ME is a network where at least one of the above-mentioned components is mobile. Mobility of sensor nodes can be accomplished in different ways (Akyildiz and Kasimoglu 2004): • Sensors can be equipped with mobilizers for changing their location. As mobilizers are generally quite expensive from the energy consumption standpoint, adding mobility to sensor nodes may be not convenient. In fact, the resulting energy consumption may be greater than the energy gain due to mobility itself. • Instead of making each sensor node mobile, mobility can be limited to special nodes, which are less energy constrained than the ordinary ones. In this case, mobility is strictly tied to the heterogeneity of sensor nodes. Moreover, instead of providing mobilizers, sensors can be placed on elements, which are mobile of their own, e.g., animals, cars, etc. Two options are possible: • All sensors are put onto mobile elements, so that all nodes in the network are mobile. • A limited number of special nodes can be placed on mobile elements, while the other sensors are stationary. Depending on the specific scenario, the support nodes might be present or not. When there are only regular nodes, the resulting WSN-ME architecture is homogeneous or flat. On the other hand, when support nodes are also present, the resulting WSN-ME architecture is non-homogeneous or tiered. Furthermore, different from traditional WSNs, which are usually limited to be dense, WSN-MEs can also be sparse. The network architecture strongly depends on the role of the ME as elaborated in Sect. 6.1.2.

6.1.2

Role of Mobile Elements in WSNs

The mobile elements forming a WSN, as previously introduced in Sect. 6.1.1, affect the network architecture according to how they play their mobile role. A comprehensive description is offered in the coming items (Di Francesco et al. 2011):

6 Energy Management Techniques for WSNs (3) …

402

Sink (basestation)

Relocatable node

Relocatable node

Fig. 6.1 Architecture of a WSN-MSE with relocatable nodes (Di Francesco et al. 2011)

• Relocatable nodes. These are mobile nodes that change their location to better characterize the sensing area, or to forward data from the source nodes to the sink. In contrast with mobile data collectors, discussed below, relocatable nodes do not carry data as they move in the network. In fact, they only change the topology of the network, assumed to be rather dense, for connectivity or coverage purposes. More specifically, after moving to the new location, they usually remain stationary and forward data along multihop paths. A WSN-ME architecture based on relocatable nodes is depicted in Fig. 6.1. Although in theory ordinary nodes might be relocatable, in most cases special MEs are used, e.g., support nodes. Relocatable nodes can also be used to address the problem of sensing coverage. In this case, the primary concern is not ensuring network connectivity, but avoiding coverage holes, areas where the density of nodes is not adequate to properly characterize a phenomenon or detect an event. Approaches targeted for sensing coverage focus on sensor deployment (Wang et al. 2007), sensor relocation and dispatch (Yoon et al. 2011), or both (Wang et al. 2008b). Relocatable nodes provide a mobility-assisted approach to WSNs, in the sense that MEs are not actively exploited for data collection; therefore, the following bullets will not further discuss relocatable nodes. Instead, there is a focus on the MEs that are actively used for data collection in mobility-based approaches. • Mobile data collectors (MDCs). These are mobile elements visiting the network to collect data generated from source nodes. Depending on the way they manage the collected data, MDCs can be either mobile sinks or mobile relays: – Mobile sinks (MSs). These are mobile nodes, which are the destination of messages originated by sensors, i.e., they represent the endpoints of data collection in WSN-MEs. They can either autonomously consume collected

6.1 Mobility in WSNs

403

Mobile sink

Mobile sink (a) Mobile sinks

Sink (basestation)

Mobile relay

Mobile-relay (b) Mobile relays

Fig. 6.2 Architecture of WSN-MEs with MDCs (Di Francesco et al. 2011)

data for their own purposes or make them available to remote users by using a long-range wireless Internet connection (Rao and Biswas 2010). The MS-based WSN-ME architecture is portrayed in Fig. 6.2a. – Mobile relays (MRs). These are support nodes that gather messages from sensor nodes, store them, and carry the collected data to sinks or basestations. They are not the endpoints of communication, but only act as mobile forwarders. This means that the collected data move along with them, until the MRs get in contact with the sink or basestation (Shah et al. 2003; Jain et al. 2006). The MR-based WSN-ME architecture is illustrated in Fig. 6.2b.

404

6 Energy Management Techniques for WSNs (3) …

Sink (basestation)

Fig. 6.3 Architecture of a WSN-ME with mobile peers (Di Francesco et al. 2011)

• Mobile peers. Unlike MDCs, which are either sinks or special relay nodes, mobile peers are ordinary mobile sensor nodes in WSN-MEs. Since they can be both originator and relays of messages in the network, their interactions are symmetrical because the sink might also be mobile. When a peer is in the communication range of the basestation, it transfers its own data as well as those gathered from other peers while moving in the sensing area. A WSN-ME architecture based on mobile peers is depicted in Fig. 6.3. In this case, the network is homogeneous and rather sparse. Mobile peers have been successfully employed in the context of wildlife monitoring applications, such as tracking zebras in the ZebraNet project presented in Chap. 7 of this book (Juang et al. 2002) or whales in the shared wireless infostation model (SWIM) (Haas and Small 2006). Sensor nodes are attached to animals and act as peers, so that not only do they generate their own data, but also carry and forward all data coming from other nodes that have been previously in contact. When mobile peers get close to a basestation, they transfer all the gathered data. Data already been transferred to a basestation are flushed by peers in order to save storage. Mobile peers can also be used for opportunistic data collection in urban sensing scenarios (Campbell et al. 2008). Typical applications include personal monitoring such as physical exercise tracking, civil defense like hazards and hotspot reporting to police officers, and collaborative applications such as information sharing for tourism purposes. In this context, sensors are not used mainly for monitoring the environment, but are rather exploited to characterize people in terms of both interactions and context (state) information. An example application is represented by handheld mobiscopes (Abdelzaher et al. 2007) where handheld devices, such as cell phones or PDAs, gather data from the surrounding environment and report them to servers, which provide services to remote users.

6.1 Mobility in WSNs

405

Most of the issues in the context of WSN-MEs based on mobile peers are similar to classic DTNs (Anastasi et al. 2009a). An interesting overview on data collection in WSNs with MEs is available in Di Francesco et al. (2011).

6.2

Mobility-Based Approach Taxonomy

The mobility-based approach, as described in the previous section, is divided into mobile sink and mobile relay approaches (Fig. 6.4): • The mobility of a mobile sink can be classified into uncontrolled mobility and controlled mobility. In uncontrolled mobility, the mobile sink can move randomly in the monitored region (Wang et al. 2005; Shi and Hou 2008) presented in Sections “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” and “Energy-Aware Routing to Maximize Lifetime in WSNs with Mobile Sink,” respectively; while in controlled mobility, the mobile sink can only move along the predefined trajectory (Basagni et al. 2008; Liang et al. 2010) detailed in Sections “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” and “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime,” respectively. • The mobile relay approach is built on a heterogeneous WSN architecture composed of a few resource rich mobile relay nodes and a large number of simple static nodes. The mobile relays have more energy than the static sensors; they can dynamically move around the network and help relieve static sensors that are heavily burdened by high network traffic, thus extending their lifetime (Jain et al. 2006; Wang et al. 2008a) described in Sects. 6.2.2.1 and 6.2.2.2, respectively. Sections 6.2.1 and 6.2.2 target a comprehensive description of the mobile sink and mobile relay protocols, respectively. A comparative assembly of the techniques offered throughout this chapter is made available in Table 6.1.

Mobility-based

Mobile sink Sect. 6.2.1

Mobile relay Sect. 6.2.2

Fig. 6.4 Mobility-based approach taxonomy (Anastasi et al. 2009b)

406

6.2.1

6 Energy Management Techniques for WSNs (3) …

Mobile Sink Protocols

For data delivery and dissemination in WSNs, the main aim is minimizing the nodes energy consumption, mostly due to radio communications, in order to increase the network lifetime, the time where the network is able to perform its intended operations. Independently of all the energy-efficient techniques developed at the different layers of the nodes protocol stack, the ultimate problem is the delivery of the sensed data from all the sensors to the sink, which imposes greater burden on the nodes closer to the sink. More specifically, when a sink is statically placed, the sensor nodes that can directly communicate with it, the sink neighbors, tend to deplete their energy faster than other nodes; not only they consume energy to communicate their own data to the sink, but also for relaying to it the data from any other node. This problem, termed the “sink neighborhood problem,” leads to a premature disconnection of the network (Basagni et al. 2008). The sink gets isolated from the rest of the network due to the death of its neighbors while most of the sensor nodes are still fully operational. One way for mitigating, if not obviating, the “sink neighborhood” problem is by exploiting the mobility of some of the network components. The key idea is changing the neighbors of the sink so that the energy consumption for data packet relaying is balanced throughout the network. Since moving the nodes would require extra power from the already limited energy of a node, the most promising way of changing the sink neighbors is to have the sink moving to different parts of the deployment area, while keeping the sensors static. Protocols proposed for sink mobility differ on the nature of the mobility itself, whether uncontrolled or controlled: • Uncontrolled sink mobility is used in those applications where the sink is sent to gather data through the network at times and along routes that are out of the control of the network (Wang et al. 2005; Shi and Hou 2008). Whether random or deterministic, the sink movement proceeds according to a schedule, which is not determined by the prevailing network conditions, such as data traffic or the nodes residual energy. Sections “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” and “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” emphasize with detail two uncontrolled sink mobility schemes. • Controlled sink mobility scheme is meant to extend network lifetime (Basagni et al. 2008; Liang et al. 2010). Proposed solutions are mostly centralized, in the sense that the proposed schemes determine optimal sink routes and sojourn times based on the knowledge of global network parameters. Section “Controlled Sink Mobility for Prolonging WSNs Lifetime (GMRE)” presents a controlled sink mobility scheme and Section “Maximizing the Lifetime of WSNs with Mobile Sink in Delay-Tolerant Applications (DT-MSM)” views controlled sink mobility in a delay-tolerant environment.

6.2 Mobility-Based Approach Taxonomy

6.2.1.1

407

Uncontrolled Sink Mobility Protocols

Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime This study is conducted to exploit the mobility of data collection points (sinks) for the purpose of increasing the lifetime of a WSN with energy-constrained nodes (Wang et al. 2005). A new linear programming formulation is suggested, for the joint problems of determining the movement of the sink and the sojourn time at different points in the network, that induce the maximum network lifetime. Differently from previous solutions, the objective function maximizes the overall network lifetime, defined as the time till the first node “fades out” because of energy depletion, rather than minimizing the energy consumption at the nodes. For WSNs with up to 256 nodes, the proposed model produces sink movement patterns and sojourn times leading to a network lifetime up to almost five times that obtained with a static sink. Simulation results are performed to determine the distribution of the residual energy at the nodes over time. These results confirm that energy consumption varies with the current sink location; the nodes more drained are those in the proximity of the sink. Furthermore, the proposed solution for computing the sink movement leads to a fair balancing of the energy depletion among the network nodes. Some simplifying assumptions are made to build the system model: • Sensors remain stationary at the nodes of a bi-dimensional square grid composed of same-size cells. • The sink can move freely on the grid from one node to another. During its sojourn time at a node, sensors can communicate with the sink. For analytical simplicity, the traveling time of the sink between two nodes is considered negligible. • Data transmission and reception are the major energy-consuming activities. • Sensor nodes are homogeneous, and wireless channels are bidirectional, symmetric, and error-free. • Each node has a limited initial energy and unlimited buffer size. • Sensor nodes communicate with the sink by sending data via multiple hops along the shortest path; a hop is of one cell side length, i.e., the distance between two adjacent nodes in the grid equals the nodes transmission range. The sensor network is modeled as a graph GðN; E Þ where N is the set of all nodes in the square grid and E is the set of all links (i, j), where i and j are neighboring nodes. A node i can communicate directly with its (at most) four neighboring nodes, Si is the set of i’s neighbors.

408

6 Energy Management Techniques for WSNs (3) …

In the proposed model, each sensor generates data packets at a fixed data rate. If a sensor node i is neither colocated with sink k, nor directly connected with it, i.e., if k is not colocated with any of the nodes in Si, then data packets generated at node i have to be relayed through multiple hops to reach the sink. The sink can only be located at one node position in the grid (the sensor locations and the possible sink locations are the same). The sink keeps moving among grid positions until the maximum network lifetime is reached, which occurs when one sensor node residual energy drops below a predefined threshold required for it to operate; when this happens, the sensor “fades out.” In the proposed model, the network lifetime is calculated as the sum of sojourn times of the sink at all visited nodes. The sojourn times are constrained by the fact that the total energy spent by each node when the sink is colocated with different nodes cannot exceed the sensor node initial energy. When a sensor node lies on the same horizontal or vertical line of the current position of the sink, a unique shortest path exists between the two nodes; otherwise, multiple shortest paths exist. For example, as illustrated in Fig. 6.5, six shortest paths exist between sensor i and sink k, each path is four hops long. Three of those paths are shown, path 1 and path 2 are along the perimeter of the rectangle defined by nodes i and k, and path 3 is one of the four interior paths. In the routing protocol, only the two paths along the perimeter of the rectangle are considered (paths 1 and 2 in Fig. 6.5). These two routes are taken at equal frequencies, or equivalently, the route alternates between the two paths. When calculating power consumption, the first-order radio model is frequently used. For receiving k1 bits/s, the power consumption, pr, at a sensor node is pr ¼ k 1  b

ð6:1Þ

where b is a factor indicating the energy consumption per bit at the receiver circuit.

Fig. 6.5 Shortest paths from a sensor to the sink (Wang et al. 2005)

k

i

Sink

6.2 Mobility-Based Approach Taxonomy

409

The power pt needed for transmitting k2 bits/s is
 pt ¼ k 2  a1 þ a2  d P

ð6:2Þ

where, a1 a2  d P

is the energy consumption factor indicating the power consumed per bit by the transmitter circuit, indicates the energy consumption on the amplifier per bit, d being the physical distance between the transmitting and the receiving node, and p the path loss exponent1 (Rouphael 2009), usually between 2 and 4, depending on the environment.

The transmission radius of a sensor node is usually very limited, a few tens of meters, so that the energy spent for the transceiver circuitry exceeds the energy consumption due to the emitted power. According to the energy model of real-life sensor nodes prototypes, an energy model is adopted such that the energy consumed when transmitting is basically constant, and the energy consumed for receiving a bit is the same as the energy consumed for transmitting a bit, denoted by e: b  a1 þ a 2  d P ¼ e

ð6:3Þ

Therefore, the total energy consumption at a node per time unit is
 pr þ p t ¼ k 1  b þ k 2  a 1 þ a 2  d P  e  ð k 1 þ k 2 Þ

ð6:4Þ

Mathematically, the power consumption for receiving and transmitting packets at sensor node i when the sink sojourns at node k (in J/s) is computed from Eq. 6.4 as follows: 0 1 X X k k k ð6:5Þ ci ¼ e  @ fij þ fji A; i; k 2 N and i 6¼ k j2Si

j:i2Sj

and

1

Path loss is intimately related to the environment where the transmitter and receiver are located. Path loss models are developed using a combination of numerical methods and empirical approximations of measured data collected in channel sounding experiments. In general, propagation path loss increases with frequency as well as distance:
 Pl ¼ 10  log10 16  p2  d n =k2

where Pl is the average propagation path loss, d is the distance between the transmitter and receiver, n is the path loss exponent which varies between 2 for free space to 6 for obstructed in building propagation, and k is the free space wavelength defined as the ratio of the speed of light in meters per second to the carrier frequency in Hz.

6 Energy Management Techniques for WSNs (3) …

410

cki ¼ e  r;

i; k 2 N and i ¼ k

ð6:6Þ

where, cki e r

denotes the power consumption for receiving and transmitting packets at node i when the sink sojourns at node k (J/s), is the energy consumption coefficient for transmitting or receiving one bit (J/ bit), is the rate, assumed the same for all nodes, at which data packets are generated (bits/s).

Equation 6.6 holds when the sink is colocated with node i; it expresses the fact that all nodes in Si communicate directly with the sink. Considering the data balance flow at each node, within each time unit, the total incoming data packets plus the data packets generated at the node equal the total outgoing data packets from the node: X

fjik þ r ¼

j:i2Sj

X

fijk ;

i; k 2 N

ð6:7Þ

j2Si

where fijk is the data transmission rate from node i to node j while the sink is at node k (bits/s). The linear programming (LP) model below determines the sojourn times tk of the sink at each node k 2 N so that the network lifetime is maximized. If the optimal value for a tk is 0, the sink does not visit node k. Every node k 2 N whose optimal tk is positive is visited by the sink for a time duration equal to tk. The sink visiting order is not important since the traveling time of the sink between nodes is considered negligible, and the data generation rate is independent of time. Specifically, the LP formulation is given by: Maximize z ¼

X

tk

ð6:8Þ

i2N

ð6:9Þ

k2N

such that X

cki  tk  e0

k2N

tk  0;

k2N

ð6:10Þ

The objective function in Eq. 6.8 maximizes the network lifetime, i.e., the sum of sojourn times of the sink at all visited nodes. In Eq. 6.9, e0 is the initial energy (Joules) of each node minus the threshold energy required for node operation. The

6.2 Mobility-Based Approach Taxonomy

411

term cki  tk represents the energy consumed at node i for receiving and transmitting data during the time interval of the sink sojourn at node k. The total energy consumed at each node is computed as the sum of the energies consumed over all sojourn times of the sink at visited nodes. The constraint Eq. 6.9 simply states that the energy consumed at each node i should not exceed the initial energy of that node. The constraint Eq. 6.10 assures the non-negativity of sojourn time tk . The calculation of cki is illustrated below using the 5  5 grid displayed in Fig. 6.6: • Each node position is defined using the ordered pair of the node column and row number (x, y), x = 0, 1, …, L − 1, y = 0, 1, …, L − 1. A pair of horizontal and vertical dotted lines is drawn enclosing the nodes associated with the row and the column of the sink; these lines partition N, the set of all nodes in the square grid, into nine subsets. • Equations 6.5, 6.6, and 6.7 are used to accumulate flows and compute cki . • According to the routing protocol defined earlier (Fig. 6.5), node i transmits its own generated packets to nodes j2 and j4, one half each. Nodes j2 and j4 relay these packets to the sink node k. In addition, node i receives half of the packets generated at nodes j1 and j3 and half of the packets generated at nodes l1 and l2. Then, node i retransmits the packets originated at nodes j3 and l2 to node j2 and those originated at nodes j1 and l1 to node j4. Note that Si ¼ fj1 ; j2 ; j3 ; j4 g and since i 2 U  R, node i receives only from nodes j1 and j3 and transmits to nodes j2 and j4. In summary, node i receives at a rate 2 * r and transmits at a rate 3 * r, having therefore power consumption cki ¼ 5  r  e. • Depending on the position (x, y) of node i and the node subset to which it belongs, the following formulas can be derived for cki :

0

1

2

3

4

x

VA

0 1 UL

2

UR

3 HL 4

LL

HR VB

LR

y

Legend: UL (upper left), UR (upper right), LL (lower left), LR (lower right), VA (vertical above), VB (vertical below), HL (horizontal left), HR (horizontal right), and node k (with which the sink is co-located).

Fig. 6.6 Dataflows received and transmitted a node i (Wang et al. 2005)

412

6 Energy Management Techniques for WSNs (3) …

8 e  r  ½ðx þ 1Þ  ð1 þ LÞ  1 > > > > e  r  ½ðL  xÞ  ð1 þ LÞ  1 > > > > e  r  ½ðy þ 1Þ  ð1 þ LÞ  1 > > > > e  r  ½ðL  yÞ  ð1 þ LÞ  1 < cki ¼ e  r  ð1 þ x þ yÞ > > e  r  ðL  x þ yÞ > > > > > e  r  ðL þ x  yÞ > > > > e  r  ð2  L  x  y  1Þ > : er

i2HL i2HR i2V A i2V B i2UL i2UR i2LL i2LR i¼k

ð6:11Þ

The computation of cki was programmed in C. The solution of the LP model for a given set of parameters was obtained thru LINGO (LINDO Systems 2019). The maximum network lifetime, zs , and the node location at which it is achieved, when the sink remains static, can be acquired by solving the following model:   e0 zs ¼ max mini k ; k ci

i; k 2 N

ð6:12Þ

Solving the LP model given in Eqs. 6.8, 6.9, and 6.10 and the model described by Eq. 6.12 on networks of 3  3, 4  4, 5  5, …, 16  16 nodes, the results depicted in Figs. 6.7 and 6.8 are acquired, for the parameters values r = 1 bit/s, e = 0.62 lJ/bit, and e0 = 1.35 J. As clarified in the figures, several findings may be drawn out: • Figure 6.7 shows that as the network size increases the network lifetime decreases. This is due to the fact that each node, acting as a relay for a higher number of nodes, has to receive and transmit a higher number of packets, which leads to faster energy depletion. In the case of static sink, the network lifetime is clearly shorter since the sensor nodes close to the sink always relay the packets of all other nodes, which drains out their energy rapidly. • Figure 6.8 displays the lifetime improvement ratio as it changes with L, for a L  L grid where L is even or odd. In both cases, the improvement increases with the network size. For the grid with even number of nodes, the improvement is higher due to a relatively lower zs; the network lifetime is maximized when the sink stays at one of the four central nodes. For odd values of L, the sink is instead colocated with the unique central node and this results in an uneven distribution of dataflows and thus in a lower network lifetime zs in case of even values of L. • Investigating the pattern of the distribution of the sink sojourn times at the different nodes and the corresponding node energy consumption, it was found that:

6.2 Mobility-Based Approach Taxonomy

413

*10 9

zm zs

8 7

Lifetime (sec)

6 5 4 3 2 1 0 0 9 16 25 36

64

81

100

121

144

169

196

225

Number of nodes

Legend: zm is the optimal network lifetime in case of a mobile sink, and zs is the maximum network lifetime for a static sink.

Fig. 6.7 Network size versus lifetime (Wang et al. 2005)

– Independently of the size of the grid, the results show a similar pattern; the sink sojourns generally at the four corners for most of the time, and in the grid central area. This implies that when the sink is at one of the four corners, the nodes close to it and along the row/column of the sink spend the most energy. By locating the sink at one of the corners, all nodes in that corner, except the one colocated with the sink, deplete their energy significantly. – In the case where the sink starts by sojourning first at the four corners (order is irrelevant), the nodes in the central area still have a relatively high residual energy. This makes it appealing for the sink to move toward the central area to extend the network lifetime. – In general, the higher the network size, the more energy the nodes spend to deliver the data to the sink, the lower sojourn times at the corners (their energy deplete faster), and the lower residual energy at the central nodes when the four corners low residual energy demands a sink relocation. Analytically and thru simulation, it was concluded that by exploiting a mobile sink and selecting the sink movement according to the solution of the suggested model, the network energy gets more balanced among the network nodes and dataflow bottlenecks can be more effectively avoided. Hence, in networks with 256 nodes, the improvement obtained on the overall network lifetime was up to 450% increase over the static sink case.

6 Energy Management Techniques for WSNs (3) …

414 5 4.5 4

(zm-zs)/zs

3.5 3 2.5 2 1.5 1 0.5

Even-number grid Odd-number grid

0 0 9 16 25 36

64

81

100

121

144

169

196

225

Number of nodes

Fig. 6.8 Improvement ratio in network lifetime when the sink is mobile (Wang et al. 2005)

Energy-Aware Routing to Maximize Lifetime in WSNs with Mobile Sink By combining the model presented in Section “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” and the LP formulation for maximum lifetime routing described in Chang and Tassiulas (2004), this work presents another centralized solution for the problem of maximizing network lifetime (Papadimitriou and Georgiadis 2006). By turning a constant of the model in Section “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime” into a variable, the model presented in this section jointly solves the problem of determining the sink sojourn times at the given sites, and the routing of packets to the current position of the sink. The focus is on the problem of maximizing the lifetime of a WSN where the sensor nodes communicate with the sink by delivering the sensed data across multiple hops with different transmission energy requirements. That is, there is flexibility of transmitter power adjustment, and the energy consumption rate per unit information transmitted is not the same for all neighbors of a sensor node, but depends on the choice of the next hop node. The lifetime of the network is commonly defined as the time until a sensor node drains out of battery energy for the first time. In the suggested setup, the sensors are realistically randomly deployed in the field; their placement does not rely on any specific pattern, e.g., grid network. The sink is mobile and can move to different places during network operation; the sensors locations and possible sink locations are not necessarily the same. The problem is, in order to maximize network lifetime, for how long the sink must stay at each place, and how the sensors must deliver their data to the sink during its sojourn time at a given location. It is shown that maximum network lifetime can be

6.2 Mobility-Based Approach Taxonomy

415

achieved by solving optimally two joint problems; typically, the scheduling problem that determines the sink sojourn times and the routing problem to find the appropriate energy-efficient paths. These two optimization problems can be defined as a LP model, capable of expressing network lifetime in terms of sink sojourn times at possible locations. The model provides the optimal solution to both of these problems and gives the best achievable lifetime. The proposed WSN model consists of a set N of sensor nodes and a sink node s collecting the information. After been randomly deployed in the field, the sensors remain stationary at their initial locations and continuously monitor the physical environment where they have been placed. Hence, there is a constant information generation rate Qi [ 0 at every sensor i 2 N, not necessarily the same for every sensor. On the contrary, the sink is mobile and can be found in different random places during network operation, not necessarily colocated with the sensors. Let L be the set of possible sink locations and tl  0 the time for which it stays at location l 2 L. It is assumed that the traveling time of the mobile sink from one location to another is small and thus can be neglected (Wang et al. 2005). Sensors communicate with the sink during its sojourn time at a given location by delivering the sensed data across multiple hops. That is, for a given location l 2 L, the sink is not necessarily within the transmission range of every sensor. Let Sli N [ fsg be the set of nodes, either sensors or the sink, that are in the transmission range of sensor i 2 N for a given location l 2 L of the sink. If j 2 Sli , then j is called a neighboring node of i for location l. Note that the only element that may be different among two sets Sli1 ; Sli2 ; l1 ; l2 2 L is the sink s, since the rest of the network, consisting of all sensors, remains static. Considering the WSN in Fig. 6.9, sensors B and C are in the transmission range of A, regardless where the sink node is located. When s is placed at location 1, sensor A can also communicate with s. Therefore, the set of neighboring nodes of sensor A for location 1 is S1A ¼ fB; C; sg. However, when s moves to locations 2, 3, 4, it is not in the transmission range of sensor A. Hence, S2A ¼ S3A ¼ S4A ¼ fB; Cg. Similarly for another sensor node, say node C, S1C ¼ S2C ¼ S4C ¼ fA; B; Dg; while S3C ¼ fA; B; D; sg. Every sensor node i 2 N has an initial amount of battery energy Ei [ 0. The sink has no energy constraint. The energy consumed at sensor i to transmit a data unit to its neighboring node j is denoted by eTij [ 0 and the energy consumed for reception by the receiver j is denoted by eRij [ 0. Note that power control is possible, i.e., the energy expenditure for an information unit transmitted by sensor i depends on the next hop node and is not necessarily the same for every neighbor j. The above description of the WSN indicates that in order to transfer the information from the sensors to the sink, two complementary algorithms are necessary: • A scheduling algorithm that determines, for every location l 2 L, the duration tl for which the sink stays at that place. • A routing algorithm to find energy-efficient paths from each sensor to sink for all locations l 2 L for which tl [ 0.

6 Energy Management Techniques for WSNs (3) …

416 Location 1 s

Location 2 s

Sensor nodes: N={A,B,C,D} Sink locations: L={1,2,3,4} _____ Permanent link between two sensor nodes

s Location 3

s

-------- Link between a sensor and sink; exists only when the sink is at the corresponding location

Location 4

Legend: Sensor nodes communicate with the sink during its sojourn time at a location by delivering the sensed data across multiple hops.

Fig. 6.9 Sensor nodes sink communication (Papadimitriou and Georgiadis 2006)

Since the sink can be found in different places, the decision of the routing algorithm depends on its location. Let qlij be the rate at which information is transmitted from sensor i to its neighboring node j, assigned by the routing algomax rithm during time tl . For each qlij , there is a constraint qlij  qmax is the ij , where qij maximum possible rate at which information can be transmitted from i to j. These bounds can be viewed as link capacity constraints determined by the network environment. Also, there is a power constraint for every sensor node; for every sink location l 2 L, the power expenditure at sensor i 2 N incurred by transmissions and receptions of sensor i during time tl cannot exceed a maximum value Pi . This value reflects the limitations imposed by the sensors hardware. The overall objective is to maximize the duration of network operation before a sensor drains out of battery energy for the first time. In the presented model, the network lifetime is equal to the sum of the sink sojourn times at all possible locations (Eqs. 6.24–6.29). The sojourn times are constrained by the fact that the total energy consumed by each sensor for all sink locations cannot exceed the sensor initial amount of energy. In the following, model optimization is formulated as a LP problem (Nash and Sofer 1996). Given the sink sojourn times tl and the information transfer rates qlij ; i 2 N; j 2 Sli ; l 2 L, the energy consumed per time unit at sensor node i when the sink is placed at location l is given by: X X eTij  qlij þ eRji  qlji ð6:13Þ j2Sli

j:i2Slj

while the corresponding energy consumption for time duration tl is given by:

6.2 Mobility-Based Approach Taxonomy

0 @

X

417

eTij  qlij þ

j2Sli

X

1 eRji  qlji A  tl

ð6:14Þ

j:i2Slj

The total energy consumed at sensor i during network operation is the sum of the quantities in Eq. 6.14 over all locations l 2 L: XX

XX

eTij  qlij  tl þ

l2L j2Sli

eRji  qlji  tl

ð6:15Þ

l2L j:i2Slj

The network lifetime is defined as the length of time until the first battery drain-out among all sensors in N. It can also be expressed as the sum of the sink sojourn times at all possible locations: X

tl :

ð6:16Þ

l2L

The goal is to find the sink sojourn times tl and the information transfer rates qlij that maximize the network lifetime. This is under the flow conservation condition and the constraint that the total energy consumed by each sensor node when the sink stays at different locations cannot exceed the sensor initial amount of energy, Ei . From the above definitions, the problem of maximizing the overall network lifetime can be written as follows: Maximize

X

tl subject to

ð6:17Þ

l2L

ð6:18Þ

l2L

tl  0; 0  qlij  qmax ij ; X

eTij  qlij þ

j2Sli

XX l2L

X

i 2 N; j 2 Sli ; l 2 L

eRji  qlji  Pi ;

ð6:19Þ

i 2 N; l 2 L

ð6:20Þ

j:i2Slj

eTij  qlij  tl þ

XX l2L

j2Sli

X j:i2Slj

qlji þ Qi ¼

eRji  qlji  tl  Ei ;

i2N

ð6:21Þ

j:i2Slj

X

qlij ;

i 2 N; l 2 L

ð6:22Þ

j2Sli

where Qi is the information generation rate a sensor node i 2 N and is Ei the initial amount of battery energy at sensor node i 2 N. Note that flow conservation condition Eq. 6.22 applies to each location l 2 L separately. That is, for every sink location, the sum of the total incoming

6 Energy Management Techniques for WSNs (3) …

418

information transfer rate and the information generation rate at a sensor is equal to the total outgoing information transfer rate from the sensor. Since sink node s is the only destination of the dataflows generated by the sensors, it holds by definition that: X qlsj ¼ 0; j 2 N; l 2 L; and Qs ¼  Qi ð6:23Þ i2N

By defining q^lij ¼ qlij  tl as the amount of information transmitted from sensor i to its neighboring node j during time tl , the optimization problem becomes Maximize

X

tl subject to

ð6:24Þ

l2L

ð6:25Þ

i 2 N; j 2 Sli ; l 2 L

ð6:26Þ

l2L

tl  0; 0  ^qlij  qmax  tl ; ij X

eTij  ^qlij þ

j2Sli

X

eRji  ^qlji  Pi  tl ;

i 2 N; l 2 L

ð6:27Þ

j:i2Slj

XX l2L j2Sli

X j:i2Slj

eTij  ^qlij þ

XX

eRji  ^qlji  Ei ;

i2N

ð6:28Þ

l2L j:i2Slj

^qlji þ Qi  tl ¼

X

^qlij ;

i 2 N; l 2 L

ð6:29Þ

j2Sli

The objective function in Eq. 6.24 maximizes network lifetime, that is, the sum of the sojourn times of the sink at all possible locations. The constraints in qlij , respecEqs. 6.25 and 6.26 ensure the non-negativity of the quantities tl and ^ tively. Multiplying the constraints, on the transmission rates in Eq. 6.19 and the power of sensors in Eq. 6.20, by tl gives Eqs. 6.25 and 6.26, respectively. The left part of the inequality in Eq. 6.28 represents the total amount of energy consumed at sensor i for transmitting and receiving data over all sojourn times of the sink at visited locations. Hence, the energy constraint in Eq. 6.28 simply states that the energy consumed at each sensor i should not exceed its initial energy Ei . Finally, the flow conservation condition Eq. 6.29 is derived by multiplying Eq. 6.22 by tl . The LP model above determines for every location l 2 L the duration tl for which the sink stays at that place and the information transfer rates ^ qlij ; i 2 N; j 2 Sli , so that the network lifetime is maximized. The information transfer rate can also be computed as qlij ¼ ^qlij =tl ; therefore, the model provides an optimal solution to the scheduling and the routing problems described earlier. If the optimal value for a tl is 0, the sink does not visit location l. Every location l 2 L, for which the optimal tl is positive, is visited by the sink for a time duration equal to tl . The sink visiting order

6.2 Mobility-Based Approach Taxonomy

419

is not important since the traveling time of the sink between two locations is considered negligible and the information generation rate independent of time. A more general version of this LP model is to consider that while the sink s is at location l 2 L, the information transfer rates qlij ; i 2 N; j 2 Sli are allowed to change. It may seem plausible that by modifying the rate qlij at which information is transmitted from sensor i to its neighboring node j, during time tl , the amounts of sensors remaining energy can be utilized more efficiently; hence, prolonging the duration of network operation. However, it is proved that this generalization of the problem does not result in an improvement of the network lifetime (Papadimitriou and Georgiadis 2006). To evaluate the performance of the optimal LP formulation (LP-opt) proposed in Eqs. 6.24–6.29, it is compared against three models: • LP model with shortest path routing (SPR). This is a generalization of Wang et al. 2005 presented on Section “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime,” so that it can be applied to general networks, instead of being restricted to square grid networks with homogeneous sensors without power control. • LP model with multiple shortest path routing (MSPR). When there are multiple shortest paths, the model in Wang et al. (2005) selects two paths and alternates the route between them. It is modified so that the routing algorithm uses all the existing shortest paths from a sensor to the sink. The LP model that determines the sink sojourn times remains the same. • LP model for the static sink case (static sink). When the sink remains static, the lifetime achieved at every location l 2 L is determined separately and the one that gives the maximum value is selected. Given the location that maximizes lifetime, the sink stays there until a sensor node runs out of battery energy for the first time and does not move to another place. For each location l 2 L separately, the LP formulation given in Eqs. 6.24–6.29 is used, replacing L by L0 ¼ flg. Therefore, the static sink model maximizes lifetime for every location, but it is not optimal for the overall objective being the sum of the sink sojourn times at all possible locations. The results were obtained from 100 randomly generated network instances for each network size considered. Random networks are generated with a specified number of sensors (20, 40, …, 100) by fixing a square grid of 100  100 points. A number of these points are randomly selected with uniform probability to represent the sensor nodes of the network. The placement of the sensors does not form a grid network, but models a random deployment of the sensors on the terrain. Two different scenarios are made for the placement of the sink (Fig. 6.10): • The coordinates of possible sink locations are (25, 25), (25, 75), (75, 25), (75, 75). That is, the grid is split to four quarters; for a given location, the sink is within the transmission range only of the sensors that lie on the corresponding quarter of the grid.

6 Energy Management Techniques for WSNs (3) …

420

(0,100)

(25,75)

(100,100)

(75,75) (50,50)

(25,25)

(75,25)

Scenario 1

(0,0)

Scenario 2 Sink node location Sensor node

(100,0)

Fig. 6.10 Sink node placement (Papadimitriou and Georgiadis 2006)

• The sink can be placed at the four corners and at the center of the grid, that is, the coordinates of possible sink locations are (0, 0), (0, 100), (100, 0), (100, 100), (50, 50). When the sink is located at one of the four corners, it can be reached by the sensors that lie on the corresponding quarter of the grid as before. When it is placed at the center, it is within the transmission range of the sensors that lie on the area (25, 25), (25, 75), (75, 25), (75, 75). In any case, the energy required for transmission of a data unit from a sensor i to the sink s is given by eTis ¼ dð2i;sÞ . The LP models are solved for a given set of parameters using LINGO (LINDO Systems 2019). Numerical results for various networks with different sizes and sink node placements are shown in Figs. 6.11 and 6.12. The performance of the proposed model was evaluated thru simulation, leaving the routing problem outside the LP formulation. The main performance metrics of interest are the duration of network operation before a sensor node drains out of battery energy for the first time and the sink sojourn time. The presented approach has always achieved higher network lifetime, leading to a lifetime up to more than twice that obtained with other models as the network size increases; it also results in a fair balancing of the energy depletion among the sensor nodes. The LP formulation also revealed the necessity to develop new heuristic algorithms that take into account at the same time, both optimization problems: the scheduling and the routing. Such algorithms can be used online in an adaptive and distributed environment where the sink sojourn times are not determined a priori. It is noted that the optimal lifetime provided by the theoretical analysis of the model can be used as a performance measure in order to test the efficiency of other heuristics that might be proposed in the future. The proposed formulation can also be used as a starting point where new algorithms can be implemented.

Average network lifetime

6.2 Mobility-Based Approach Taxonomy

421

900 800 700 600 500 400 300 200 100

0

20

40

60

80

100

Number of sensors

Average network lifetime

(a) For sink locations (25,25), (25,75), (75,25), (75,75) 900 800 700 600 500 400 300 200 100 0

20

40

60

80

100

Number of sensors (b) For sink locations (0,0), (0,100), (100,0), (100,100), (50,50) Legend: The symbol on top of each bar represents the corresponding standard deviation.

Fig. 6.11 Average lifetime for various networks sizes (Papadimitriou and Georgiadis 2006)

6.2.1.2

Controlled Sink Mobility Protocols

Controlled Sink Mobility for Prolonging WSNs Lifetime (GMRE) This study presents the advantages of using controlled mobility in WSNs for increasing their lifetime; that is, the period of time where the network is able to provide its intended functionalities (Basagni et al. 2008). More specifically, for WSNs that comprise a large number of statically placed sensor nodes transmitting data to a collection point (the sink), it is shown that by controlling the sink movements, remarkable lifetime improvements are obtainable. Out of the concern to prolong the lifetime of WSNs, the mobility of the network sink was exploited so that, by sojourning in the vicinity of different sensors, energy consumption is more uniformly distributed throughout the nodes; as a consequence, nodes lifetime and network lifetime are increased. Three schemes that represent different solutions for sink mobility were introduced:

6 Energy Management Techniques for WSNs (3) …

422

Average sink sojourn times

300 250 200 150 100 50 0

20

60 SPR

100

20

60 100 MSPR

20

60 LP-opt

100

(a) For sink locations (25,25), (25,75), (75,25), (75,75) Average sink sojourn times

600 500 400 300 200 100 0

20

60 SPR

100

20

60 100 MSPR

20

60 LP-opt

100

(b) For sink locations (0,0), (0,100), (100,0), (100,100), (50,50) Legend: • The symbol on top of each bar represents the corresponding standard deviation. • Results are obtained for networks with 20,60,100 sensor nodes. • The lifetime achieved by Static Sink corresponds only to one location where the sink node stays for whole network operation. • In Figure 6.12a the time spent is almost the same for every location. • In Figure 6.12b the sink node stays most of the time at location (50,50).

Fig. 6.12 Average sink sojourn times for various network sizes (Papadimitriou and Georgiadis 2006)

• Mathematically presenting a new mixed integer linear programming (MILP) model that determines sink routes and sojourn times at the sink sites (specific locations the sink can visit). This scheme, termed OPT, computes optimal sink routes and sojourn times based on a MILP formulation that considers realistic parameters of wireless sensor networking and sink mobility; it achieves the best possible network lifetime by deciding sink movements based on nodal transmission costs in a centralized way that is typical of LP models. Differently from previous solutions, parameters and constraints are included to model realistic requirements of a WSN. For instance, considering the cost of moving the sink

6.2 Mobility-Based Approach Taxonomy

423

from a site to another, both from a data latency point of view, and from an energy consumption point of view. From a data latency point of view, there is the effect of buffering data packets during the sink movement. While from an energy consumption point of view, there is the cost of building and releasing data routes from the sensors to the current position of the sink. Moreover, constraints are introduced for considering the mobility rate of the sink, and imposing a minimum sojourn time for the sink at the different sites. This allows investigating how lower or higher sink mobility affects network lifetime. Although data routing optimization could be easily incorporated in the MILP model, it is elected to have it routing-independent. Primarily, since a centralized solution for routing is not viable for WSNs. Moreover, deriving routing as part of the model solution optimizes it only with respect to the sole metric of network lifetime, whereas there are other relevant metrics to be considered when designing efficient routing for WSNs, such as packet throughput, data latency, control overhead. ILP-based solutions are notoriously hard to compute, and often these models can be solved only for particularly small input scenarios (Garey and Johnson 1979). However, the simplicity of the proposed model for sink mobility makes it possible to obtain optimal sink routes for non-trivial, quite realistic cases, such as networks with hundreds of sensor nodes and several dozens of sink sites. This allows using the MILP-generated optimum sink mobility as an upper bound for more suitable distributed and localized solutions to be devised and deployed. MILP models provide centralized solutions. For example, in order to find lifetime-optimal routes and sojourn times for the mobile sink, one has to provide a global view of network topology, communication costs, etc. Centralized solutions, however, are highly time-consuming and energy-consuming for most WSN applications. • The development of decentralized, simple heuristic, greedy maximum residual energy (GMRE), a completely distributed and localized protocol for sink mobility. Throughout the network lifetime, the sink moves as drawn by sink sites in energy-rich areas of the network. More specifically, the sink keeps monitoring surrounding sites with respect to the energy of the nodes around them. When a site, different from the current, is in area with higher energy, the sink greedily moves at that new site. This simple heuristic takes explicitly into account crucial parameters, such as the costs of data route release and establishment when the sink moves, the different sink mobility rates, as well as possible constraints on sink mobility. Moreover, it is completely adaptive to diverse data routing protocols. • For sake of benchmarking, random movement (RM) heuristic was introduced, a simple distributed scheme according to which the sink moves randomly, hence uncontrollably, among the nodes.

424

6 Energy Management Techniques for WSNs (3) …

Starting with the problem definition, a scenario is considered where a large number jN j of resource constrained, static sensor nodes with sensing and wireless communication capabilities are scattered in a given geographic area. A periodic generation of data packets is assumed at the sensor nodes. Node i 2 N transmits at a given data rate, ri , packets that are “converge-casted,” to the sink for processing. While the sensor nodes are static, the sink can be mobile. More specifically, there is a set S ¼ f1; . . .; qg of q sink sites, which are the points the sink can visit within the geographic area. A typical scenario is shown in Fig. 6.13a. Because of the “sink neighborhood problem” (Sect. 6.2.1), the sink moves throughout the network in an attempt to balance the energy consumption among the nodes. Every time the sink reaches a new site, it floods a packet f to all the network nodes making them aware of its current site. A node that receives f starts sending/ relaying its packets toward the new site of the sink. Every routing scheme that works with the topological information provided by f, such as geographic path or shortest path-based routing, is a viable routing for data delivery to the sink. Noticeably, the independence from the particular routing protocol yields a twofold advantage: • Guaranteeing the longest possible network lifetime given the specific routing. • Allowing the network users to design or choose the routing algorithm that best meets the WSN application requirements in terms of different metrics of interest, not just the lifetime. Every time the sink leaves a site, it again floods a packet to all nodes, informing that it is no longer reachable at that site. Upon receiving this second packet, a node stops forwarding data while remaining packets are buffered and waits to receive a new packet f from the sink, carrying its whereabouts. When the new packet f arrives and routes to the new site of the sink are formed, packet transmission is resumed. There is virtually no bound on how far the sink can travel between two sites. However, while the sink is traveling, the sensors do not transmit. Therefore, if new data are sensed, they are buffered. This implies the possibility of high delays for data packets. In order to contain this delay, a new parameter dmax is introduced to represent an upper bound on the distance that the sink can travel from a site to the following one. Thus, the pair ðS; dmax Þ uniquely defines a graph of sink sites where there is a link between two sites if and only if their Euclidean distance is  dmax . Figure 6.13b illustrates an interpretive scenario. In case of high sink mobility and low data traffic, it is observed that the energy cost for route construction and release can be significant; therefore, this cost is explicitly taken into account. In order to evaluate the impact of different, higher or lower, sink mobility rates, the parameter tmin is introduced to represent a mandatory minimum time the sink has to sojourn at a site.

6.2 Mobility-Based Approach Taxonomy

425

Legend: • 32 nodes, represented by circles, are scattered randomly and uniformly on a bi-dimensional area. • 16 sink sites, represented by squares, are organized according to a 4×4 grid. • A solid link between two nodes indicates that those two nodes are neighbors, i.e., they can hear each other transmissions. • A dotted link between two sites indicates that the sink can move from one site to the other and vice-versa.

(a) Sensor nodes and sink sites

Legend: • The sink (the triangle) can reach, from its current position, three sites (darker squares). • The dotted lines between the sites indicate that the sink can only move horizontally or vertically in the 4×4 grid. • Routes from selected sensors to the current site of the sink are shown.

(b) Sink movements and routes to the sink Fig. 6.13 Typical WSN scenarios (Basagni et al. 2008)

6 Energy Management Techniques for WSNs (3) …

426

With regard to the problem definition, the following problem is solved via mathematical modeling and by designing distributed protocols: Determine the starting site and the route for the mobile sink over the graph ðS; dmax Þ, together with the sojourn times tk  tmin of the sink at each visited site k 2 S so that network lifetime is maximized. A mixed integer linear programming (MILP) formulation of the problem is developed: Max

X

ð6:30Þ

tk subject to

k2S

X

cik  tk þ

k2S

X

fik  yk  e0

ði 2 N Þ

ð6:31Þ

k2S

tmin  yk  tk  M  yk X x0k ¼ 1

ð k 2 SÞ

ð6:32Þ ð6:33Þ

k2S

X

xk;q þ 1 ¼ 1

ð6:34Þ

k2S

X

X

xjk ¼

j 2 S [ f0g ðj; kÞ 2 O [ A

xkj

ð k 2 SÞ

ð6:35Þ

j 2 S [ fq þ 1g ðk; jÞ 2 A [ D

X

xjk ¼ yk

ð k 2 SÞ

ð6:36Þ

ððj; k Þ 2 AÞ

ð6:37Þ

j 2 S [ f 0g ðj; kÞ 2 O [ A uj  uk þ q  xjk  q  1 tk ; uk  0 yk 2 f0; 1g xjk 2 f0; 1g

ð k 2 SÞ

ð6:38Þ

ðk 2 S Þ

ð6:39Þ

ððj; k Þ 2 X Þ

ð6:40Þ

where, A O

is the set of directed edges joining sink sites whose distance is less than or equal to dmax , i.e., A ¼ fðj; kÞ 2 S  S: j 6¼ k; djk  dmax , is the set of directed edges ð0; k Þ; k 2 S, joining a fictitious site 0 (origin) with the sites in S,

6.2 Mobility-Based Approach Taxonomy

D X

427

is the set of directed edges ðk; q þ 1Þ; k 2 S, joining the sites in S with a fictitious site q þ 1 (final destination), is the union of A, O, and D.

The MILP description and more variables definitions are below detailed: • The objective function in Eq. 6.30 maximizes the sink total time at sojourning P sites, k2S tk , which is the effective lifetime. • The constraint P Eq. 6.31 states that the combined energy spent at node i for data delivery, k2S cik  tk , and for data route consumption and release, P f  y , during the time before depletion of the first node, should node ik k k2S exceed the node initial energy, e0 . The variable cik represents the power consumption (Watt) for receiving and transmitting packets at node i 2 N when the sink sojourns at site k 2 S, and fik is the energy consumption (Joules) at node i 2 N for setting-up/releasing routes when the sink moves to site k 2 S. • The right part, of the double inequality in Eq. 6.32, forces yk to take the value 1 if the sink sojourns at site k (tk [ 0), thus linking the binary variable yk (the constraint Eq. 6.39) with the continuous variable tk ; M is a significantly large number. The left part of the double inequality in Eq. 6.32 restricts the sojourn time tk to be at least equal to the mandatory minimum sojourn time tmin if the sink sojourns at site k (yk ¼ 1), and at the same time forces yk to take the value 0 if the sink does not sojourn at site k (tk ¼ 0). The first sojourning site in the sink movement route is allowed to be any site in S. For sake of implementation, a fictitious fixed initial site 0 (origin) is introduced. At the beginning of the WSN lifetime, the sink moves in zero time and cost from the origin to some site a 2 S, determined by the model. • As revealed in Eq. 6.33, this is such that particular site x0a ¼ 1, namely it is the optimum starting point of the sink journey. Then, the sink sojourns at that first site and at subsequent other sites in S to be determined by the model. Finally, in Eq. 6.34, from the last sojourning site x, the sink moves again in zero time to a second fictitious site q + 1. The site x completes the sink route started at site a. This is the last site at which the sink sojourns and indicates the end of the WSN lifetime. The arcs ðj; k Þ 2 X on the sink route are associated with binary variables xjk equal to 1. The variable xjk is equal to 0 for all ðj; kÞ 2 X that do no belong to the route. Equivalently, one can think of a unit of flow moving from the origin to the destination. • The constraint Eq. 6.33 induces a unit flow from the origin to some node a 2 S, while the constraint Eq. 6.34 causes the destination to absorb a unit flow coming from some node k 2 S; it induces a unit of flow from the last node in the sink route x to the fictitious final node q þ 1. The constraint Eq. 6.35 forces flow conservation at all sites k 2 S, thus ensuring the generation of a route. • The constraint Eq. 6.36 ensures that the sites k 2 S on the generated route are sites at which the sink sojourns (kjyk ¼ 1). Specifically, if yk in Eq. 6.36 equals 1, then the sink sojourns at site k; therefore, there must be one and only one arc

428

6 Energy Management Techniques for WSNs (3) …

on the sink movement route reaching site k. On the other hand, if yk equals 0, then there will not be any incoming arc to that site. Applying the proposed MILP, Fig. 6.14 shows a possible optimum sink route that goes from the initial site a ¼ 2 to the final site x ¼ 24. The run of events can be illustrated as: • The constraints in Eqs. 6.33 and 6.34 ensure that, independently of a and x, respectively, there is only one initial site and one final site for the sink route, since the corresponding arcs (x0a and xxq þ 1 ) must be 1. • The combination of constraints Eqs. 6.35 and 6.36 takes care of generating the route between a and x that pass through all the sites where the sink has to sojourn for maximizing the network lifetime. In particular, the constraint Eq. 6.35 mandates that there must be one outgoing arc xkl for every incoming one xjk (with the ordinary exception of the two fictitious sites 0 and q þ 1). For instance, this is the case of arcs x34 and x49 in Fig. 6.14a, which are both set to 1. The remaining arc out of site 4, namely arc x45 , is forced to be 0. According to the constraint Eq. 6.36, for every site k where the sink sojourns (yk ¼ 1) there must be a way to get there, i.e., there must be exactly one site j, which includes the fictitious site 0, from which k is reachable (xjk ¼ 1). At the same time, the sink should not pass through sites where it does not sojourn. For instance, the sites with k = 2, 3, 4, 9, 10, 15, 20, 25, and 24 in Fig. 6.14a are all, and only those for which yk ¼ 1, i.e., these are all and only those sites that can be in the sink route. All other sites h are such that yh ¼ 0. • Flow conservation constraints in Eqs. 6.35 and 6.36 do not prevent the formation of cycles disjoint from the route from the origin to the destination. The disjoint non-simple route, depicted in Fig. 6.14b, comprising nodes 2, 3, 4, 9, 10, 15, 20, 25, and 24 and nodes 16, 17, 22, and 21 (cycle) is possible, according to Eq. 6.30 up to Eq. 6.36, since none of these constraints is violated by having y16 ¼ y17 ¼ y21 ¼ y22 ¼ 1 as well as x16;17 ¼ x17;22 ¼ x22;21 ¼ x21;16 ¼ 1. As quite unrealistic, this situation is undesirable. It is practically impossible, for instance, to have the sink moving from site 9 (a site in the connected route from site 2 to site 24) to site 17 (a site in the cycle); sites 9 and 17 are not directly connected, i.e., their distance is  dmax . The constraint Eq. 6.37 ensures that no such cycles are formed (Miller et al. 1960). According to Eq. 6.38, a site k is associated with a weight uk  0. The constraint Eq. 6.37 imposes that the sites visited by the sink are traversed in increasing order, i.e., if xjk ¼ 1, then uj \uk . This renders the impossibility to return to the same node and hence to form cycles like the one in Fig. 6.14b. The parameter tmin has been introduced to assess the effect of different (higher or lower) sink mobility rates on network performance. For a given tmin , the model will produce the sink route and sojourn time tk  tmin at site k that maximizes network lifetime. By varying tmin , a number of tradeoffs can be explored. For instance, at higher tmin lower overhead is expected (e.g., for route construction and release). Shorter tmin results in better choices of sojourn times at different sites, which might

6.2 Mobility-Based Approach Taxonomy

429

Fig. 6.14 Sink optimum routes obtained by constraints Eqs. 6.33–6.36 (Basagni et al. 2008)

x24,q+1 x25,24 x20,25

x15,20

x10,15 x9,10 x4,9 x2,3

x3,4

x0,2

(a) A sink route

x24,q+1

x22,21

x25,24

x17,22

x20,25

x21,16 x16,17

x15,20

x10,15 x9,10 x4,9 x2,3

x3,4

x0,2

(b) A disjoint cycle in the sink route

430

6 Energy Management Techniques for WSNs (3) …

result in a longer network lifetime, at the price of increasing overhead. Depending on prevailing network conditions, there is a value of tmin for which the advantage of a finer granularity of sojourn times is out powered by the energy consumption increase due to the extra overhead. After the mathematical problem formulation as an MILP, two distributed protocols for sink mobility, GMRE and RM, are proposed and compared with the optimal routing strategy, OPT, provided by the MILP model. The two protocols differ on the strategy used by the sink sojourning at a site for choosing the next one: • In the greedy maximum residual energy (GMRE) protocol the sink greedily selects the site, within dmax , surrounded by nodes that have the most energy left. The idea is that, with time this should most likely result into a balanced energy consumption throughout the network, and hence into a longer network lifetime. After spending a time tmin at a site, a sink evaluates whether to move toward one of the adjacent sites or to stay where it is. Two sites are adjacent if their distance is  dmax . In order to decide whether to move or not, the sink gathers information about the residual energy at the nodes around each of the potential future sites and compares it with the residual energy at the current site. If there are adjacent sites with a residual energy higher than that at the current site, the sink moves to the site with the highest residual energy (selecting randomly among sites with the same residual energies in case of ties); otherwise, the sink stays at the current location. Communicating, to the sink, the residual energies at the adjacent sites are a key to the definition and implementation of GMRE. This communication proceeds in two phases: – In the first phase, for each of the adjacent sites, the sink identifies one sentinel sensor node that will be in charge of measuring and reporting the residual energy at that site, when requested by the sink. To implement this phase, advantage is made of the flooding performed by the sink when it makes the nodes aware of its new location. – The second phase concerns the sink interrogation of the sentinels. This is performed whenever the sink has to decide whether to move or not. At this time, the sink interrogates the selected sentinels about the residual energy at their sites. This is accomplished by sending a small packet to the sentinels. When interrogated, the sentinels query their neighboring sensor nodes about their residual energy and communicate back to the sink the minimum of the obtained values, or any suitable function that can express how critical for the network lifetime is to place the sink in that area. For this heuristic protocol, it is assumed that a node is aware of being in the “transmission vicinity” of a site (its Euclidean distance from a site is less than or equal to the nodes transmission range). This can be obtained by endowing the nodes with a suitable localization mechanism (Savvides and Srivastava 2004). The flooding message contains the coordinates of the current location of the sink. Upon receiving the flooding packet, a node

6.2 Mobility-Based Approach Taxonomy

431

knows if it is in the vicinity of a possible future sink site. In this case, it sends to the sink a (small) packet for its candidacy, as sentinel. Upon receiving such packets, the sink decides which is the sentinel for a given site. This mechanism also allows the sink to identify those sites that are isolated (no packet is received from nodes around that site). In this case, the sink will not consider that site as possible in the future. • The second proposed simple protocol for sink mobility captures the case of uncontrolled, random mobility of the sink, the random movement heuristic (RM). Every tmin the sink selects randomly and uniformly the new location among all the sites within distance dmax from the current. In case a site different from the current one is selected, the sink moves to that site. RM generalizes data gathering protocols previously proposed in the literature, e.g., the data mules approach (Jain et al. 2006) as detailed in Sect. 6.2.2.1, to the case of multihop data routing; it is used mainly as a benchmark for assessing the effectiveness of GMRE in prolonging network lifetime. Performance evaluation, via the ns-2 simulator, involved the previously presented protocols as follows: • The sink is static. This is a degenerate mobility scheme, as the sink does not move. In this case, named STATIC, the sink is placed at the geographical center of the deployment area, which is the position that maximizes the network lifetime. • OPT mobility where the sink moves along the optimum route derived by the MILP model. • The sink moves according to the RM heuristic. • The sink moves as specified by the GMRE heuristic. The performances of the four schemes have been compared with respect to several metrics: • Network lifetime. The time until the first node fully depletes its energy. • Per node residual energy over time. Investigating this metric allows to determine the actual energy consumption associated with the different mobility schemes as well as to verify how balanced is the consumption itself. • End-to-end packet latency. This is the time that goes from packet generation at a sensor node to the successful delivery of that packet at the sink. • Overhead (bps). The overhead incurred by a protocol is defined as the number of bits/s sent on average by each node for route maintenance (building and releasing routes when the sink moves), and for gathering information needed by the sink for deciding whether to move or not and where. • Delivery ratio. The percentage of packets generated at the sensor nodes that are successfully delivered to the sink. In the low traffic scenarios considered here, all packets are successfully transmitted.

432

6 Energy Management Techniques for WSNs (3) …

• Sink sojourn times at the different sites. This metric captures the time spent by the sink on average at the different sites and is key to correlate sink behavior with network performance (overhead, latency, and energy consumption). • The mobility pattern followed by the sink for a given sink mobility scheme over different runs to identify common patterns and obtain an in-depth understanding of the rationale behind the sink movements. Simulation has been run according to three scenarios. The outcomes presented are those obtained out of a scenario in which n = 400 sensor nodes, generating data at the constant rate of 0.5 bps (a packet is injected into the network around every 13 min), are deployed in a grid-like topology over a square area of side L = 400 m. The sensor nodes transmission range R is fixed at 25 m. This means that all nodes, which are not on the perimeter of the area, have a “cross-like visibility” of their neighbors (they have four neighbors). A single sink moves through jSj possible sites. Sink sites are arranged into a 4  4; 6  6; 8  8 grid. Data are delivered to the sink according to the routing protocol presented in Basagni et al. (2004). The route construction process is initiated by the sink. The sink floods a packet via which the nodes calculate their hop distance from the sink. Forwarding is based on a simple and unchanging information; typically, a node that is i hops away from the sink will send data packets to one of its neighbors whose distance is i − 1. The specific neighbors can be different each time, and one is chosen among the neighbors randomly and uniformly. Channel capacity and MAC settings are typical of sensor networking (250 Kbps and CSMA/CA, respectively), and the sensor nodes are assumed alike. Sensor nodes initial energy is set to 50 J. The energy model used in the experiments follows the specifications of the TR1000 radio transceiver (Murata Electronics 2015); the energy consumptions corresponding to transmission, reception, and sleep mode are 14.8, 12.5 and 0.016 mW, respectively. In this basic scenario, the model parameter dmax has been set to 190 m. The set of sites the sink can move to from its current position depends on dmax and on the cardinality of S. Figure 6.15 shows the set of adjacent sink sites when the sink can select where to sojourn among 4  4 and 8  8 sites. The first circle encloses adjacent sink sites when dmax ¼ 190 m. This set of experiments aims at demonstrating the bare effectiveness of the proposed MILP and heuristic solutions for extending network lifetime with respect to the static case. Figure 6.16 shows the WSN lifetime in networks with different numbers of sink sites and for values of tmin that vary between 5 * 104 s and 1 * 106 s. Each figure compares the lifetime obtained by OPT, GMRE, RM, and STATIC. The lifetime in the static case is equal to 7,013,801 s, independently of tmin . This is the time when one of the four sink neighbors, i.e., the nodes that relay all the network data to the sink, fades out because of energy depletion. The other three nodes remain with negligible amount of energy, fading out immediately after the first one. This leaves the sink isolated from the rest of the network. Network lifetime improvement due to sink mobility comes at the price of increased data latency. Two reasons provide the justification:

6.2 Mobility-Based Approach Taxonomy

(a)

case

433

(b)

case

Legend: The sink is indicated as a triangle.

Fig. 6.15 Neighboring sink sites (Basagni et al. 2008)

• The packets that are newly generated while the sink is moving and those in transit toward the sink have to wait until routes to the new position of the sink are established. • In order to balance energy depletion, the sink will spend time not only at the center of the deployment area but also along borders. This imposes longer average routes and hence higher packet latency than the latency experienced when the sink is statically placed at the center. This is actually the dominant reason for increased latency in low sink mobility scenarios. It is now possible to fully understand the latency performance of the different schemes. Figure 6.17 depicts the average latency per packet in OPT, GMRE, RM, and STATIC when the number of sink sites varies from 16 to 64. The outcomes may be digested in several bullets: • When the sink sojourns at perimeter or at corner sites, which is typical of GMRE and OPT, the lifetime increases. However, these are also the cases when the average length of the routes to the sink increases, which implies, in turn, higher packet latency. It is thus reasonable to expect that lower latencies might be experienced when the sink is statically placed at the center of the sensor deployment area. • The RM heuristic, which tends to move the sink to sites located centrally, is the first best mobility scheme in terms of latency. The increase of RM-induced latency with respect to STATIC is expectedly lower when the number of sink sites increases; since in this case, central sites are closer to the geographical center of the deployment area. This increase never tops 40%. As tmin increases,

6 Energy Management Techniques for WSNs (3) …

434 3.5*107

Network lifetime (sec)

3*107 2.5*107 2*107 1.5*107 1*107 5*106 0 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(a)

sink sites

3.5*107

Network lifetime (sec)

3*107 2.5*107 2*107 1.5*107 1*107 5*106 0 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(b)

sink sites

Fig. 6.16 Average network lifetime versus tmin (Basagni et al. 2008)

the sink tends to stay less at central sites, leading to higher average latencies experienced by RM packets. • The opposite trend is observed for GMRE. For small tmin , the sink stays at sites on the corners and on the perimeter, which leads to average latencies up to 30% higher than those experienced in the RM case. When tmin increases, the sink sojourns less at corner sites, which implies a decrease in the average latency.

6.2 Mobility-Based Approach Taxonomy

435

0.34

Latency per packet (sec)

0.32 0.3 0.28 0.26 0.24 0.22 0.2 0.18 0.16 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(a)

sink sites

0.34

Latency per packet (sec)

0.32 0.3 0.28 0.26 0.24 0.22 0.2 0.18 0.16 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(b)

sink sites

Fig. 6.17 Average data latency versus tmin (Basagni et al. 2008)

• OPT sink mobility is not significantly affected by varying tmin in the selected range, and the latency values are similar to those observed for GMRE. Also, the overhead of the considered mobility schemes was investigated. The overhead is the cost incurred by the protocols for route management, i.e., for route setup and release when the sink changes site, as well as the cost required for gathering information about residual energy at adjacent sink sites. For OPT, it is ideally assumed that the needed input, cik and fik in Eq. 6.31, to the MILP

6 Energy Management Techniques for WSNs (3) …

436 0.08

Overhead per node (bps)

0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(a) GMRE, OPT and STATIC 0.08

Overhead per node (bps)

0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0

1*105 2*105 3*105 4*105 5*105 6*105 7*105 8*105 9*105 1*106

tmin (sec)

(b) RM, OPT and STATIC

Fig. 6.18 Average overhead per node (Basagni et al. 2008)

formulation is known. Therefore, the OPT overhead is associated with route maintenance when the sink follows the mobility pattern output by the model. Figure 6.18 shows the average number of control bits that each node transmits per second (bps) according to OPT, GMRE, RM, and STATIC. In particular, Fig. 6.18a shows the overhead incurred by GMRE for varying number of sink sites and tmin , and the overhead imposed by OPT and STATIC. Figure 6.18b depicts the same metric for the RM heuristic and compares it with OPT and STATIC. Both OPT and STATIC impose negligible overhead (barely visible in the figure). When the sink is kept static, routes need to be computed only once. In the case of OPT mobility, the sink moves from one site i to the next one j only when the whole sojourn time ti at site i has passed, resulting in a reduced number of movements. According to the GMRE and RM heuristics, instead, the

6.2 Mobility-Based Approach Taxonomy

437

sink makes a decision on whether to move or not every tmin . As expected, the higher the tmin , the fewer the movements, and the lower the overhead per second. In all considered scenarios, it was observed that moving the sink always increases network lifetime. In particular, controlling the mobility of the sink leads to remarkable improvements, which are as high as sixfold compared to having the sink statically (and optimally) placed, and as high as twofold compared to uncontrolled mobility.

Maximizing the Lifetime of WSNs with Mobile Sink in Delay-Tolerant Applications (DT-MSM) This delay-tolerant mobile sink model (DT-MSM) is suggested to maximize the lifetime of a WSN by taking advantage of sink mobility while giving care to the case where the underlying applications tolerate delayed information delivery to the sink (Yun and Xia 2010). One of the application examples is battlefield surveillance with sensor nodes deployed to monitor the movement of enemy vehicles or troops. A mobile sink attached to an unmanned aerial vehicle flies over the monitored region regularly to harvest the collected intelligence. To avoid being intercepted or detected by enemy forces, the mobile sink needs to operate in only a few safe locations within a limited operation time. Another example is habitat monitoring where a mobile robot is used to collect information from the sensor nodes in the field. If much of the habitat area is not accessible by the robot or if it is desirable to minimize disturbance to the targeted animal species, the mobile robot will trace predetermined paths and stop by a set of pre-arranged locations regularly for data collection. Within a prescribed delay tolerance level, each node does not need to send the data immediately as it becomes available. Instead, the node can store the data temporarily and transmit it when the mobile sink is at the location most favorable for achieving the longest network lifetime. To find the best solution within the proposed framework, the optimization problems are formulated to maximize the lifetime of the WSN subject to constraints on the delay bound, node energy, and flow conservation. The proposed framework is compared with other lifetime-maximization proposals. Computational experiments have shown that the proposal increases the lifetime significantly when compared to the stationary and mobile sink models without delay tolerance. It integrates multiple energy-saving techniques: explicitly, multipath routing, mobile sink, delayed data delivery and active region control, into a single optimization problem. Such sophistication comes though at a cost. Whether the proposal should be adopted in practice depends on the tradeoff between the lifetime gain and the actual system cost. The system cost includes all costs/complexity in implementing the framework and in actual operations. These may include extra communication protocols for coordination and control, e.g., new routing and rate control protocols, extra memory for keeping delayed data and memory management costs, and application-level costs incurred by delayed

438

6 Energy Management Techniques for WSNs (3) …

information delivery. Even if the decision is not to adopt it due to excessive cost or high complexity, the framework is still useful because it can supply the practitioners with a performance benchmark, e.g., how much lifetime improvement opportunity is possible. There can be many variants of the problem formulation, some of which can be highly tough, often involving NP-hard combinatorial subproblems. Two models are to be recalled as a matter of differentiation: • The model involving mobile agents that move around and collect data from nearby sensor nodes on behalf of the immobile sink (Shah et al. 2003; Jain et al. 2006). In this model, mobile agents forward the collected data to the sink when they move to its vicinity. Differently from the proposed multihop communication framework, communication occurs only from the sensor nodes to the mobile agents or from the mobile agents to the sink via a single hop; the sensor nodes do not relay traffic. Mobile agents are assumed to have plenty of energy, and the movement of each is modeled as a random walk. Also, queues in the mobile agents and sensor nodes are finite, and the delay of the collected data is bounded. • The LP formulation, detailed in Section “Exploiting Sink Mobility for Maximizing Sensor Networks Lifetime,” determines how to move the mobile sink and how long to park it at each stop along its path so as to maximize the lifetime of the WSN (Wang et al. 2005). Dataflows are not decision variables of the lifetime optimization problem; while in the model presented in this section, the sink sojourns at different sink stops and the routing scheme are decision variables. The analysis and experiments were conducted under a simple structured network topology where the sensor nodes are deployed in a grid-like pattern. In order to maximize the lifetime of WSNs with mobile sink in applications that can tolerate a certain amount of delay, the DT-MSM is now to be introduced. In this setting, each node can postpone the transmission of data until the sink is at the stop most favorable for extending the network lifetime. Thus, the nodes can collectively achieve a longer network lifetime. In contrast, the static sink model (SSM) and the mobile sink model (MSM) do not exploit such possibility. Let D be the maximum tolerable delay, or the delay tolerance level. It is assumed that the sink finishes one round of visit to all the stops (where the sink stays for a positive duration to collect data) in D time units and then repeats with another round again and again. Note that two consecutive visits to the same stop take a time D. As an example that shows how the proposed framework can perform better, consider the two nodes shown in Fig. 6.19. Ignoring the receiving energy requirement and supposing that the transmission energy per unit of data is equal to the square of the distance between the sender and the receiver, it is shown that: • Both nodes N1 and N2 generate data at 1 bps and have initially 100 units of energy. • If the sink is located at O in the SSM, both nodes spend 4 units of energy for sending a bit of data. It is obvious that the optimal lifetime is 25 s.

6.2 Mobility-Based Approach Taxonomy

1

1

439

1

1

Legend: and are two sensor nodes, O is the sink location in SSM, the mobile sink.

and

are the candidate stops of

Fig. 6.19 SSM, MSM, and DT-MSM instances (Yun and Xia 2010)

• In the MSM with sink locations {L1 , L2 }, due to the symmetry of the structure, the sink stays at both L1 and L2 for the same amount of time to achieve the maximum lifetime. Each node spends 1 or 9 units of energy for sending 1 bit of data depending on whether the sink is at L1 or L2 . The average energy consumption per bit is 5 units. Thus, the lifetime is 20 s. • In the DT-MSM, assuming that the sink alternates between the two stops and stays for 1 s at each stop in each cycle. Hence, this is the case where D = 2 s. When the sink stays at L1 , only N1 sends 2 bits of data to the sink; when the sink moves to L2 , only N2 transmits 2 bits of data (N2 keeps its data while the sink is at L1 ). Both nodes spend 2 units of energy every 2 s or 1 unit of energy per second on average. Thus, the lifetime is 100 s, a significant increase compared to the SSM and MSM. This is because, in the DT-MSM, the nodes do not always participate in the communication for all the sink stops and each waits until the sink location is most favorable for energy saving, and then sends data at the higher rate. Recalling that the traffic rate is assumed sufficiently small compared to the capacity of the wireless link; hence, sending data at a higher rate does not alter the energy consumption per bit. Unlike the MSM or SSM, the sink in the DT-MSM can collect data from only a subset of the set of all sensor nodes, N , at each stop. Let Rl be the subset of N such that only nodes in Rl can participate in the communication toward the sink when the sink is at l 2 L. As Rl is the coverage of the sink location l, the union of Rl over l 2 L must be the set of all sensor nodes, N . In other words, any sensor node should be covered by at least one sink location. When node i is in Rl , it is said to be active at l 2 L. A simple method of constructing Rl is considered. Letting the radius of coverage of the sink be r, for each l 2 L, if d ði; lÞ  r, where i 2 N , then i 2 Rl . The radius of coverage of the sink, r, should be large enough so that every sensor node belongs to at least one Rl . To be noticed, the minimum r depends on the locations of the sink stops. In both SSM and MSM, the sink collects data from each node i at the same rate at which node i generates the data. However, in the DT-MSM, the data transmission rate at node i during the collection time is no longer the same as the constant data generation rate di . When node i is not active, i.e., not covered by the current sink location, it continues to gather data and should store the newly generated data. Hence, data buffering is required for the proposed framework. Within a cycle of

6 Energy Management Techniques for WSNs (3) …

440

D time units, the total stored data at each node i is at most D  di . For ease of presentation, the sink is assumed to visit all locations in L in the order 1 ! 2 ! jLj ! 1 . The sink may stay at some location for zero time. With slight alteration of terminology, the network lifetime T is defined to be the number of cycles made by the sink until the first node fades out due to energy exhaustion. The actual lifetime is T  D. Once traffic is allowed for buffering, there are different strategies on whose traffic is buffered. Which strategy gets adopted in practice probably depends on the application, added practical concerns, and the designer preference. Since these factors are not known in advance, two strategies, or two variants of the model are to be described, namely the subflow-based model and the queue-based model. The selections lead to different performance/complexity tradeoffs: • In the subflow-based model, the nodes in the current coverage Rl are not allowed to buffer the relayed traffic from other nodes; as soon as a node in Rl receives the data from other nodes, it immediately forwards the data to its neighboring nodes. To model this constraint at each node i, it is required to differentiate the data generated by node i from the data originally generated by other nodes but ðc;lÞ forwarded to node i. Again, let xij be the rate assignment from node i to node j, ðlÞ

while the sink is at l, for the traffic generated by node c (commodity c). Let xij be the aggregated rate of traffic that needs to be forwarded to node j from node i when the sink is at l. That is ðlÞ

xij ¼

X

ðc;lÞ

xij ;

8l 2 L; 8i; 8j 2 Nl ðiÞ

ð6:41Þ

c2Rl

Defining Nl ðiÞ to be the set of the downstream neighbors of node i that are in the coverage Rl : Nl ðiÞ ¼ Rl \ N ði; lÞ

ð6:42Þ

where N ði; lÞ is the set of the downstream neighbors of node i that are in the coverage Rl :   j 6¼ ig N ði; lÞ ¼ fj 2 N [ flgd ði; jÞ  d; ð6:43Þ Since at node i 2 N , the commodity or subflow of other nodes c 2 Rl , c 6¼ i must be forwarded as soon as it has been received: X k:i2Nl ðk Þ

ðc;lÞ

xki

¼

X j2Nl ðiÞ

ðc;lÞ

xij

ð6:44Þ

6.2 Mobility-Based Approach Taxonomy

441

The flow conservation at node i is the same as in the MSM except that the ðlÞ amounts of traffic originating from node i, wi ; l 2 L; i 2 Rl , are decision variables: 0 zl  @

X

ðlÞ xij

X



j2Nl ðiÞ

1 ðlÞ xki A

ðlÞ

¼ wi

ð6:45Þ

k:i2Nl ðk Þ

The data buffered during the previous sink movement cycle must be cleared in the current cycle. This requirement can be written as: X

ðlÞ

w i ¼ D  di

ð6:46Þ

l:i2Rl

Due to Eq. 6.44, it may appear that there is a need for per commodity-based formulation of the problem. Similar to the SSM problem, there is also a simpler, equivalent, aggregate-traffic formulation, using only the aggregate arc flow ðlÞ variables xij , as given in Eq. 6.41:

0 zl  @

X

ðlÞ

xij 

j2Nl ðiÞ

8 jLj Vth1

SHDN

Vsensor Vambi

Hysteresis comparator

Vbattery Li-polymer battery

Current limit switch (battery Charges charger)

Legend: Vth, Vth1: Preset threshold voltages. Vcap: Terminal voltage of a reservoir capacitor. Vbattery: Battery voltage. Vambi: Short-circuit voltage. Vsensor: Sensor output signal. SHDN: Shutdown.

Fig. 7.33 Architecture of an energy harvesting subsystem in AmbiMax platform

AmbiMax board (32 mm × 24 mm):

(a) With super-capacitors

(b) Without super-capacitors

Fig. 7.34 AmbiMax board (Park and Chou 2006a)

own maximum power point. As shown in Fig. 7.32, RCs of different sources compose RCA. AmbiMax can be powered by the ambient sources or the battery. It is powered solely by the ambient sources when the RCA has a higher voltage at its terminal than a certain threshold. It draws power from the battery when the RCA terminal voltage drops below the threshold. Eventually, the RCA terminal voltage rises again. When the ambient power sources can generate more energy than needed to

552

7 Energy Harvesting Projects for WSNs

drive the system, they also charge the battery. In AmbiMax, all of the above activities are autonomously controlled by purely analog circuitry, without the use of a digital controller. AmbiMax is also expandable to additional harvesting sources, simply by connecting their reservoir capacitor to the RCA. This section focuses on the rules of operation for each subsystem. Energy Harvesting Subsystem As shown in Fig. 7.33, the EH subsystem consists of a set of energy harvesting units, and each unit includes an ambient power source, a PWM switching regulator, and MPPT circuitry. In the ensuing text, Voc, Vambi, and Vop denote, respectively, the open-circuit voltage, short-circuit voltage, and operating voltage at MPP of an ambient power source. Also, Vcap denotes the terminal voltage of a reservoir capacitor. Principles of Operation The energy harvesting subsystem entails several principles as elucidated below: • PWM switching regulator. Different from designs that connect the solar panel either directly to a super-capacitor or to a battery though a diode (Raghunathan et al. 2005; Jiang et al. 2005), the PWM switching regulator is put between the ambient power source and the reservoir capacitors (Fig. 7.33). This, for several reasons, comparatively improves the harvesting efficiency: – The PWM switching regulator prevents the super-capacitor from degrading the operation efficiency of the ambient power source. Such isolation enables energy harvesting to continue even when Voc < Vcap, a condition that stopped harvesting in Prometheus (Sect. 7.2.2) and Heliomote (Sect. 7.2.4) platforms. Consequently, AmbiMax can harvest energy at all times except when the battery is fully charged and the sensor node is turned off, so as not to consume energy. – The PWM switching regulator also functions as a diode to block the reverse current flow from the super-capacitor to the ambient power source, but without the overhead of the usual 0.7 V drop. • MPPT circuitry. An ordinary PWM switching regulator by itself cannot perform MPPT when it charges a capacitor (Simjee and Chou 2006). However, a switching regulator with a comparator can be used to perform MPPT. Functionally, the comparator controls the operation of the switching regulator by comparing Vambi, and an output signal of a sensing device (Vsensor), that can be generated based on the status of ambient power sources. Once Vambi drops below Vsensor–VHysterisis indicating that the ambient power source is out of the maximum power point, the comparator turns off the regulator. Also, when Vambi increases so that it is higher than Vsensor + VHysterisis, the comparator turns ON the regulator again.

7.2 Energy Harvesting Projects

553

(a) Power-voltage characteristics of ambient power sources and maximum power window

(b) Hysteresis band and regulator status

Fig. 7.35 MPPT using a switching regulator and hysteresis comparator (Park and Chou 2006a)

Figure 7.35 clarifies how MPPT works on the AmbiMax platform. Previous platforms used an MCU to sample the source status and generate the control signal, but such approach is not applicable if the MCU must sleep. AmbiMax uses sensing devices to autonomously monitor the status of ambient power sources without digital control. For solar panels, a light intensity sensor such as the resistive sensor S1087 (HAMAMATSU 2014) or the TSL250RD voltage output sensor (TAOS 2007b) may control the ON/OFF status of the switching regulator. The configurations of the comparator and sensing devices are described in Fig. 7.36. Both the sunlight intensity and the sensor’s output signal are almost linearly proportional to the operating voltage at MPP (Vop). Therefore, it is possible to exactly monitor the MPP of the power source by scaling both the short-circuit voltage (Vambi) and the output voltage of the sensor (Vsensor) using R1–R4. For other types of ambient power sources such as a wind generator and vibration-to-electricity converter, it is likely to choose an accelerometer (Silicon Designs 2016) and a frequency-to-voltage (FV) converter (Microchip 2007).

Fig. 7.36 Hysteresis comparator configuration of MPP tracker (Park and Chou 2006a)

Legend: SHDN: Shutdown voltage

Comparator configuration: (a) With resistive sensors (b) With voltage output sensors

554

7 Energy Harvesting Projects for WSNs

Energy Harvesting Subsystem Implementation The implementation spans the ambient power sources and the energy harvesting unit as below described: • Ambient power sources. The Solar World’s solar cell module (4-4.0-100) (SolarWorld Americas 2016) is adopted, and its maximum output voltage and current are 4.0 V and 100 mA, respectively. Also, chosen, the Jointiff wind generator (Jointiff Limited 2016) that generates up to 500 mW maximum at 2000 rpm. The ambient power sources are picked up based on their output power levels, while considering several design parameters of WSNs such as expected lifetime, power consumption level, and the type and characteristics of the regulator and battery. These parameters should match the availability of ambient energy at the deployment site. Based on the estimated local conditions, the solar panel and wind generator generate 400 mW at the peak and 200 mW for at least 6 h a day. • Energy harvesting unit. The energy harvesting unit implementation consists of a switching regulator and an MPP tracker. The switching regulator controlled by the MPP tracker charges the super-capacitors of the RCA. A boost switching regulator is chosen, the LTC3401 (Linear Technology 2001b), it has a wide input voltage range of 0.5–5.5 V, and a conversion efficiency (see Footnote 16) higher than AmbiMax’s operating power range. The conversion efficiency of LTC3401 is over 85% in the 10–50 mA output current range, when its output voltage is set to 4.1 V. The input voltage and current ranges are well matched with the power characteristics of the ambient power sources. The output voltage is set to 4.1 V to provide similar input voltage range (2.7–4.2 V) to the sensor nodes designed to use a single cell Li-polymer or two AA type batteries. The hysteresis band of the comparator is set to 100 mV, and accurate 0.1% resistors are used for configuration resistors, R1–R4. The S1087 photodiode (HAMAMATSU 2014) is the source sensor for the solar panel, and the TC9401 (Microchip 2007) is the voltage-to-frequency converter for the wind generator.

Reservoir Capacitor Array As shown in Figs. 7.32 and 7.37, the reservoir capacitors (RCs) from different energy harvesting subsystems form a reservoir capacitor array (RCA). The RCs resolve the imbalance between power generation and consumption. They smooth out the wide dynamic range and possible sudden, unpredictable change of the ambient power source, which can cause the system to operate unreliably or fail. Also, the wide dynamic range makes it hard to directly charge a lithium polymer battery. RCs also enable drawing power from the capacitors instead of the battery whenever possible to slow down battery aging due to deep discharging cycles. An RC is connected to the RCA via a diode, which has a voltage drop but provides the necessary isolation among the super-capacitors and the battery.

7.2 Energy Harvesting Projects

555

Fig. 7.37 Control and charger (Park and Chou 2006a)

Legend: PS: Power sources forming the RCA

Super-capacitors of this network can have different capacities. When choosing the capacity, it is needed to consider the characteristics of the power source as well as power consumption rate of the load system. The super-capacitor for the ambient power source with a low power output should have a smaller capacity. Otherwise, such super-capacitor will not deliver current, because its terminal voltage is always lower than other super-capacitors in the network. For RCs implementation, super-capacitors from Panasonic are used (Panasonic 2016b). Two 22 F super-capacitors are connected in series for the solar panel, and two 10 F super-capacitors are in series for the wind generator. The maximum voltage of each super-capacitor is 2.3 V. Because the output voltage of the harvesting regulator is set to 4.1 V, two capacitors had to be serially connected to increase the voltage level. A high-conductance 1N4148 (NXP Semiconductors 2004) fast diode is adopted.

Control and Charger The control and charger (CC) subsystem includes a window comparator, threshold detector, load switch, battery charger, and a battery protector (Figs. 7.33 and 7.37). By comparing, Vcap, the terminal voltage of RCA, and, Vth, the user-configured threshold voltage, this subsystem determines which power source, the battery or RCA, should power the load. When Vcap > Vth, the RCA powers the load system. Otherwise, the threshold detector turns ON the load switch so that the battery powers the system. Also, the window comparator, which takes Vcap and Vbattery,

556

7 Energy Harvesting Projects for WSNs

decides if the RCA charges the battery; when Vcap is higher than the threshold voltage (Vth1), and Vbattery is not fully charged, the comparator turns ON the charger. An adjustable current limit switch is the battery charger instead of COTS battery charging ICs that consume nontrivial power. The current limit switch allows only programmed current level to pass through the switch when it turns ON, that is, it behaves like what the battery charger does in constant-current mode. To implement the control and charger unit, two LTC1441 window comparators (Linear Technology Corporation 1996) are chosen, and an ST890 (STMicroelectronics 2013) adjustable current limit switch. The Vth is set to 2.7 V so that when Vcap is lower than Vth, the battery powers the load system. The 2.7 V are set because the conversion efficiency of Eco’s switching regulator (Linear Technology 2000a), used for outdoor experimentation, drops sharply under 2.7 V. It is desirable to maintain the output voltage of AmbiMax higher than 2.7 V to achieve high overall energy conversion efficiency. For different wireless sensor nodes and regulators, Vth can be adjusted accordingly. Also, Vth1 is set to 3.7 V so that when Vcap is higher than 3.7 V, AmbiMax starts to charge the battery until either Vth1 drops below Vth1 or the battery is fully charged.

7.2.6.2

Experimentation Outcome

AmbiMax was tested both indoor and outdoor; interesting findings can be identified: • It is a low cost, low hardware complexity platform. • AmbiMax uses entirely analog control without involving a microcontroller; hence, it does not consume precious I/O pins or ADC/DAC channels and does not incur software or memory overhead. • In addition to platform independence, AmbiMax’s modular structure is easily expandable to more energy harvesting sources, such as water flow, vibration, etc. • Indoor Experimentation of AmbiMax prototype. To emulate sunlight indoors, a halogen lamp with 1000 lm maximum output light is used; the light output is varied using a dimmer. To evaluate a wind energy harvesting system, a fan is used to generate wind at a maximum speed of 16 km/h. The following information is acquired: – For solar energy harvesting as compared to Prometheus (Sect. 7.2.2): Under the same light intensity, AmbiMax maintains the Vcap at 4.1 V until it starts to discharge, while Prometheus’s Vcap is always less than 3.0 V, and it drops further as, Vsolar, the solar panel terminal voltage decreases. Most of the time, the amount of current harvested by Prometheus is much lower than that of AmbiMax. AmbiMax can charge the super-capacitor 12.5 times faster (0.4 s) and can harvest three times more energy (1318.3 J) under identical sunlight intensity.

7.2 Energy Harvesting Projects

557

Fig. 7.38 AmbiMax prototype (Park and Chou 2006a)

Actually, AmbiMax is more efficient because it can continue harvesting energy when Vsolar < Vcap while Prometheus cannot. – For wind energy harvesting. To charge the 1.5 F 5 V super-capacitor using a wind generator, the speed of the fan is set at approximately 8.3 m/s and the super-capacitor is connected to AmbiMax. The obtained MPP at this speed is about 2.7 V. AmbiMax tracks the MPP by limiting the current draw at 17– 18 mA, even though the super-capacitor demands much more current. – Overhead. The current consumed by AmbiMax’s analog control circuitry and switching regulator is less than 500 lA when 3.3 V are supplied. • Outdoor Experimentation. Experiments were conducted near the Vincent Thomas Bridge in San Pedro, CA from 9:00 am to 11:00 pm. The configuration as shown in Fig. 7.38 was assumed to power up the Eco wireless sensor node for this 14 h period; it embraces a 70 mAh Li-polymer battery and two 10 F super-capacitors for each of the two ambient power sources. The Eco wireless sensor nodes (Park and Chou 2006b), as presented in Chap. 3 of this book, has a 2.4 GHz radio and an 8051 compatible MCU core in the same package. The entire sensor node including the battery measures barely 13 mm
10 mm 8 mm; it consumes 22 mA in receive mode and less than 10 mA in transmit mode (at 1 Mbps and 0 dBm power). The Eco node was programmed to keep sending 10 Bytes of data every 1 ms. Experimentation illustrated that the Eco node battery took over after the 14 h period, when the RCAs terminal voltage of the AmbiMax dual ambient energy harvesting system dropped to 2.5 V.

558

7.2.7

7 Energy Harvesting Projects for WSNs

Sunflower: Low-Power, Energy Harvesting System with Custom Multichannel Communication Interface

Sunflower developed at the Department of Electrical and Computer Engineering, Carnegie Mellon University, is a self-powered computing system that uses a combination of a PIN photodiode array, switching regulators, and a super-capacitor, to provide a small footprint renewable energy source (Stanley-Marbell and Marculescu 2007). The design provides software-controlled power-adaptation abilities, for both the main processor and its peripherals. The system’s power consumption is characterized, and its energy scavenging efficiency is quantified with field measurements under a variety of weather conditions. Developing the Sunflower hardware platform aimed at designing a general-purpose, miniature computing element, that could be used as the basis for evaluating the basis of designing macro-sensor electro-mechanical systems (MSEMS). Several points are to be remembered when comparing MSEMS to MEMS, and to sensor networks: • Like their microscale counterpart MEMS, MSEMS combine electronics with mechanical actuation; and like sensor networks, they involve multimodal sensing on multiple computing nodes. • Unlike MEMS, in which electric circuits and mechanical systems are integrated at the die-level23 (Ayers 2009), MSEMS incorporate computational devices (e.g., microcontrollers), sensors and mechanical actuators (e.g., shape memory alloy strands) over a large surface area, such as a sheet of plastic. • Unlike sensor networks, which typically involve discrete nodes manually placed in an environment, spanning a large area and thus necessarily communicating over wireless networks, MSEMS are integrated into the materials where they control actuation, and thus can communicate over wired channels. With several nodes in proximity, a multichannel wired interconnection can provide significantly reduced communication medium contention, as well as improved energy efficiency. Considering the Sunflower miniature hardware platform, several constraints are imposed and must be accounted for: • The low power consumption. • The availability of multiple wired communication interfaces. This constraint arises from the desire to interconnect multiple systems in topologies that enable path diversity.

23

A die in the context of integrated circuits is a small block of semiconducting material, on which a given functional circuit is fabricated. Typically, integrated circuits are produced in large batches on a single wafer of electronic grade silicon (EGS) or other semiconductor through processes such as photolithography. The wafer is cut (“diced”) into many pieces, each containing one copy of the circuit. Each of these pieces is called a die.

7.2 Energy Harvesting Projects

559

• Small overall size and weight including that of the energy source. Such constraint arises from the need to use actuators to achieve mechanical control of the materials in which the nodes are embedded; the use of weighty energy sources, such as AAA or CR2032 batteries, is therefore undesirable. The desire to embed these hardware devices permanently in materials precludes the use of components whose performance may degrade significantly over time, which pushes forward the investigation of alternatives to electrochemical batteries.

7.2.7.1

System Components and Design Objectives

Overview The aforementioned constraints guided the design of the Sunflower system architecture, shown in Fig. 7.39. Central to the design is an MSP430F1232 microcontroller (Texas Instruments 2004a). The first constraint is addressed by incorporating hardware facilities to enable the software to take advantage of all power saving modes of the microcontroller and other system components. The second design constraint is encountered by implementing a custom communication interface for the microcontroller, within a Xilinx XC2C32A complex programmable logic device (CPLD) (Xilinx 2008). All components in the system were carefully chosen to minimize printed circuit board (PCB) area, device count, and weight, addressing thus the third design constraint listed above. The Sunflower system incorporates an unconventional energy scavenging subsystem, composed of an array of PIN photodiodes24 (Radio-Electronics.com 2016), a boost switching regulator, and a miniature surface-mount super-capacitor, enabling a PCB implementation in less than 2 cm2. A wealth of sensors is embraced; specifically, a programmable color detector, a microphone, a two-axis accelerometer and a temperature sensor, as well as up to 24

The PIN diode, P-I-N diode is essentially a refinement of the ordinary PN junction diode. The PIN diode differs from the basic PN junction diode in that the PIN diode includes a layer of intrinsic material between the P and N layers. As a result of the intrinsic layer, PIN diodes have a high breakdown voltage and they also exhibit a low level of junction capacitance. Also, the larger depletion region of the PIN diode is ideal for applications as a photodiode. The PIN diode is used in a number of areas as a result of its structure proving some properties that are of particular use: • High voltage rectifier. The PIN diode can be used as a high voltage rectifier. The intrinsic region provides a greater separation between the PN and N regions, allowing higher reverse voltages to be tolerated. • RF switch. The PIN diode makes an ideal RF switch. The intrinsic layer between the P and N regions increases the distance between them. This also decreases the capacitance between them, thereby increasing the level of isolation when the diode is reverse biased. • Photodetector. As the conversion of light into current takes place within the depletion region of a photodiode, increasing the depletion region by adding the intrinsic layer improves the performance by increasing the volume in which light conversion occurs.

560

7 Energy Harvesting Projects for WSNs

Flash memory Core

Microphone 2.7 V

I/O

SPI

ADC 1.8 V

SRAM

GPIO Color sensor

Flash memory

ADC Accelerometer

2.7 V

Energy scavenging subsystem

Voltage regulator

GPIO

Temperature sensor

CPLD

GPIO UART

LED

UART2 … UART0

Microcontroller

Fig. 7.39 Sunflower architecture (Stanley-Marbell and Marculescu 2007)

8 MB flash memory, and a software controllable LED. Due to the differing power rail requirements of these diverse devices, a dual-output high-efficiency switching regulator is used to provide two voltage rails. The microcontroller takes advantage of the features of this power regulation subsystem, as well as the power-down features of the individual components in the system, to provide multiple opportunities for dynamic power adaptation and saving. The hardware design was carried out using Cadence CIS for schematic capture (Cadence 2016a), and OrCAD PCB for layout (Cadence 2016b). To satisfy the goal of minimal size, the hardware is implemented in a six-layer printed circuit board process, on 0.032″ FR4 laminate25 (AirBorn Electronics 2013), with four signal/ routing layers, and two plane layers for ground and one of the voltage rails. The partially populated PCB size is 0.9″
1.2″
0.032″, and was manufactured using a lead-free process. A description of the hardware platform is shown in Fig. 7.40. As Sect. 7.2.7.3 details, in measurements performed on the prototype hardware, the photodiode array and boost converter can charge the super-capacitor to a level sufficient to start the system’s voltage regulators in approximately one hour. Under ideal lighting conditions, the super-capacitor can be fully charged in approximately five hours. In the absence of sunlight, it can power the system for up to four hours 25

FR-4 PCB Laminate is the most commonly used base material for printed circuit boards. The “FR” means flame retardant, and Type “4” indicates woven glass reinforced epoxy resin.

7.2 Energy Harvesting Projects

561

(a) Top surface

(b) Bottom surface

(c) Side view

Fig. 7.40 Sunflower PCB (Stanley-Marbell and Marculescu 2007)

when executing a low-duty-cycle application, and longer if all the devices in the system are in their low-power states. The system can also be charged remotely, and at much higher speed than by sunlight, by illuminating the PIN photodiode array with a light source such as a laser, from a distance. Comparing the Sunflower hardware to several existing research and commercial sensor platforms is presented in Table 7.3. Even though the Sunflower hardware was designed primarily for use in investigating the construction of high-density MSEMS platforms with multichannel wired communication interfaces, such comparison is meant to illustrate the related application domains. Due to its expansion interface, the Sunflower hardware may be repurposed as a wireless sensor node. The upcoming sections detail the implementation of the system’s communication interface, the power regulation system design, and the energy scavenging.

Size with battery

Energy scavenging

Sunflower 1.2″
0.9″
0.2″ Yes Mica2 2″
1.5″
1.2″ No mote Mica2dota 1″ diameter, 0.5″ height No No Eco nodeb 0.5″
0.5″
0.3″ Telos 1.3″
3.2″
1.1″ No mote Heliomote 2″
1.5″
1.2″ Yes a Crossbow (2002b) b Eco node platform (Park and Chou 2006b)

Platform 4 0 0 1 3 0

\5 years \5 years \5 years \5 years

2 2 1.1 5

Default number of sensors

Unlimited \5 years

Maximum system lifetime

0.7 5

Minimum CPU current (lA)

Table 7.3 Comparison of Sunflower to contemporary sensor platforms (Stanley-Marbell and Marculescu 2007)

1, wireless

1, wireless 1, wireless 1, wireless

3, wired 1, wireless

Number of communication interfaces

562 7 Energy Harvesting Projects for WSNs

7.2 Energy Harvesting Projects

563

Table 7.4 Operating voltage ranges and power budgets of the active components (Stanley-Marbell and Marculescu 2007) Device

Voltage range

Power consumption at 2.7 V (1.8 V for CPLD core)

MSP430F1232 (microcontroller)

1.8–3.6 V

ADXL320 (accelerometer) TCS230 (color sensor)

2.4–5.25 V 2.7–5.5 V

XC2C32A (CPLD/communication interface) core XC2C32A (CPLD/communication interface) LVCMOS2.5 I/O AT45DBxx (off-chip flash memory)

1.7–1.9 V

0.270 lW (LPM4), 0.540 mW (active, 1 MHz) 0.945 mW 18.9 lW power down, 5.4 mW active 59.4 lW (quiescent)

2.3–2.7 V

–

2.5–3.6 V

SP0103 (microphone) TPS61070 switching boost regulator TPS61100 dual-output boost/LDO

1.5–5.5 V 0.9–5.5 V 0.8–3.3 V

0.025 mW power-down, 0.063 mW standby, 18.9 mW read 0.473 mW 0.051 mW 0.175 mW

Communication Interface One of the design goals of the Sunflower hardware platform was to provide the hardware with multiple, addressable, wired communication interfaces. This is desirable for the target MSEMS hardware platform, where hundreds of devices are to be comprised in tiny areas. In such scenarios, wired communication interfaces become desirable over wireless ones, as at such high spatial densities of devices, there could be significant interference in a shared wireless channel. On the other hand, a multilink wired channel enables efficient isolation of local communications, and also offers the possibility of significant reduction in communication power consumption. A miniature 20-pin connector provides an interface for programming the microcontroller and loading designs into the CPLD, via their respective JTAG (Joint Test Action Group, IEEE standard 1149.1) interfaces26 (Corelis 2016). Since general-purpose I/O pins of the microcontroller, as well as both voltage rails are available at this interface, it can be used as an expansion connector for connecting a daughter-card. For instance, as listed in Table 7.4, this interface is being used for a daughter card with an IEEE 802.15.4 radio, to enable the repurposing of the Sunflower hardware in applications such as WSNs. In addition to the devices listed

26

The IEEE-1149.1 standard, also known as JTAG or boundary-scan, has for many years provided an access method for testing printed circuit board assemblies, in-system-programming, and more.

564

7 Energy Harvesting Projects for WSNs

in the table, there are over 50 surface-mounted passive devices in the design (resistors, capacitors, and inductors). The multichannel wired interface was implemented by multiplexing the MSP430F1232 hardware UART using a circuit implemented within the Xilinx XC2C32A CPLD (Xilinx 2008). The XC2C32A provides programmable logic in the smallest available board space and has very low power dissipation. The microcontroller’s UART is multiplexed into three communication interfaces, which are software addressable, with added hardware flow control support. The design uses only 9 of the CPLD’s 32 macrocells, and 16 of the 80 CPLD’s functional blocks. None of the CPLD’s 32 registers are used since the design is fully combinational.

Power Regulation Subsystem The power regulation subsystem provides two regulated voltage rails: 1.8 V for the CPLD core voltage and 2.7 V for the CPLD I/O voltage and all other devices (Fig. 7.39). It was implemented with a TPS61100 dual-output regulator (Texas Instruments 2004b) composed internally of a switching boost converter (see Footnote 14) and a low dropout voltage (LDO) regulator27 (Williams 1989; Day 2002). The TPS61100 provides two signals for monitoring its input supply voltage, i.e., the output voltage of the energy storage subsystem. These signals, which are connected to the microcontroller, provide four threshold alert signals about the conditions of the system’s energy store and enable adaptive power management based on its state. For instance, given the current level of the energy store, the software can change the microcontroller’s operating frequency to reduce power consumption to a level optimal for the voltage regulator efficiency.

27

There are two types of linear regulators: standard linear regulators and low dropout linear regulators (LDOs). The difference between the two is in the pass element and the amount of headroom, or dropout voltage, required to maintain a regulated output voltage. The dropout voltage is the minimum voltage required across the regulator to maintain regulation. A 3.3 V regulator that has 1 V of dropout requires the input voltage to be at least 4.3 V. The input voltage minus the voltage drop across the pass element equals the output voltage. LDOs are a simple inexpensive way to regulate an output voltage that is powered from a higher voltage input. They are easy to design with and use. For most applications, the parameters in an LDO datasheet are usually very clear and easy to understand. However, other applications require the designer to examine the datasheet more closely to determine whether or not the LDO is suitable for the specific circuit conditions. The advantages of a low dropout voltage regulator over other DC/DC regulators include the absence of switching noise (as no switching takes place), smaller device size (as neither large inductors nor transformers are needed), and greater design simplicity (usually consists of a reference, an amplifier, and a pass element). A significant disadvantage is that, unlike switching regulators, linear DC regulators must dissipate power across the regulation device in order to regulate the output voltage.

7.2 Energy Harvesting Projects

7.2.7.2

565

Power-Adaptive Design

The core of the Sunflower system is the 16-bit MSP430F1232 microcontroller (Texas Instruments 2004a). The device features an 8 KByte on-chip flash memory for code and data storage, and a 256-Byte RAM. It operates at up to 8 MHz software-selectable clock speeds, with operating voltage in the range of 1.8–3.6 V; in the Sunflower hardware design, the operating voltage is fixed at 2.7 V. This is the lowest voltage at which the microcontroller could be operated without resorting to voltage level converters for interfacing the connected peripherals, as depicted in Table 7.4. Such design decision positively impacts the system’s power consumption, as the leakage power of the microcontroller is lower at this reduced operating voltage, and the dynamic power is also reduced depending on workload. Further eliminated, the power cost of additional level converter ICs for interfacing peripherals when operating at different voltages, which might be comparable to that of the microcontroller.

Microcontroller Power Adaptation The MSP430F1232 microcontroller features six different operating states; notably, an active mode and five low power modes (referred to as LPM0 through LPM4) that can be individually activated by setting appropriate bits in a system status register. In addition to these low power modes, the operating frequency can be set under software control. These modes, as well as frequency setup, affect energy in several respects: • Frequency scaling does not lead to energy reduction for compute-intensive applications. While the average power is reduced by frequency scaling, the overall time to complete a computation is increased and the energy consumed remains unchanged. • If the computation involves communicating with peripherals, then as computation progresses more slowly, the peripherals will need to be active longer, leading to an increase in energy consumption. • The low power modes provide the ability to reduce energy consumption by deactivating components in the MSP430F1232 microarchitecture. Underlying the first four low power modes, LPM0–LPM3, is the use of an appropriate clock source for the microcontroller’s timer unit, which can be programmed to generate periodic interrupts to wakeup the CPU from sleep. • The microcontroller can synthesize its own clock using an on-chip digitally controlled oscillator (DCO), which can be provided as the clock for the timer unit. The DCO is however disabled in power modes LPM1–LPM41. In LPM1– LPM3, an external crystal may be used to provide a reference for an on-chip oscillator; however, this source is also deactivated in the lowest power mode, LPM4. The lowest power sleep mode, LPM4, may be viewed as more of a power-down or stop-mode than a sleep mode. In this mode, while the power

566

7 Energy Harvesting Projects for WSNs

consumption is about 3000 times less than that in active mode, and as all clocks are disabled, the processor can only be woken up by an external interrupt source, e.g., as a result of a user pressing a button. The Sunflower hardware design facilitates the use of all five low power modes, including LPM4: • LPM1–LPM3 are made easy by the inclusion of an external 32.768 kHz crystal to provide a clock reference. The microcontroller uses this reference to synthesize on-chip clock frequencies of up to 8 MHz. • LPM4, the lowest-power mode, is assisted via using an external R-C circuit to generate a delayed interrupt trigger. Before entering LPM4, the software configures an I/O pin connected to an R-C network, to generate interrupts on a rising edge. Another I/O pin, also connected to this R-C network, is subsequently driven high. After a delay dependent on the R-C time constant, the microcontroller receives an interrupt due to the rising edge at the R-C circuit; this delay is approximately set to 0.5 s. Using this simple technique, a software routine is implemented to enable an LPM4 sleep of any chosen duration (with 0.5 s granularity). The power dissipated in the R-C circuit as well as the charge lost therein each cycle, negates though some of the benefits of this sleep circuit approach.

System-Level Power Adaptation Reduction of overall system energy consumption requires intelligent power control of all components, including the microcontroller and peripherals. Minimizing power dissipation is a main objective that goes along more than a direction: • To minimize power dissipation, the Sunflower hardware design enables the microcontroller to shutdown the color sensor and external flash memory under software control. Notably, the accelerometer and microphone do not provide for device shutdown. • The voltage regulator is also a source of power dissipation. The TPS61100 (Texas Instruments 2004b) combined switching/LDO dual-output regulator is used to provide two voltage rails for the system operating through the energy scavenging subsystem (Section “Overview”). It delivers energy conversion efficiencies (see Footnote 16) in the range 65–95%, depending on the output current. Poor efficiencies at low load currents necessitate the use of techniques to deactivate or bypass the voltage regulator while the system components are idle. In order to conserve the energy store of the system, Sunflower shuts down the switching regulator by deactivating its internal pulse width modulation (PWM) circuit, such that the output load runs directly off the regulator’s output super-capacitor. In the Sunflower hardware design, after software running on the microcontroller shuts down unused peripherals, it can further conserve energy

7.2 Energy Harvesting Projects

567

by enabling shutdown mode in the voltage regulator. When the regulator output super-capacitor drops below a threshold, the TPS61100 automatically reactivates.

7.2.7.3

Sunflower Potential and Forecast

Energy Scavenging Subsystems Compared Energy scavenging techniques are diverse; among many there are solar, piezoelectric, mechanical, and thermal gradient approaches (Sect. 2.3). Compared to contemporary integrated self-powered platforms such as Everlast (Sect. 7.2.5), Sunflower employs a switching regulator to charge the super-capacitor from the photodiodes, and it is over an order of magnitude smaller in volume, albeit intended for different purposes. In contrast to energy harvesting systems such as the Heliomote add-on board for the Mica2 platform (Sect. 7.2.4), the Prometheus add-on for the Telos mote hardware (Sect. 7.2.2), and the AmbiMax energy scavenging hardware (Sect. 7.2.6), the energy harvesting circuit in Sunflower is completely integrated into the system’s already miniature design. This integration is enabled by the use of an innovated combination of an array of PIN photodiodes, a boost switching converter, and a miniature surface-mount super-capacitor, to provide a small footprint renewable energy source. Although it is possible to use the photodiode array to directly charge an energy storage device (in this case a 0.2 F super-capacitor), the potential drop across each photodiode is typically only 0.4 V in daylight. Figure 7.41 shows the measured average potential drop per photodiode. From these measurements, using the array of four PIN photodiodes in the Sunflower platform to directly charge the super-capacitor would mean charging it to only one half of its capacity (i.e., approximately 1.6 V vs. its rated voltage of 3.4 V), regardless of how long the system was exposed to sunlight. To make full use of periods of available sunshine, the four PIN photodiode array in the Sunflower platform is connected to the super-capacitor through a TPS61070 switching boost converter (see Footnote 14) (Texas Instruments 2015). The TPS61070 has a startup voltage of 0.8 V, thus it can fully charge the super-capacitor storage cell in lighting conditions, which yields as little as 0.2 V per photodiode. The additional circuit board overhead of the boost converter subsystem is approximately that of a single photodiode. Figure 7.42 shows the charging curve for the system’s energy store, the 0.2 F super-capacitor being driven by the energy scavenging subsystem (the PIN photodiode array and boost converter), in outdoor lighting conditions.

7 Energy Harvesting Projects for WSNs

Voltage

568

Time (hr) Legend: A 12-days measurement period (May 6th 2006, 7:23pm – May 18th 2006). The measurements were taken at longitude 40.4422 N, latitude 79.9464 W, elevation 900 ft, in an indoor environment, close to a window, but not in direct sunlight.

Voltage

Fig. 7.41 Average measured voltage drop per PIN diode (Stanley-Marbell and Marculescu 2007)

Time (hr)

Fig. 7.42 Voltage at output of TPS61070 regulator, the super-capacitor (Stanley-Marbell and Marculescu 2007)

Remote Charging via Infrared Laser The energy scavenging subsystem affords opportunities for remote charging. This might be useful, when extended periods of darkness might cause the energy store to be fully depleted. Although effective in terrestrial sunlight, the PIN photodiode array employed in the scavenging circuit has peak sensitivity at 850 nm in the infrared portion of the electromagnetic spectrum. While remote charging can be achieved via illumination with a bright spotlight (e.g., from an incandescent light source), the amount of energy needed to drive such a spotlight could be large, in the order of hundreds of watts. Figure 7.43 shows the capacitor charging curve for a single photodiode, i.e., one-quarter of the full scavenging unit, illuminated with the laser, driving the 0.2 F energy storage super-capacitor. Using this matched light source, the energy scavenging subsystem can be fully charged in approximately five minutes compared to five hours in sunlight.

569

Voltage

7.2 Energy Harvesting Projects

Time (min) Legend: • Charging a 0.2 F super-capacitor with a single PIN photodiode. • The photodiode is illuminated via a 5 mW infrared laser (optical output) that is operated from a 2.0 V supply, and drawing 30 mA.

Fig. 7.43 Charging the super-capacitor with a single PIN photodiode (Stanley-Marbell and Marculescu 2007)

Future Betterments The Sunflower platform is a miniature self-powered computing system built upon a combination of an array of PIN photodiodes, a switching regulator, and a super-capacitor, to provide a small footprint energy scavenging power source. It was designed as a platform for the investigation of ideas related to the construction of MSEMS. From experimentation several directions for improving the design were found worth considering: • The accelerometer and microphone peripherals may be power-managed by powering them directly from an I/O pin of the microcontroller, as they consume small current to make this option feasible; or their operating voltage rails may be gated with a low-quiescent current FET switch. • The utilization of the microcontrollers deep sleep mode may be made more efficient by the use of an external watchdog timer for waking up the microcontroller from sleep in place of the initially adopted R-C sleep circuit. • Other directions would consider the integration of additional sensors such as humidity sensors.

7.2.8

Micro-Solar Power Sensor Networks for Forest Watersheds

A framework is proposed at the University of California, Berkeley for the design of micro-solar subsystems in WSNs (Taneja et al. 2008). Its motivation is to design a micro-climate network for the HydroWatch Project intended for the studies of hydrological cycles in forest watersheds (Keck Hydrowatch Center 2008).

570

7 Energy Harvesting Projects for WSNs

Remarkably, many tools and calculators are available for macro-solar installations in residential and commercial applications, but on the contrary for in situ micro-solar power, only sketchy proposals are revealed in the sensor network literature. Noticeably, the basic components, such as solar panels, regulators, and batteries, are well documented (Messenger and Abtah 2010), but the selection, sizing, and composition of the components are not considered. The design issues are rather different from those in macro-solar systems because of the very small power transfers involved, precisely, microwatts to milliwatts rather than kilowatts to megawatts. A micro-solar system operates at very different efficiencies and every bit of power conditioning or monitoring impacts the overall performance. Also, the ease of installing the micro-solar panels on a convenient rooftop with ample exposure is not available; habitually, they are installed where the measurements are to be taken, regardless of how much they might be shaded. Nonetheless, new degrees of design freedom are realized by the tiny magnitude of the energy requirements. This framework presents a formulation of a general model of micro-solar systems that is sufficient for constructing a capacity planning “calculator” to guide the sizing of the various elements. A concrete design is developed for the HydroWatch application, a smartly engineered climate monitoring node and network with a flexible power subsystem that can support various specific design details and provide visibility into the solar performance in real-application settings.

7.2.8.1

Solar Panels

A solar panel transforms available incident solar radiation to electrical power. A given panel is characterized by its IV curve and, in particular, three points: open-circuit voltage (Voc), short-circuit current (Isc), and its maximum power point (MPP). Internally, these parameters are determined by the serial and parallel composition of the solar cells and the total area of the panel. Increasing temperature depresses the IV curve somewhat, reducing the power output. The portion of incident solar energy that is available at the panel is determined by a variety of environmental factors: • The absorption by the atmosphere, typically, the weather factors such as clouds, fog, etc. • Any particular point of installation will have various obstructions and shadows. Attenuation factors can only be characterized empirically. Experience with many deployments in different settings can provide statistical models. Care in choosing deployment sites can potentially improve the expected solar energy availability.

7.2 Energy Harvesting Projects

571

Macro-solar Panels Versus Micro-solar Panels For the large, expensive panels used in macro-solar installations parameters are accurately characterized in data sheets and are well validated. On the other hand, for the small, inexpensive panels used in micro-solar applications, empirical characterization is often required. More importantly, the operating point of the IV curve is determined by the load experienced at the panel, which is determined by the input regulator or the storage facility and downstream load in the absence of a regulator. For most panels, the IV curve is nearly flat for voltages less than that of the MPP, so power increases nearly linearly with V. As shown in Fig. 7.44, the input regulator conditions the output of the panel to meet the operational constraints of the particular battery, including voltage limits, current limits, and charge duration. While macro-solar inverters operate for about 95% efficiency in the subwatt range, regulator efficiencies are in the range of 70– 80%. The product of such low efficiencies translates into a significant overall supply/demand ratio. For storing charge, a wide range of battery construction and chemistries are available, as well as super-capacitors. They have differing operating voltages, charge algorithms, and complexities. From a system design perspective, it is necessary for the power subsystem to be able to charge a fully discharged battery without software in the loop, so that when placed in sunlight the device is guaranteed to eventually become active.

External environment

Occlusions

Solar panel

Psol

Output regulator (Effreg-out)

Input regulator (Effreg-in) Pbat-chg

Pmote

Load

Pbat-dis Energy storage

Legend: Psol: Power generated from the solar panel. Pbat-chg: Power input to charge the energy storage. Pbat-dis: Power discharged from the energy storage. Pmote: Power consumed by the load. Pshunted: Power being shunted when in excess. Effreg-in: Power efficiency of the input regulator. Effreg-out: Power efficiency of the output regulator. Effbat: Charge-discharge efficiency of the energy storage.

Fig. 7.44 Micro-solar system architecture and parameters (Taneja et al. 2008)

572

7 Energy Harvesting Projects for WSNs

The portion of energy transferred into the battery during the day and discharged during the night incurs additional transfer efficiency, Effbat, of 66% for NiMH chemistries. The capacity of the battery determines the potential lifetime in darkness, but also how much energy can be harvested while the sun shines. The output regulator matches the battery characteristics to the requirements of the mote. It is characterized by its efficiency, Effreg–out, particularly at two different operating points, 10 s of microwatts most of the time, and 10 s of milliwatts during short active periods. For a typical bimodal Pmote, an effective efficiency of 50% or less is expected. This determines the load experienced by the supply and storage components of the power subsystem. Generally, the daily battery and solar panel power cycle has five phases: • From sundown to sun up, the battery discharges, supplying the device load. • When the panel is initially illuminated, a transition period occurs during which the battery provides only a portion of the device load. • With sufficient illumination, the panel supports the entire load and delivers charge into the battery. • If this recharge period is sufficiently long, the battery becomes fully charged and the system operates in saturation, thus power is shunted. • Eventually, a dusk transition occurs similar to dawn. The efficiency coefficients impose the net change in battery capacity over the daily cycle, knowing the starting capacity, supply power, and demand power. The framework sizing guideline assumes that the recharge period would need to be no more than half an hour, possibly distributed throughout the day. Merely, saturation preserves capacity. Obviously, a series of cloudy days may result in a progressive drop in battery capacity, which would then increase the recharge duration when the weather clears up. In the micro-solar setting, given the ratio of mote load and typical battery capacities, it is wise to come out with design schemes that captivate entire seasonal variations in weather patterns. Calculating solar availability during 2% of operation (i.e., a half hour of radiation during the day), and a 3:1 supply/demand ratio from the product of efficiencies Effreg–out, Effbat suggests that the solar panel needs to be sized at 150 times the average demand. This makes serious every aspect of the micro-solar subsystem design.

7.2.8.2

Network and Node Design

The HydroWatch Project (Keck Hydrowatch Center 2008) targets gathering extensive, high frequency, and automated observations of the water lifecycle progressing through a forest ecosystem. A robust network of low-maintenance sensor nodes is deployed to collect scientifically significant data indefinitely while enduring a defiant wet forest environment. In this section, a focus is accorded on the network architecture and mechanical design, and on the micro-power solar subsystem mounted on each node.

7.2 Energy Harvesting Projects

573

Network Architecture The sensor network architecture follows the canonical habitat monitoring form (Szewczyk et al. 2004). The sensor node is based on the TelosB-compatible Tmote Sky (Moteiv 2006). The mote software provides periodic data acquisition, thresholding, power management, remote command processing, and health monitoring; it is a modified Primer Pack/IP based on TinyOS 2.0 (Arch Rock Corporation 2015). The patch network is an implementation of IPv6 using 6LoWPAN over IEEE 802.15.4 radios (Montenegro et al. 2007), it utilizes a packet-based low-power listening (Polastre et al., Versatile Low Power Media Access for Wireless Sensor Networks 2004) to minimize idle listening. Data collection is implemented as UDP packets; the routing layer adopts hop-by-hop retransmissions and dynamic rerouting in a redundant mesh (up to three potential parents) to provide path reliability on lossy links. Trickle-based (Levis et al. 2004) route updates are used for topology maintenance. Source-based IPv6 routing is used to communicate directly to specific nodes, while dissemination is performed as a series of IPv6 link-local broadcasts. In the initial HydroWatch deployment at the Angelo Reserve in Northern California, the sensor patch embodied 19 nodes over a 220
260 m2 area. The transit network between the basestation and the patch is implemented using the same node and link technology as the patch; so, there is no specific gateway node in the patch. To provide redundancy in the transit network, multiple micro-solar router nodes cover a 120 m stretch. These nodes are just patch nodes without the environmental sensors. The network depth is 5 hops or greater. The IEEE 802.15.4 bridge node attached to the basestation uses a high-gain (19 dBi) parabolic antenna. The backhaul network28 (WhatIs.com 2006) is a Wi-Fi-based IP network with repeaters on peaks and treetops to reach a T1 line29 (TechTarget 2016). The basestation is a Linux-class gateway server that provides a Web services front-end, and a PostgreSQL database (PostgreSQL 2013) for information storage and retrieval, it also offers a Web-based management console. The basestation is 28

Backhaul has several usages in information technology:

• In satellite communication, backhaul is used to mean getting data to a point from which it can be distributed over a network. • Manufacturers of network switching equipment use the term to mean “getting data to the network backbone“ (which is similar to its use in the satellite communication industry). • Backhauling is sending network data over an out-of-the-way route (including taking it farther than its destination) in order to get the data there sooner or because it costs less. This kind of backhauling involves understanding changing network conditions and economics. • Backhauling may sometimes be used to mean the use of the back channel on a bidirectional communications line. 29 The T1 (or T-1) carrier is the most commonly used digital transmission service in the USA, Canada, and Japan. In these countries, it consists of 24 separate channels using pulse code modulation (PCM) signals with time-division multiplexing (TDM) at an overall rate of 1.544 million bits per second (Mbps). T1 lines originally used copper wire but now also include optical and wireless media.

574

7 Energy Harvesting Projects for WSNs

also an IP router that allows end-to-end connectivity to the patch nodes. Moreover, it eases such tasks as remotely monitoring overall network health, diagnosing misreporting or missing nodes, and checking the quality of links between a node and its neighbors, a critical function during the deployment phase.

Engineering the Node The mechanical design of the node aims at providing accurate sensor readings over an extended time span (Fig. 7.45a). To provide trustworthy data, the sensors must be properly exposed to the environment while protecting the electronic parts. The internal electronic components are protected from environmental damage, by setting a limit on the number of node features requiring holes in the enclosure, preventing thus water vapor from infiltrating the interior. The mote platform composed of the microcontroller, radio, flash, system software, and networking is fairly common across many applications. On the other hand, the sensor suite, power subsystem, and mechanical design of the node are application-specific and are highly interrelated. The sensor suite for this microclimate monitoring application is mostly the one developed for tracking weather fronts in Redwoods (Tolle et al. 2005) and available natively on the TelosB platform, specifically, total solar radiation (TSR), photo-synthetically active radiation (PAR), temperature, and relative humidity (RH). The TelosB compatible Tmote Sky mote (Moteiv 2006) is used with an attached SMA connector 30 (Wellshow Technology 2016) for the external antenna; there are no onboard sensors. The mote is connected to external sensorboards via custom cables with IDC connectors31 (Jaycar Electronics 2004) to provide some freedom to control sensor orientation on the node. The sensors are affixed higher than their surroundings to avoid water and to obtain unobstructed indications of solar illumination (Tolle et al. 2005). A Sensirion SHT15 provides relative humidity (RH) and temperature (Sensirion 2016), it is factory calibrated to exhibit a maximum ±2% RH and ±0.3 °C error. To accurately measure humidity the sensor must be exposed to naturally aspirated air flow, whereas to measure temperature it should be shaded and decoupled from large thermal masses and sources of self-heating. Two tiny photodiodes are used: a Hamamatsu S1087 to measure incident photo-synthetically active radiation

30

SMA connectors are coaxial RF connectors developed in 1960s. SMA is the abbreviation of SubMiniature version A. SMA connector has a 50 O impedance, 1/4″-36 thread type coupling mechanism and offers excellent electrical performance from 0 to 18 GHz. SMA connectors are available in various quality classes for different applications with different material options. Various SMA connectors are widely used in vehicle tracking systems, wireless LAN, WiMax, telecommunication, aero, and precise testing instrumentation. 31 IDC technology is based on the idea of displacing or “pushing aside” some of the insulation around the cable conductors (wires), and making a direct electrical connection that way.

7.2 Energy Harvesting Projects

(a) HydroWatch node components

575

(b) HydroSolar power board organization

Fig. 7.45 HydroWatch and HydroSolar (Taneja et al. 2008)

(PAR) and Hamamatsu S1087-01 for the total solar radiation (TSR) (HAMAMATSU 2014). Since natural environments may cause considerable fluctuations in link quality causing wireless networks to fail severely and unpredictably, nodes are equipped with a 7 dBi omnidirectional antenna. Nodes were attached to the top of 3 and 4 ft metal fence posts. Micro-Power Subsystem The HydroSolar power board was designed to study a variety of power subsystem options. As shown in Fig. 7.45b, the core of the node design is a flexible power subsystem board that ties together a solar panel, an optional input regulator, a battery, and a switching output regulator. Several parameters are monitored out of such a configuration, specifically, time-series logs of solar panel voltage and current, battery voltage, in addition to the logs of sensor data from the application and neighboring nodes. These data are collected and stored by the gateway server for in depth analysis of the performance of the node and network under varying solar conditions.

7.2.8.3

Micro-Solar Panels Design Considerations and Implementation

Energy Storage For a measured average consumption of 0.53 mA at 3.3 V and a 50% efficiency of the output regulator, the daily energy requirement from the energy storage element is 79.2 mWh. This energy requirement guides the energy storage selection process. Table 7.5 compares each type of storage based on its capacity. Except the super-capacitor, other alternatives can provide energy for more than 30 days of operation without recharging, which is sufficient to operate a node for several days in the absence of solar radiation.

576

7 Energy Harvesting Projects for WSNs

Table 7.5 Node lifetime using energy storage elements without recharge (Taneja et al. 2008) Type

Lifetime

Lead Acid (LC-R061R3P) Two NiCd (KR-1100AAU) Two NiMH (NH15-2500) Li-ion (UBP053048) Li-polymer (UBC433475) Supercap (BCAP0350)

98.5 days (=7800 mWh/79:2 mWh/day) 33.3 days (=2
1320 mWh/79:2 mWh/day) 75.8 days (=2
3000 mWh/79:2 mWh/day) 35.4 days (=2800 mWh/79:2 mWh/day) 42.9 days (=3400 mWh/79:2 mWh/day) 3.8 days (=304 mWh/79:2 mWh/day)

Evoking Table 2.1 in Chap. 2, several rechargeable energy storage alternatives can be used for micro-solar power systems. Several characteristics are considered, including capacity, operating range, energy density, and charging method. For the micro-solar system, lead–acid batteries are inconvenient because of low energy density. NiCd batteries have a similar footprint and charging method as NiMH batteries, but with a much smaller capacity. Moreover, NiCd chemistries are less environmentally friendly and are far more liable to the memory effect, which considerably reduces battery capacity over time. For the choice between Li-based chemistries and NiMH, experience is drawn from the Trio deployment (Dutta et al. 2006). The selection of a NiMH battery with its straightforward charging logic was due to two considerations: • To avoid having software in the charging loop which allows nodes to simply charge when placed in the sun entirely independent of their software state. • The complexity of integrating a hardware Li-ion charger. Such choice does have some drawbacks, though. Its chemistry suffers from a self-discharge rate of 30% and an input–output efficiency of roughly 66%, both are worse than for any other battery chemistry considered. This implies that for every three units of energy that are input to a battery, only two units of energy are output. This cost was overcome by the simplicity of the charging logic. A two-cell configuration enables the operation without an input regulator. For increased capacity, it is possible to put two-cell packs in parallel. Additionally, since the discharge curve of NiMH batteries is relatively flat, most of the discharge cycle produces a near-constant voltage.

Solar Panel Several considerations are essential to have a solar panel suitable for a micro-solar subsystem. Typically: • Care is required when selecting a panel that will operate near its MPP. It is, hence, essential to determine whether the load it is expected to support is a combination of an input regulator and energy storage or energy storage only.

7.2 Energy Harvesting Projects

577

• The cell composition, that is how many cells are present and their serial/parallel arrangement, becomes a concern when the solar panel is partially occluded. • The physical dimensions of the panel should be compatible with the choice of the enclosure. For the HydroSolar power subsystem, a 4 V–100 mA panel (Silicon Solar 2015) was picked up. The MPP of this panel occurs at 3.11 V, which makes it appropriate for charging two NiMH cells directly. Using the developed astronomical model, it was possible to vary the latitude, day of year, time of day, panel orientation, and angle of inclination to match the conditions expected in the field deployments. Using a rule of thumb of 30 min of sunlight per day, the solar energy generated by this panel at its MPP is 139 mWh, which fulfills the 120 mWh (= 79.2 mWh/66% NiMH charge–discharge efficiency) per day requirement of HydroSolar.

Input Regulator To select the input regulator, the essential parameters are the operating range of the solar panel and batteries, and the method and logic used to charge the battery. In the proposed design, the trickle batteries charge is chosen because it requires only a simple circuit and no software control. In the initial design of the HydroSolar board, an input regulator was used to limit the voltage to the battery (Fig. 7.45b). However, the input regulator forced the solar panel to operate at a point far from its MPP. Dispensing of the input regulator results in significantly more energy harvested from the solar panel because the input impedance of the regulator is less than that of the battery. Moreover, energy is no longer consumed by the input regulator, which empirically has about a 60% efficiency factor. This substantial gain in total system energy as well as efficiency led to the removal the input regulator. Removing the input regulator, though, is just an option because the operating voltage of the solar panel matches the charging voltage of the batteries.

Output Regulator The basic criteria for choosing an output regulator are the operating ranges of the batteries and the load, as well as the efficiency of the regulator over the load range (Fig. 7.45b). When choosing two NiMH AA batteries, the nominal voltage of the energy storage is 2.4 V. A boost converter (see Footnote 14) is thus required to match the 2.7–3.6 V operating range of TelosB motes. The output regulator has also to provide a stable supply voltage to ensure the fidelity of sensor data. The LTC1751 regulator (Linear Technology 2000b) is picked up, its efficiency is nearly 50%, and it requires few discrete parts and has low constant switching noise.

578

7.2.8.4

7 Energy Harvesting Projects for WSNs

Evaluating the Design

A Sensor Network in an Urban Neighborhood The purpose of this deployment was to ascertain that nodes could sense, charge, and operate continuously for days, as well as to assess whether the developed model accurately estimated the generation and consumption of energy in a variety of solar conditions. A deployment of 22 nodes in an urban neighborhood in Berkeley is implemented. To emulate the forest watershed environment, nodes are placed in the vicinity of significant obstructions and the orientation of the solar panels was varied. Several findings are interestingly revealed: • In a fairly sunny day, a widest distribution of received solar energy (roughly 100–1700 mWh) results among the nodes, with the high-end nodes, the more exposed to sun, receiving more energy. • As the days became cloudier, the variance of the distribution lessened (roughly 200–1150 mWh); nodes at the high end of the distribution received slightly more than half the solar energy when compared to a sunny day. • Interestingly, nodes on the lower end of the distribution received more solar energy (100–200 mWh) on cloudier days; this is because the diffusion of light, caused by clouds layers, scatters the light source and boosts the chance of the normally occluded solar panel to harvest more solar energy. • Every node harvests a surplus of energy on both sunny and cloudy days; from 1.4 to 14 surplus battery days in a sunny day, and from 1.6 to 8.35 surplus battery days in a cloudy day. Surplus battery days are calculated as:

Surplus battery days ¼ Surplus of energy flowing into the battery  Charge-discharge efficiency ð66%Þ Daily consumption ð79:2 mWhÞ ð7:14Þ

A Sensor Network in a Forest Watershed Compared to the urban deployment, several remarks came under focus in a forest watershed: • The blend of solar profiles seen by the nodes in the forest watershed was far less diverse. Most of the nodes received no more than 500 mWh of energy on any of the days of the deployment. • Remarkably, much less solar energy was harvested in the forest watershed. The best-performing node on a sunny day did receive 4.15 surplus battery days,

7.2 Energy Harvesting Projects

579

which is much less than the solar energy harvested by a median node on a cloudy day in the urban neighborhood. • Sun-starved nodes harvested less than the node consumption each day. Such daily energy deficit results in a negative number of surplus battery days. These nodes are suffering different degrees of sun starvation; some are only consuming about half a day’s worth of battery energy daily, while others are consuming a full day’s worth of energy daily. Noticeably, a majority of the nodes did not receive enough solar energy for sustained operation, which caused a finite network lifetime. Consequently, solar patterns considerably impact the micro-solar panels design considerations along the coming several dimensions: • The primary limitation of available solar energy in the forest is not due to the fewer amounts of light, but to its speckled nature. The spot of light that falls on the small panels is not large enough to illuminate the entire panel. Cloudy days diffuse the shadows, thus reducing spotting. • An individual solar cell produces about 0.5 V; so, several cells are to be placed in series within the panel to provide a useful output voltage. The cell current is determined by its area, so cells must be interconnected in various serial–parallel networks. However, when a cell in a serial chain is not well illuminated, it limits the current flow through the entire chain. Accordingly, enlarging the panel does not necessarily increase the power output in speckled light. Instead, many small panels should be connected in a parallel configuration. • Increasing the battery size also has surprising implications. With the low daily consumption of a well-engineered environmental monitoring application, it is reasonable to size batteries to last for several seasons. In deciduous forests, where leaves seasonally fall off the trees, this would allow nodes to store up all their energy after the leaves fall. Even in coniferous (evergreen) forests, energy can be collected when the interaction of the canopy and the sun angle are most favorable. • Additional improvements are possible through utilizing more efficient regulators. Also, exploration of novel collectors and storage profiles leads to further improvements in the models as well as the physical design.

7.2.8.5

Gained Experience

From the deployments of the HydroSolar board, it was found that the prediction of available sunlight was accurate for the urban environment, but highly optimistic for a forest watershed. Hence, the proposed micro-solar power model was further studied to identify potential designs of nodes that may operate indefinitely in forested or other solar challenging environments. The conducted study stressed on the unique design issues that differentiate micro-solar power systems from the familiar macro-solar power systems.

580

7.2.9

7 Energy Harvesting Projects for WSNs

Energy Harvesting from Hybrid Indoor Ambient Light and Thermal Energy Sources

In this project developed at the Department of Electrical and Computer Engineering, National University of Singapore, a hybrid energy harvesting (HEH) project from solar and thermal energy sources is proposed to enhance the performance of wireless sensor nodes in indoor environment (Tan and Panda 2011). Whenever solar energy is not available, the alternate approach is to harvest from the readily available thermal energy source so as to continue powering the operation of the wireless sensor node before exhausting the energy storage device. Another key characteristic of this HEH system is the ability to harvest simultaneously from both energy sources whenever they are available, while using one power management circuit, instead of harvesting from individual energy source one at a time (Lhermet et al. 2008; Tadesse et al. 2009; Guilar et al. 2009; Khaligh et al. 2010); thus boosting the performance of the indoor wireless sensor node.

7.2.9.1

Characterization of Indoor Energy Sources

Indoor Solar Energy Harvesting System Several mathematical models describe the operation of photovoltaic (PV) cells (Celik and Acikgoz 2007; Sera et al., PV Panel Model Based on Datasheet Values 2007; Villalva et al. 2009). In (Celik and Acikgoz 2007; Tan and Panda 2011), an electrical circuit with a single diode (single exponential) is considered as the equivalent photovoltaic model consisting of 15 PV cells in series, ns, as shown in Fig. 7.46. Assuming that the shunt resistance, Rsh, is infinite, the current-voltage (IV) characteristic of the PV module can be described with a single diode as the four-parameter model given by Eq. (7.15) (Celik and Acikgoz 2007): IPV ¼ IL  Io  ½eððVPV þ IPV Rs Þ=ns Vt Þ  1

Fig. 7.46 Equivalent electrical circuit for a PV module (Tan and Panda 2011)

ð7:15Þ

IPV ID IL

Rs

Ish Rsh

VPV

7.2 Energy Harvesting Projects

581

where, IL is the light-generated current (A), Io is the dark/reverse saturation current of the PN diodes (1 * 10−9 A), Rs is the series resistance of the PV module. Vt is the junction terminal thermal voltage (V) depending on the cell absolute temperature is defined as:

Vt ¼

k  TC q

ð7:16Þ

where, TC is the cell absolute temperature (K)32 (WhatIs.com 2011a), k is the Boltzmann constant33 (Encyclopædia Britannica 2015), q is the charge of the electron34 (Encyclopædia Britannica 2011)

32

The Kelvin (abbreviation K), less commonly called the degree kelvin (symbol, °K), is the standard international (SI) unit of thermodynamic temperature. One kelvin is formally defined as 1/ 273.16 (3.6609 * 10−3) of the thermodynamic temperature of the triple point of pure water (H2O). The kelvin scale differs from the more familiar celsius or centigrade (°C) temperature scale; there is no such thing as a below zero kelvin figure. A temperature of 0 K represents absolute zero, the absence of all heat. However, the size of the kelvin “degree” is the same as the size of the celsius “degree.” A change of plus-or-minus 1 °C is the same as a change of plus-or-minus 1 K. At standard earth-atmospheric sea-level pressure, water freezes at 0 °C or +273.15 K, and boils at +100 °C or +373.15 K. A temperature of 0 K thus corresponds to −273.15 °C. A temperature of 273.15 K corresponds to 0 °C. To convert a kelvin temperature figure to celsius, subtract 273.15. To convert a celsius temperature figure to kelvin, add 273.15. 33 Boltzmann constant, (symbol k), a fundamental constant of physics occurring in nearly every statistical formulation of both classical and quantum physics. The constant is named after Ludwig Boltzmann, a nineteenth century Austrian physicist, who substantially contributed to the foundation and development of statistical mechanics, a branch of theoretical physics. Having dimensions of energy per degree of temperature, the Boltzmann constant has a value of 1.38064852
10−23 J per Kelvin (K), or 1.38064852 * 10−16 erg per Kelvin. The physical significance of k is that it provides a measure of the amount of energy (i.e., heat) corresponding to the random thermal motions of the particles making up a substance. For a classical system at equilibrium at temperature T, the average energy per degree of freedom is k * T/ 2. In the simplest example of a gas consisting of N non-interacting atoms, each atom has three translational degrees of freedom (it can move in the x-direction, y-direction, or z-direction), and so the total thermal energy of the gas is 3 * N * k * T/2. 34 Electron charge, (symbol e), is a fundamental physical constant expressing the naturally occurring unit of electric charge; it is equal to 1.6021765 * 10−19 C, or 4.80320451 * 10−10 electrostatic unit (esu, or statcoulomb). In addition to the electron, all freely existing charged subatomic particles, thus far discovered, have an electric charge equal to this value or some whole-number multiple of it. Quarks, which are always bound within larger subatomic particles such as protons and neutrons, have charges of 1/3 or 2/3 of this value.

582

7 Energy Harvesting Projects for WSNs

The main goal is to determine whether the power harvested by the PV module is able to power the wireless sensor node; thus, it is essential to estimate the electrical power throughput of the PV module by leveraging on the relationship between the current and voltage of the PV module expressed by Eq. (7.15). From the equation, it can be deduced that the voltage drop across the series resistance, VRs ¼ IPV  Rs , is much lower than the output PV voltage, VPV, due to the very low PV current, IPV, of the order of lA flowing through the small series resistance, Rs, of few ohms. Hence, the IPV  Rs term in Eq. (7.15) can be neglected during the formulation of the output power of the solar panel, PPV(VPV) expressed as follows: h i PPV ðVPV Þ ¼ VPV  IPV ¼ VPV  IL  VPV  I0  eðVPV =ns Vt Þ  1 VPv  Isc  VPV  I0  eðVPV =ðns kTcÞ=q Þ

ð7:17Þ

Note that the term eðVPV =ns Vt Þ  1, and the light generating current, IL Isc, the short-circuit current (Celik and Acikgoz 2007). The harvested PV power, PPV(VPV), as expressed in Eq. (7.17), is formulated as a function of the PV voltage, VPV, and it can be simulated based on the technical characteristics of the PV module, i.e., ns, k, q, and Io and the environmental variables such as light irradiance, which is related to IL, hence Isc, and ambient temperature, TC, of 295 K (21.85 °C). For various light irradiance of 380 lx35 (WhatIs.com 2011b), 425, 480, 640, and 1010 lx, the respective Isc currents are measured to be 60, 65, 74, 96, and 150 lA. For this indoor research work, an amorphous type of solar panel, SC-01 (Praher 2016), is used under indoor conditions, i.e., artificial lighting from fluorescent lamps and at room temperature where the variation of temperature is relatively less than outdoors. At very low light illumination, such as G = 380 lx ( 380/120 W/ m2 = 3.17 W/m2) (Randall and Jacot 2003). The technical characteristics of the solar panel are tabulated in Table 7.6. The corresponding solar panel’s efficiency can be determined as: g¼

35

PPV  100% GA

ð7:18Þ

The lux (symbolized lx) is the unit of illuminance in the international system of units (SI). It is defined in terms of lumens per meter squared (lm/m2). One lux is the equivalent of 1.46 milliwatt (1.46 * 10−3 W) of radiant electromagnetic (EM) power at a frequency of 540 terahertz (540 THz or 5.40 * 1014 Hz), impinging at a right angle on a surface whose area is one square meter. A frequency of 540 THz corresponds to a wavelength of about 555 nm, which is in the middle of the visible-light spectrum. The lux is a small unit. An alternative unit is the watt per meter squared (W/m2). To obtain lux when the illuminance in watts per meter squared is known, multiply by 683. To obtain watts per meter squared when the illuminance in lux is known, divide by 683 or multiply by 0.00146. Illuminance varies inversely with the square of the distance from the source on a free-space line of sight. If the distance is doubled, the illuminance is cut to 1/4; if the distance increases by a factor of 10, the illuminance becomes 1/100 (0.01 times) as great.

7.2 Energy Harvesting Projects Table 7.6 Technical characteristics of solar panel used (Tan and Panda 2011)

583 Parameter

Unit

Value

Physical dimension Cross-sectional area, A Open-circuit voltage, Voc (at 380 lx) Short-circuit current, Isc (at 380 lx) MPPT voltage, Vsc (at 380 lx) MPPT current, Isc (at 380 lx)

mm cm2 V lA V lA

55
30
1 16.5 4.14 60 3.5 51.44

Electrical power (μW)

where G is the light intensity; A is the cross-sectional area. Based on Eq. (7.18), the calculated efficiency of the solar panel is found to be around 3.4%, which is lower than the outdoor solar panel efficiency (Randall 2006). Due to the low efficiency of the solar panel in the indoor condition, the power harvested is also low; hence, it is necessary to optimize the indoor solar energy harvesting system in order to maximize the power harvested from the solar panel. Further tests were carried on at different lighting conditions; Figs. 7.47 and 7.48 record the indoor solar panel’s PV and PR curves at different lux illuminations ranging from 380 to 1010 lx. Taking into consideration the solar panel parameters, the model of the solar panel expressed by Eq. (7.17) has been simulated and the simulation results plotted in Fig. 7.47. The simulated PV curves are verified against the measured PV curves, for varying solar irradiance conditions performed by a characterization setup on a fluorescent light source. It is shown that the power curves peak near the solar panel output voltage of 3.6 V. Moreover, in order to compare the solar panel and the thermoelectric generator (TEG) characteristics, the harvested power as a function of the load resistance is plotted. From the power curve (PR) of the solar panel as plotted in Fig. 7.48, it can

Voltage (V)

Fig. 7.47 PV curves of solar panel at different lux conditions (Tan and Panda 2011)

7 Energy Harvesting Projects for WSNs

Electrical power (μW)

584

Load resistance (k )

Fig. 7.48 PR experimental curves of solar panel at different lux conditions (Tan and Panda 2011)

be observed that the maximum power points (MPPs) of the solar panel vary between 27 and 68 kX. Hence, by setting the output voltage of the solar panel at a fixed 3.6 V, the maximum output power can be harvested from the solar panel under different solar irradiances. Noticeably, within the indoor lighting conditions of 380–1010 lx, the maximum electrical power that the solar panel can harvest ranges from 180 to 480 lW, respectively.

Thermal Energy Harvesting System In the thermal energy harvesting (TEH) system, a miniaturized thermoelectric generator housed in the thermal energy harvester is used to convert thermal energy into electrical energy. The thermal energy, generated from the heat source at certain high temperature TH, is channeled through the enclosed TEG via a thin film of thermally and electrically conductive silver grease connecting them to the heat sink. The residual heat accumulated in the heat sink is then released to the surrounded ambient air at a lower temperature TC. An equivalent thermal circuit model of the thermal energy harvester that illustrates its thermal and electrical characteristics is provided in Fig. 7.49. Referring to Fig. 7.49, it can be observed that the temperature difference, ΔTTEG, across the junctions of the TEG is lower than the temperature gradient, ΔT = TH − TC, that is externally imposed across the thermal energy harvester. This is due to the thermal contacts and thermal grease resistances residing in the cold and hot sides of the TEH, i.e., Rcon(H), Rcon(C) and Rg(H), Rg(C), respectively. To minimize this negative effect, the thermal resistance, RTEG, of the TEG is chosen to be as high as possible and/or conversely, the rest of the thermal resistances of the thermal energy harvester are designed to be as small as possible. Taking these design considerations into account, the miniaturized TEH, having 20 mm
20 mm

7.2 Energy Harvesting Projects

585 TH

Fig. 7.49 Equivalent electrical circuit of the thermal energy harvester (Tan and Panda 2011)

Rcon(H) Rg(H) P

Hot side, THJ TEG = P* TEG

Rs,TEG

ITEG

TEG TEG

VOC = VTEG S* TEG

RL

Cold side, TCJ Rg(C) Rcon(C) TC

20 mm physical size, is designed such that most of the heat is channeled through the TEG in order to maximize TEH. Analysis and characterization were conducted on the designed thermal energy harvester to evaluate the performance of the TEH system when powering the wireless sensor node. According to Seebeck’s effect, the open-circuit voltage, Voc, of the TEG enclosed in the TEH and composed of n thermocouples connected electrically in series and thermally in parallel, is given as: Voc ¼ S  DT ¼ n  a  ðTH  TC Þ

ð7:19Þ

where a and S represent the Seebeck’s coefficient of a thermocouple and a TEG, respectively. When connecting a load resistance, RL, electrically to the TEG via the TEH as shown in Fig. 7.49, an electrical current, ITEG, flows in accordance to the applied temperature difference, ΔT, which is given as: ITEG ¼

Voc  VTEG n  a  ðTH  TC Þ  VTEG ¼ Rs;TEG Rs;TEG

ð7:20Þ

where Rs,TEG is the internal electrical resistance of the TEG. Based on the described in Eq. (7.20) current-voltage (IV) characteristic of the TEG, the output power, PTEG(VTEG), delivered by the TEG to the load, RL, can be determined. By substituting ITEG with Eq. (7.20), the electrical power, PTEG(VTEG), harvested by the thermal energy harvester as a function of its output voltage, VTEG, is derived as:

7 Energy Harvesting Projects for WSNs

Electrical power (μW)

586

Voltage (V)

Fig. 7.50 PV curves of TEG at different thermal gradients (Tan and Panda 2011)

PTEG ðVTEG Þ ¼ VTEG  ITEG ¼

2 VTEG  n  a  ðTH  TC Þ  VTEG Rs;TEG

ð7:21Þ

Following the technical specifications provided for the Thermo Life low-power thermoelectric generator (TECTEG 2019), the TEG used in this HEH project is made up of 5200 thermocouples. Each thermocouple has a Seebeck coefficient, a, of 0.21 mV/K and its internal electrical resistance, Rs,TEG, is 82 kX. For a given temperature difference, ΔT = TH − TC, between 5 and 10 K, the model illustrated by Eq. (7.21) is simulated for different TEG’s output voltage, VTEG, as illustrated in Fig. 7.50. Experiments were carried out to characterize the TEG by applying a temperature difference between the thermal contact faces, and then measuring both the electrical output voltage and current with different loads connected. A hotplate is used to emulate the heat source, and a Fluke 52-II digital thermometer (Fluke 2016b) via its thermocouple probe is used to measure the surface temperature. The output voltage and current are measured using two separate Fluke 8845A 6.5 digit precision multimeters (Fluke 2016a); the measured parameters are used to calculate the electrical power. This experimental setup was intended to harvest energy under different thermal conditions with temperature differences in the range 5–10 K. The experimental results obtained out of the TEG are plotted in Fig. 7.50. By analyzing the power curves PV and PR in Figs. 7.48, 7.50, and 7.51 several deductions are attained: • Referring to the power curve (PV) in Fig. 7.50, it can be seen that the simulation results obtained using the model expressed in Eq. (7.21) are comparable to the measurement results collected from the characterization of the thermal energy harvester under varying temperature differences. It is clear that the maximum obtainable power for each thermal gradient corresponds to an output voltage of

587

Electrical power (μW)

7.2 Energy Harvesting Projects

Load resistance (k )

Fig. 7.51 PR experimental curves of TEG at different thermal gradients (Tan and Panda 2011)

the thermal energy harvester, i.e., PMPPT,ΔT=5 K = 96 lW at 2.8 V, PMPPT,ΔT=10 K = 247 lW at 4.5 V, etc. This is unlike the indoor solar energy harvesting case where all their power curves (PV) peak near a particular output voltage of the solar panel (Fig. 7.47). • However, for the power curves (PR) plotted for solar and thermal energy harvesting as shown in Figs. 7.48 and 7.51, respectively, a noticeable finding is distinguished. It is observed that the MPPs of the solar panel vary between load resistances in the range 27–68 kX, whereas the MPPs of the thermal energy harvester are fixed at the 82 kX internal impedance of the thermal energy harvester. • When the load resistance matches the source resistance of the thermal energy harvester, as displayed in Fig. 7.51, the harvested power is always maximum for different temperature differences. Hence, it can be concluded, from both power curves in Figs. 7.48 and 7.51, that no common MPPT approach exists between the solar and thermal energy harvesting systems.

7.2.9.2

Hybrid Energy Harvesting from Solar and Thermal Energy Sources

The concept of HEH has been newly considered in the literature (Park and Chou 2006a; Lhermet et al. 2008; Tadesse et al. 2009; Guilar et al. 2009; Khaligh et al. 2010), as a potential micro-power supply solution to minimize the size of the energy supply as well as to extend the operational lifetime of the wireless sensor node. Researchers have introduced a number of HEH methods to combine different small scale energy harvesting (EH) sources. For these HEH methods, each EH source requires a unique power management circuit to condition the power flow

588

7 Energy Harvesting Projects for WSNs

from the energy source to its output load. The concern is that the needed number of converters has to increase as much as the number of augmented energy sources. In Tan and Panda (2011), the proposed HEH system requires only one electronic power converter with a simple low power control circuitry, to condition the combined output power harvested simultaneously from the solar and thermal energy sources. By avoiding the use of different power management units for multiple energy sources, the number of components used in the HEH system is lessened, and the system’s form factor, cost, and power losses are accordingly reduced. However, the challenge faced by this approach is that there could be an impedance mismatch issue between the integrated energy sources as further discussed in the subsections to come.

Characteristics of Solar Panel and Thermal Energy Harvester Connected in Parallel For the HEH approach proposed in Tan and Panda (2011), the terminal output voltages of the solar panel, VPV, and the thermal energy harvester, VTEG, are directly connected to the load, each via a Schottky diode, DPV or DTEG, to block reverse-biased current flow. A display of the equivalent electrical circuit of the hybrid energy harvester is shown in Fig. 7.52. According to Figs. 7.47 and 7.50, the output voltages of the two energy sources are not that low, typically of few volts, hence the series energy sources configuration is not used to step-up the voltage across the load, VRL . Instead, a parallel energy sources configuration, VRL ¼ VPV þ VDPV ¼ VTEG þ VDTEG , is employed to produce more current flows, i.e., IRL ¼ IPV þ ITEG . VDPV and VDTEG are the 0.2 V forward voltage drops across each of the two series diodes. The powers harvested from the solar panel, PPV(VPV), and from the thermal energy harvester, PTEG(VTEG), expressed by Eqs. (7.17) and (7.21), respectively, are summed together; when subtracting the power losses of the two series diodes, the resultant power is used to drive the load. The electrical power throughput of the hybrid energy harvester, PHEH(VRL), as a function of its output voltage, VRL , is thus given by:

Rs ID IL

ITEG + DTEG + Rs,TEG

DPV RL

VTEG

VOC

Fig. 7.52 Equivalent electrical circuit of the proposed HEH system (Tan and Panda 2011)

7.2 Energy Harvesting Projects

589

PHEH ðVRL Þ ¼ jPPV ðVRL Þj þ jPTEG ðVRL Þj  V  V 2     RL oc;TEG  VRL  VRL =ðns kTC =qÞÞ  ð     VRL  Isc;PV  VRL  I0  e þ  Rs;TEG ð7:22Þ

Electrical power (μW)

Based on the technical specifications of the solar panel and the thermal energy harvester given in Sections “Indoor Solar Energy Harvesting System” and “Thermal Energy Harvesting System” respectively, the harvested power expression of the hybrid energy harvester, as expressed by Eq. (7.22), is simulated over a range of output voltages, VRL , for different solar irradiance and temperature differences

Voltage (V)

Electrical power (μW)

(a) AT 380 lux and T= 5

Voltage (V)

(b) AT 1010 lux and T= 10

Fig. 7.53 Experimental and simulated electrical power harvested from parallel solar and thermal energy sources (Tan and Panda 2011)

590

7 Energy Harvesting Projects for WSNs

that correspond to the solar panel’s short-circuit current, Isc,PV, and the thermal energy harvester’s open-circuit voltage, Voc,TEG. As shown in Fig. 7.53, a set of simulation and experimental results are extracted under the minimum 380 lx and ΔT = 5 K, and the maximum 1010 lx and ΔT = 10 K power harvesting conditions. From Fig. 7.53, simulation and experimentation yield several observations: • The measured power curve, (Measured S + T), of the hybrid energy harvester, is the summation of the individual power curves, i.e., solar panel (Solar (S)) and thermal energy harvester (Thermal (T)) being superimposed into the figure minus the negligible small power loss in the Schottky diodes. • The simulated waveforms (Simulated S + T) based on the model expressed by Eq. (7.22) and the measured waveforms (Measured S + T) obtained from experimentations are quite similar. This verifies the expression model, derived in Eq. (7.22), to be used for determining the electrical power throughput of the hybrid energy harvester, PHEH, so as to sustain the operational lifetime of the wireless sensor node. • At MPPT voltage, VRL ;MPPT , of 3.6 V, the output voltages of the solar panel and thermal energy harvester are slightly higher than VRL ;MPPT , such that the two isolation diodes are conducting in the forward bias condition. Hence, from (Measured S + T), it can be seen that the hybrid energy harvester can generate power at the minimum of 252 lW (PPV = 167 lW, PTEG = 85 lW) and at the maximum of 693 lW (PPV = 466 lW, PTEG = 227 lW), respectively. • When VRL  VPV, the solar panel operates in the open-circuit mode, therefore no solar power is harvested. As well, this situation happens to the thermal energy harvester if VRL  VTEG (3.6 V). More analysis and characterization works were conducted on the HEH system to evaluate its performance in powering the wireless sensor node. Referring to Figs. 7.54 and 7.55, several findings can be seen from the PR curves, (Thermal ΔT5) − (Thermal ΔT10): • The MPPs of the stand-alone thermal energy harvester are fixed at its 82 kX internal resistance under temperature differences in the range of 5–10 K. • When the thermal energy harvester is paralleled with the solar panel under 380 lx weak illumination and 1010 lx strong illumination, it can be observed from the PR curves, (Measured S + T(ΔT5)) − (Measured S + T(ΔT10)), that the MPPs are not longer fixed, but vary with the combined internal impedance of the parallel solar panel and thermal energy harvester. This is the effect of the impedance mismatch between the two energy sources of the hybrid energy harvester. Although the impedance mismatch issue arises in the proposed HEH system, it is sought possible to combine the two energy sources together, without dedicating each individual energy source with a power converter to perform MPPT, as does the AmbiMax project presented in Sect. 7.2.6. • The PV curves, (Measured S + T(ΔT5)) − (Measured S + T(ΔT10)), confirm that all the MPPs of the hybrid energy harvester are fixed at around its 3.6 V output voltage.

591

Electrical power (μW)

7.2 Energy Harvesting Projects

Voltage (V)

Electrical power (μW)

(a) PV curves

Load resistance (k )

(b) PR curves

Fig. 7.54 PV and PR curves of HEH system at fixed solar irradiance of 380 lx (3 W/m2) and different thermal conditions of 5–10 K (Tan and Panda 2011)

Further interesting experimentation:

observations

are

obtained

out

of

this

extensive

• For an illumination level of 380 lx and above in common indoor lighting condition, as revealed in Figs. 7.56 and 7.57, the hybrid energy harvester tested under fixed thermal condition of 5 and 10 K indicates that all PV power curves (Measured S + T(380 lx)) − (Measured S + T(1010 lx)) peak around 3.6 V. • Checking through all the power curves shown in Figs. 7.54, 7.55, 7.56, and 7.57, it can be observed that all the MPPs of the PV curves tend to cluster around a fixed voltage of 3.6 V, whereas the MPPs of the PR curves are scattered out in the range of 20–50 kX. Hence, by setting the terminal voltage of the

7 Energy Harvesting Projects for WSNs

Electrical power (μW)

592

Voltage (V)

Electrical power (μW)

(a) PV curves

Load resistance (k )

(b) PR curves

Fig. 7.55 PV and PR curves of HEH system at fixed solar irradiance of 1010 lx (3 W/m2) and different thermal conditions of 5–10 K (Tan and Panda 2011)

hybrid energy harvester to a value in the peak power range (VRL ;MPPT = 3.6 V), it is possible to extract maximum output power from the hybrid energy harvester with a simple and ultra-low-power control circuit, so as to place the panel at its MPPs, rather than using energy hungry tracking techniques which require high computational power and cost. There are such techniques as perturb and observe (Femia et al. 2005; Abdelsalam et al. 2011; Elgendy et al. 2012), and incremental conductance (Safari and Mekhilef 2011; Elgendy et al. 2013; Sera et al. 2013).

593

Electrical power (μW)

7.2 Energy Harvesting Projects

Voltage (V)

Electrical power (μW)

(a) PV curves

Load resistance (k )

(b) PR curves

Fig. 7.56 PV and PR curves of HEH system at fixed thermal condition of ΔT = 5 K and varying solar irradiances of 380–1010 lx (Tan and Panda 2011)

Design and Implementation of Ultra-Low Power Management Circuit The schematic diagram of a self-autonomous indoor wireless sensor node powered by the proposed HEH system and its ultra-low power management circuit is illustrated in Fig. 7.58. The designed power management circuitry with fixed voltage reference MPPT approach essentially consists of three main building blocks: • A boost converter (see Footnote 14) with MPP tracker and its control, and pulse width modulation (PWM) generation circuit that manipulates the operating point of the HEH system to keep harvesting power at near MPPs.

7 Energy Harvesting Projects for WSNs

Electrical power (μW)

594

Voltage (V)

Electrical power (μW)

(a) PV curves

Load resistance (k )

(b) PR curves

Fig. 7.57 PV and PR curves of HEH system at fixed thermal condition of ΔT = 10 K and varying solar irradiances of 380–1010 lx (Tan and Panda 2011)

• An energy storage element, a super-capacitor, to buffer the energy transfer between the source and the load. • A regulating buck converter (see Footnote 14) to provide constant voltage to the wireless sensor node and other electronic circuitries. For indoor environment, the ambient energy sources such as solar and thermal gradient are not available at all times or at a steady level. Hence, there is a need to incorporate a super-capacitor as an energy storage device in the HEH system to store the excessive energy harvested from the solar panel and/or thermal energy harvester when energy sources are unavailable. Moreover, by drawing power simultaneously from both solar and thermal energy sources, the throughput power of the HEH system is increased, which enhances the performance of the indoor

7.2 Energy Harvesting Projects

595 Boost converter with voltage reference MPPT

Hybrid energy harvesting source IPV Rs

L (100 mH) 10 M

DPV 3.3 F

IL

D

V0 +

NMOS Si1563DH

10 M

PWM

Energy storage

Regulating buck converter 2.8 SYNC SW 10 H VIN 22 F RUN

Supercapacitor (0.1 F, 5.5 V)

LTC1877 ITH VFB GND

ITEG Voc

Rs,TEG

DTEG

47 F

1M 470 k

Wireless sensor node TI CC2500

VRopt Vfb VMPPT

180 k Verr PI controller

4.7 F

Microcontroller TI MSP430F2274 (low power mode, 100 kHz)

LTC6906 Pulses V+ OUT 18 k DIV

GND SET

1M

LMC7215 15 nF

PWM generation (10 kHz)

Fig. 7.58 Functional block diagram of HEH system (Tan and Panda 2011)

wireless sensor node. A super-capacitor is preferred in this project because it has relatively favorable characteristics over batteries as described in Simjee and Chou (2008) and Sect. 2.2.4 of this book. These characteristics include more than half a million full charge cycles, long lifetime of 10–20 years of operation, and power density of an order of magnitude higher than a battery. Unlike the discrete capacitors, which have very small capacitance values of pF–lF range, the super-capacitors have very large capacitance values of Farads, which suits well energy storage for a long duration. The Linear Technology LTC1877 switched-mode voltage regulator (Linear Technology 2000c) is inserted after the super-capacitor to provide a 2.8 V constant DC operating voltage (VDC) to the wireless sensor node and other electronic circuitries. The efficiency of the regulating buck converter is experimentally tested and was about 80–90% while consuming an operating current of 12 lA. The operation of the wireless sensor node deployed in an application field comprises two functions: • Sensing some external analog signals from sensory devices such as temperature, and humidity. • Communicating and relaying the sensed information to the gateway node every 5 s interval. Upon receiving the data at the basestation, the collected data are processed into usable information for any follow up action.

596

7.2.9.3

7 Energy Harvesting Projects for WSNs

Experimentation Outcomes

The near-optimal HEH wireless sensor node has been successfully implemented as a hardware prototype for laboratory testing. Several experimental tests have been conducted to analyze the performance of the HEH system and its simple ultra-low power fixed reference voltage MPPT scheme that powers the connected load consisting of the super-capacitor, the sensing, control and PWM generation circuitries, and the wireless sensor node.

Performance of Parallel HEH Configuration As mentioned previously, when the differently characterized solar and thermal energy sources are combined together, there is an impedance mismatch among the integrated energy sources. For accurate insight, the performance of the parallel hybrid energy harvester, which comprises the combined characteristics of the solar panel as well as the thermal energy harvester, is investigated: • Referring back to Figs. 7.54, 7.55, 7.56, and 7.57, it is illustrated that the fixed reference voltage method is able to operate the hybrid energy harvester near its MPPs for different light intensities and temperature differences, but at the expense of some percentage loss in the harvested power. It is thus important to examine the significance of these power differences between the actual harvested power, PHEH,actual, with respect to the MPPs, PHEH,MPPT, of the hybrid energy harvester. • Thus considered, an example extreme operating condition, which is at low light illumination of 380 lx and small temperature difference of 5 K. It was found that the power, PHEH,actual, harvested at the fixed 3.6 V reference MPP voltage, and the maximum obtainable power, PHEH,MPPT, of the hybrid energy harvester, are 252 and 260 lW, respectively. Noticeably, the power difference is only 8 lW, which is about 3% of its harvested power. This 8 lW power loss is due to the impedance mismatch issue between the solar panel and the thermal energy harvester when they are connected directly without the use of separate power converters. • Similarly, it is realized that the power differences between the actual harvested power, PHEH,actual, with respect to the MPPs, and the maximum obtainable power, PHEH, MPPT, for all the other operating conditions, range between 8 and 35 lW, which is 3–6% of the harvested power. Although the proposed HEH power management unit would incur such power loss in the overall harvested power, this power loss is significantly small compared to other MPPT techniques that require higher computational power and cost to fulfill their objective of precise and accurate MPP tracking. It is thus justifiable to utilize the simple and ultra-low power fixed reference voltage method for the HEH. • For indoor applications like hospitals and factories, the ambient conditions are 1010 lx solar irradiance and 10 K temperature difference. Going back to

7.2 Energy Harvesting Projects

597

Figs. 7.47 and 7.50 with operating conditions of 1010 lx solar irradiance and 10 K temperature difference, the maximum power obtained by summing the individual MPP of the thermal energy harvester, PTEG, and solar panel, PPV, is 727 lW, and the actual harvested power, PHEH,actual, measured from the two paralleled energy sources is 690 lW. Also obtained, a 35 lW power difference between the calculated and measured powers, due to impedance mismatch amongst the two paralleled energy sources.

Power Conversion Efficiency of the HEH System Other than the regulating buck converter, there are two main contributors of power losses to the HEH system; namely the boost converter itself, which acts as a MPP tracker, and its associated sensing, control, and PWM generation circuits. The efficiency of the boost converter, ηconv, can be expressed as a function of its output load power, Pload, over its input DC power, PDC, under varying solar irradiance, temperature differences, ΔT, and loading, RL, conditions. For a lux illumination, temperature difference, and output load resistance, the efficiency of the boost converter is given by: gconv ¼

Pout V 2 =Rload  100% ¼ out  100% Pin Vin  Iin

ð7:23Þ

As an example, for 380 lx illumination, 5 K temperature difference, and 68 KX output load resistance, the efficiency of the boost converter is given by: ¼

4:95V 2 =68 KX  100% ¼ 91:8% 3:6 V  109 lA

Expressive observations, as plotted in Figs. 7.59 and 7.60, are acquired when calculating the efficiencies of the boost converter using Eq. (7.23) for other temperature differences, solar irradiance, and loading conditions: • The efficiency of the designed boost converter ranges between 80 and 94% over a 50–330 kX load resistances range. At heavy load condition, roughly 50 kX, signifying the discharge state of the super-capacitor, it can be seen that the efficiency of the boost converter is high, about 94%. This high efficiency boost converter is favorable and desirable to ensure optimal transfer of energy from the hundreds of microwatts micro-power sources, or lower, to the energy storage. • As the loading gets lower with the super-capacitor charging up, it can be seen that the efficiency of the converter decreases to nearby 82% at a load resistance of 300 kX. This decreasing efficiency trend is due to the power loss in the boost converter.

7 Energy Harvesting Projects for WSNs Efficiency of boost converter (%)

598

Load resistance (k )

Efficiency of boost converter (%)

Fig. 7.59 Efficiency of HEH boost converter at fixed voltage reference-based MPPT and varying temperature difference (Tan and Panda 2011)

Load resistance (k )

Fig. 7.60 Efficiency of HEH boost converter at fixed voltage reference-based MPPT and varying solar irradiance (Tan and Panda 2011)

• Even though the efficiency at light load is lower, it is not as critical as the heavy load condition because the super-capacitor, by then, is already near full charge state and any surplus energy would not be stored. • For indoor applications like hospitals and factories, and taking into consideration both the power difference and the power losses in the voltage regulating and MPPT converters as shown in Figs. 7.59 and 7.60, the net harvested power output to power the indoor wireless sensor node, through the boost converter with 90% efficiency, is 621 lW. This harvested power from the HEH system is more than what is harvested by the single-source energy harvesting system, either from 432 lW ambient light or 223 lW thermal energy source. For a fully charged 0.1 F 5.5 V super-capacitor, rechargeable either from ambient light or thermal energy source, the wireless sensor node with an average power

7.2 Energy Harvesting Projects

599

consumption of 1 mW can last for nearly 0.74 or 0.54 h, respectively. However, by using the proposed HEH system, the wireless sensor node lifetime is increased, by roughly two times, to 1.11 h. Another worth consideration source of power loss in the HEH system is the power consumption of the associated electronic circuits for sensing, Psense, control, Pctrl, and PWM generation, Pgenerate. Based on the voltage and current requirements of each individual component in the HEH system shown in Fig. 7.58, the total power consumption of the electronic circuits can be calculated as follows: Pconsumed ¼ Psense þ Pctrl þ PPWMgenerate ¼ 2:7 V  ð3 lA þ 15 lA þ 32 lAÞ ¼ 135 lW

ð7:24Þ

Once all the power losses in the HEH system are identified, the performance of the designed HEH system, for enhanced performance indoor wireless sensor node, might be smartly evaluated. These power losses include the power difference factor due to impedance mismatch between two paralleled energy sources, and the power losses in the voltage regulating and MPPT converters.

Concluding Recap A near optimal HEH system has been proposed to enhance the performance of indoor wireless sensor node. Theoretical studies on individual as well as hybrid solar and thermal energy harvesting systems were conducted and simulated to understand the characteristic of the HEH system; subsequent experimentation verified the obtained results. The proposed HEH system using one power management circuit was successfully implemented into hardware prototype for laboratory testing. Based on the power analysis, the efficiency of the power management unit, with fixed reference voltage-based MPPT scheme, is about 90%; its sensing, control, and PWM generation circuitries consumption was around 135 lW. Experimental results disclosed that the HEH system can harvest an average electrical power of 621 lW from both energy sources at 1010 lx average solar irradiance and 10 K thermal gradient. Such harvested power is almost three times higher than the conventional single thermal energy harvesting, which helps enhancing the performance of the indoor wireless sensor node.

7.3

Conclusion for Radiance

Sensor networks with battery-powered nodes cannot meet in a single shot the design goals of lifetime, cost, sensing reliability, sensing, as well as transmission coverage. With a finite energy source, these performance parameters can be hardly optimized simultaneously; unavoidably, higher battery capacity causes increased

600

7 Energy Harvesting Projects for WSNs

cost, low duty-cycle leads to decreased sensing reliability, higher transmission range goes toward higher power requirement, and lower transmission range implies transmission paths with more number of hops resulting in energy usage at more number of nodes. The harnessed electrical energy powers the sensor nodes; if the harvested energy source is large, and periodically or continuously available, a sensor node can be powered almost perpetually. Based on the periodicity and magnitude of harvestable energy, system parameters of a node such as sampling frequency, duty-cycle, and transmission power can be tuned to increase node and network performance. Since a node is energy-limited just till the next harvesting opportunity (recharge cycle), it can optimize its energy usage to maximize performance during that interval. For instance, a node can increase its sampling frequency or its duty-cycle to increase sensing reliability, or increase transmission power to decrease length of routing paths. As a result, energy harvesting techniques have the potential to address the tradeoff between performance parameters and lifetime of sensor nodes. The challenge lies in estimating the periodicity and magnitude of the harvestable source and dynamically deciding which parameters to tune while avoiding energy depletion before the next recharge cycle. The capability of a wireless sensor node to harvest energy might simultaneously address the conflicting design goals of lifetime and the aforementioned performance metrics. This chapter covers in full details the energy harvesting projects that innovated ideas and techniques to prolong the lifetime of wireless sensor nodes by benefiting from energy available in the environment. These projects are models that highlight the basic concepts of harvesting systems, such as energy harvesting architectures, types of harvestable energy sources, storage technologies, and hardware support that encompasses microcontrollers, voltage regulators, and interfaces. Table 7.7 assembles and compares the projects presented. By studying and detailing inventive energy harvesting projects and the underlying concepts and technologies, this chapter motivates further research and industry activities toward the usage of energy harvesting wireless sensor nodes and promoting their applications. The upcoming exercises offer several leads for fostering projects and research trends. Going all along this target, the next chapter focuses on energy management projects for WSNs built upon the techniques thoroughly presented in Part II of this book.

7.3 Conclusion for Radiance

601

Table 7.7 Energy harvesting projects compared Platform

Application

Energy harvesting source

Power source

Autonomy of harvesting control (Yes/ No)

MPPT

ZebraNet (Sect. 7.2.1) Prometheus (Sect. 7.2.2)

Monitoring wildlife General

Solar

Li-ion battery

Yes

Yes

Solar

No

Yes

Solar Biscuit (Sect. 7.2.3) Fleck1 (Sect. 7.2.3) Heliomote (Sect. 7.2.4) Everlast (Sect. 7.2.5) AmbiMax (Sect. 7.2.6)

Environmental monitoring General

Solar

Super-capacitor, Li-polymer battery Super-capacitor

No

No

No

No

Ecosystem sensing General

Solar

Yes

No

No

Yes

Yes

Yes

No

No

Sunflower (Sect. 7.2.7)

Solar

Solar

General

Solar, wind

General MSEMS platforms Forest watersheds

Solar, laser

Two AA NiMH batteries Two AA NiMH batteries Super-capacitor Super-capacitor, Li-polymer battery Super-capacitor

Micro-solar power Solar Two AA NiMH No No sensor network batteries (Sect. 7.2.8) Energy harvesting Indoor Ambient Super-capacitor Yes Yes from hybrid light, indoor sources thermal (Sect. 7.2.9) energy Legend All projects use the harvest-store-use energy harvesting architecture (Sect. 2.2.1 in Chap. 2)

7.4

Exercises

1. Differentiate between MEMS, MSEMS, and WSNs. 2. Compare the environmental monitoring projects detailed in this chapter. 3. Find and study more energy monitoring projects. Fill in Table 7.7 entries with the obtained findings. 4. Identify and compare the projects built on battery storage. 5. Identify and compare the projects built on super-capacitor storage. 6. Identify and compare the projects built on tiered storage. 7. Identify and compare the projects that focus on hybrid energy harvesting. 8. Identify the indoor energy harvesting projects presented in this chapter.

602

7 Energy Harvesting Projects for WSNs

9. Identify and compare the projects based on the MPPT. 10. What are the concepts that characterize each of the projects described in this chapter? What are their weaknesses? 11. How should the microcontrollers selection be justified when comparing the projects presented along this chapter? Would it differ if other choices were made? 12. Identify and compare the voltage regulators used in the projects illustrated in this chapter.

References Abdelsalam, A.K., A.M. Massoud, S. Ahmed, and P.N. Enjeti. 2011. “High-Performance Adaptive Perturb and Observe MPPT Technique for Photovoltaic-Based Microgrids.” Transactions on Power Electronics (IEEE) 26 (4): 1010–1021. ACM/IEEE. 2005. The 4th International Symposium on Information Processing in Sensor Networks (IPSN’05). ACM SIGBED and IEEE Signal Processing Society. http://ipsn.acm.org/ 2005/. Accessed 16 Apr 2016. Ad Hoc Electronics. 2007. 9XStream—900 MHz 100 mW. MaxStream, Inc. http://www. adhocelectronics.com/9XStream-900MHz-100mW. Accessed 16 Aug 2016. Aerogel. 2016. Organic and Carbon Aerogels. Aerogel.org. http://www.aerogel.org/?p=71. Accessed 4 Apr 2016. AirBorn Electronics. 2013. FR-4 PCB Laminate = Flame Retardant. AirBorn Electronics. May 1, 2013. http://airborn.com.au/method/fr4pcb.html. Accessed 17 Oct 2016. Analog Devices. 2000. Low-Cost 2g Dual-Axis Accelerometer with Duty Cycle output. Analog Devices, Inc. https://www.sparkfun.com/datasheets/ADXL/ADXL202E_a.pdf. Accessed 25 May 2016. Arch Rock Corporation. 2015. Arch Rock Primer Pack/IP—Product Datasheet. http://www.cs. tufts.edu/comp/150CB/papers/primerpack-ip_datasheet.pdf. Accessed 14 Jan 2016. Atmel. 2008. AT45DB041B: 4-Megabit 2.5-Volt or 2.7-Volt DataFlash. Atmel Corporation. http:// www.atmel.com/images/doc3443.pdf. Accessed 16 Aug 2016. Atmel. 2011. 8-bit Atmel Microcontroller with 128KBytes In-System Programmable Flash. Atmel. http://www.atmel.com/Images/doc2467.pdf. Accessed 21 Mar 2014. Ayers, J.E. 2009. Digital Integrated Circuits: Analysis and Design. Washington, D.C.: CRC Press, Taylor & Francis Group. Balaz, P. 2013. Versatile Sample & Hold Circuit for Industrial and T&M Applications. Texas Instruments, Inc. http://www.ti.com/lit/ug/tidu155a/tidu155a.pdf. Accessed 26 May 2016. Cadence. 2016a. Allegro Design Entry Capture and Capture CIS. Cadence Design Systems, Inc. https://www.cadence.com/content/cadence-www/global/en_US/home/tools/pcb-design-andanalysis/design-authoring/allegro-design-entry-capture-cis.html. Accessed 8 Oct 2016. Cadence. 2016b. OrCAD PCB Designer. Cadence Design Systems, Inc. http://www.orcad.com/ products/orcad-pcb-designer/overview. Accessed 8 Oct 2016. Celik, A.N., and N. Acikgoz. 2007. “Modelling and Experimental Verification of the Operating Current of Mono-Crystalline Photovoltaic Modules Using Four- and Five-Parameter Models.” Applied Energy 84 (1): 1–15. Coates, E. 2016. Buck-Boost Converters. Buck-Boost Converters. Accessed 16 Apr 2016. Corelis. 2016. JTAG. EWA Company. http://www.corelis.com/education/JTAG_Tutorial.htm. Accessed 12 Oct 2016.

References

603

Corke, P., P. Valencia, P. Sikka, T. Wark, and L. Overs. 2007. “Long-Duration Solar-Powered Wireless Sensor Networks.” In Proceedings of the 4th Workshop on Embedded Networked Sensors (EmNets), 33–37. Cork, Ireland: ACM. Crossbow. 2002a. MICA2. http://www.eol.ucar.edu/isf/facilities/isa/internal/CrossBow/ DataSheets/mica2.pdf. Accessed 3 Feb 2014. Crossbow. 2002b. MICA2DOT. http://www.eol.ucar.edu/isf/facilities/isa/internal/CrossBow/ DataSheets/mica2dot.pdf. Accessed 5 Feb 2014. Crossbow. 2007. MTS/MDA Sensor Board Users Manual. http://www.investigacion.frc.utn.edu.ar/ sensores/Equipamiento/Wireless/MTS-MDA_Series_Users_Manual.pdf. Accessed 7 Jan 2014. Day, M. 2002. Understanding Low Drop Out (LDO) Regulators. Texas Instruments, Inc. http:// focus.ti.com/download/trng/docs/seminar/Topic%209%20-%20Understanding%20LDO% 20dropout.pdf. Accessed 10 Oct 2016. Dictionary.com. 2016. Butyl. Dictionary.com. http://www.dictionary.com/browse/butyl. Accessed 1 Aug 2016. Dutta, P., et al. 2006. “Trio: Enabling Sustainable and Scalable Outdoor Wireless Sensor Network Deployments.” In The 5th International Conference on Information Processing in Sensor Networks (IPSN), 407–415. Nashville, TN: ACM/IEEE. Dyo, V., et al. 2010. “Evolution and Sustainability of a Wildlife Monitoring Sensor Network.” In The 8th ACM Conference on Embedded Networked Sensor Systems (SenSys), 127–140. Zurich, Switzerland: ACM. Electropaedia. 2005. Electropaedia Battery Technology Glossary. Woodbank Communications Ltd. http://www.mpoweruk.com/glossary.htm#acceptance. Accessed 21 Apr 2016. Elgendy, M.A., B. Zahawi, and D.J. Atkinson. 2012. “Assessment of Perturb and Observe MPPT Algorithm Implementation Techniques for PV Pumping Applications.” Transactions on Sustainable Energy (IEEE) 3 (1): 21–33. Elgendy, M.A., B. Zahawi, and D.J. Atkinson. 2013. “Assessment of the Incremental Conductance Maximum Power Point Tracking Algorithm.” Transactions on Sustainable Energy (IEEE) 1 (4): 108–117. Encyclopædia Britannica. 2011. Electron Charge. Encyclopædia Britannica, Inc. https://www. britannica.com/science/electron-charge. Accessed 25 Dec 2016. Encyclopædia Britannica. 2015. Boltzmann Constant. Encyclopædia Britannica, Inc. https://www. britannica.com/science/Boltzmann-constant. Accessed 25 Dec 2016. Encyclopædia Britannica. 2016. Luminous Intensity. Encyclopædia Britannica, Inc. http://www. britannica.com/science/luminous-intensity. Accessed 26 May 2016. Esram, T., and P.L. Chapman. 2007. “Comparison of Photovoltaic Array Maximum Power Point Tracking Techniques.” Transactions on Energy Conversion (IEEE) 22 (2): 439–449. Femia, N., G. Petrone, and G. Spagnuolo. 2005. “Optimization of Perturb and Observe Maximum Power Point Tracking Method.” Transactions on Power Electronics (IEEE) 20 (4): 963–973. Fluke. 2016a. 8845A/8846A 6.5 Digit Precision Multimeters. Fluke Inc. http://en-us. uke.com/ products/bench-instruments/8845a-8846a-6-5-digit-precision-multimeters.html. Accessed Dec 2016. Fluke. 2016b. Fluke 52 II Dual Input Digital Thermometer. Fluke Inc. http://en-us. uke.com/ products/thermometers/ uke-52-ii-thermometer.html. Accessed 26 Dec 2016. Greek Mythology. 2016a. Helios. Greek Mythology. http://www.greekmythology.com/Other_ Gods/Helios/helios.html. Accessed 20 Apr 2016. Greek Mythology. 2016b. Prometheus. Greek Mythology. https://www.google.com/search?client= safari&rls=en&q=prometheus&ie=UTF-8&oe=UTF-8. Accessed 10 Mar 2016. Guilar, N.J., T.J. Kleeburg, A. Chen, D.R. Yankelevich, and R. Amirtharajah. 2009. “Integrated Solar Energy Harvesting and Storage.” Transactions on Very Large Scale Integration (VLSI) Systems (IEEE) 17 (5): 92–101. HAMAMATSU. 2014. Si Photodiodes—S1087/S1133 Series. HAMAMATSU PHOTONICS K. K., Solid State Division. March 1, 2014. https://www.hamamatsu.com/resources/pdf/ssd/ s1087_etc_kspd1039e.pdf. Accessed 1 Jan 2016.

604

7 Energy Harvesting Projects for WSNs

Hofmann-Wellenhof, B., H. Lichtenegger, and J. Collins. 2013. Global Positioning System: Theory and Practice. Wien: Springer-Verlag Wien GmbH. IMDb. 2012. Prometheus. http://www.imdb.com/title/tt1446714/. Accessed 10 Mar 2016. Jaycar Electronics. 2004. IDC Cable, Headers & Connectors. Jaycar Electronics. http://www. jaycar.com/images_uploaded/IDCconnJ.pdf. Accessed 15 Jan 2016. Jiang, X., J. Polastre, and D. Culler. 2005. “Perpetual Environmentally Powered Sensor Networks.” In Fourth International Symposium on Information Processing in Sensor Networks (IPSN), 463–468. Los Angeles, CA: ACM/IEEE. Jointiff Limited. 2016. WindLab Junior. Jointiff Limited. http://jointiff.com/wind-turbines/ windlab-junior/. Accessed 23 Feb 2016. Juang, P., H. Oki, Y. Wang, M. Martonosi, L.S. Peh, and D. Rubenstein. 2002. “Energy-Efficient computing for Wildlife Tracking: Design Tradeoffs and Early Experiences with ZebraNet.” In The 10th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-X), 96–107. San Jose, CA: ACM. Keck Hydrowatch Center. 2008. About the Project. http://hydrowatch.cs.berkeley.edu/Welcome. html. Accessed 15 Jan 2016. Khaligh, A., P. Zeng, and C. Zheng. 2010. “Kinetic Energy Harvesting Using Piezoelectric and Electromagnetic Technologies-State of the Art.” Transactions on Industrial Electronics (IEEE) 57 (3): 850–860. Kuperman, A., and I. Aharon. 2011. “Battery–Ultracapacitor Hybrids for Pulsed Current Loads: A Review.” Renewable and Sustainable Energy Reviews (Elsevier) 15 (2): 981–992. Levis, P., Neil Patel, David Culler, and S. Shenker. 2004. “Trickle: A Self-Regulating Algorithm for Code Propagation and Maintenance in Wireless Sensor Networks.” In The First Symposium on Networked Systems Design and Implemention (NSDI). San Francisco, CA: USENIX. Lhermet, H., C. Condemine, M. Plissonnier, R. Salot, P. Audebert, and M. Rosset. 2008. “Efficient Power Management Circuit: from Thermal Energy Harvesting to above-IC Microbattery Energy Storage.” Journal of Solid State Circuits (IEEE) 43 (1): 246–255. Linear Technology Corporation. 1996. LTC1440/LTC1441/LTC1442: Ultralow Power Single/ Dual Comparator with Reference. Linear Technology Corporation. http://cds.linear.com/docs/ en/datasheet/144012fd.pdf. Accessed 23 Feb 2016. Linear Technology. 2000a. LTC3402: 2A, 3 MHz Micropower Synchronous Boost Converter. Linear Technology Corporation. http://cds.linear.com/docs/en/datasheet/3402fb.pdf. Accessed 23 Feb 2016. Linear Technology. 2000b. LTC1751/LTC1751-3.3/LTC1751-5: Micropower, Regulated Charge Pump DC/DC Converters. Linear Technology Corporation. http://cds.linear.com/docs/en/ datasheet/1751f.pdf. Accessed 17 Jan 2016. Linear Technology. 2000c. LTC1877: High Efficiency Monolithic Synchronous Step-Down Regulator. Corporation, Linear Technology. http://cds.linear.com/docs/en/datasheet/1877fb. pdf. Accessed 27 Dec 2016. Linear Technology. 2001a. LTC3400/LTC3400B: 600 mA, 1.2 MHz Micropower Synchronous Boost Converter in ThinSOT. Linear Technology Corporation. http://cds.linear.com/docs/en/ datasheet/3400fa.pdf. Accessed 19 Aug 2016. Linear Technology. 2001b. LTC3401: 1A, 3 MHz Micropower Synchronous Boost Converter. Linear Technology Corporation. http://cds.linear.com/docs/en/datasheet/3401fb.pdf. Accessed 23 Feb 2016. Linear Technology. 2001c. LTC3440: Micropower Synchronous Buck-Boost DC/DC Converter. Linear Technology Corporation. http://cds.linear.com/docs/en/datasheet/3440fc.pdf. Accessed 16 Aug 2016. Liu, T., C.M. Sadler, P. Zhang, and M. Martonosi. 2004. “Implementing Software on Resource-constrained Mobile Sensors: Experiences with Impala and ZebraNet.” In The 2nd International Conference on Mobile Systems, Applications, and Services (MobiSys), 256–269. Boston, MA: ACM.

References

605

Maxim Integrated. 2009. Current Monitor and Accumulator with Integrated Sense Resistor. Maxim Integrated Products. https://datasheets.maximintegrated.com/en/ds/DS2741.pdf. Accessed 20 Apr 2016. Maxim Integrated. 2019. Low-Voltage, Single-Supply Dual SPST/SPDT Analog Switches. Maxim Integrated Products. https://www.maximintegrated.com/en/products/analog/analog-switchesmultiplexers/MAX4544.html. Accessed 1 June 2019. Messenger, R., and A. Abtah. 2010. Photovoltaic Systems Engineering. Boca Raton, FL: CRC Press. Microchip. 2000. PIC18 Microcontroller Family. http://ww1.microchip.com/downloads/en/ DeviceDoc/30327b.pdf. Accessed 26 Jan 2014. Microchip. 2007. TC9400/9401/9402: Voltage-to-Frequency/Frequency-to-Voltage Converters. Inc. Microchip Technology. http://ww1.microchip.com/downloads/en/DeviceDoc/21483d.pdf. Accessed 23 Feb 2016. Microchip. 2013. PIC16F7X7: 28/40/44-Pin, 8-Bit CMOS Flash Microcontrollers with 10-Bit A/D and NanoWatt Technology. Microchip Technology, Inc. http://ww1.microchip.com/ downloads/en/DeviceDoc/30498D.pdf. Accessed 25 May 2016. Minami, M., T. Morito, H. Morikawa, and T. Aoyama. 2005. “Solar Biscuit: A Battery-Less Wireless Sensor Network System for Environmental Monitoring Applications.” In The 2nd International Workshop on Networked Sensing Systems (INSS). San Diego, CA: TRF. Montenegro, G., N. Kushalnagar, J. Hui, and D. Culler. 2007. Transmission of IPv6 Packets over IEEE 802.15.4 Networks-RFC4944. https://tools.ietf.org/html/rfc4944. Accessed 29 Aug 2013. Moteiv. 2006. Tmote Sky Data Sheet: Ultra Low Power IEEE 802.15.4 Compliant Wireless Sensor Module. http://www.eecs.harvard.edu/*konrad/projects/shimmer/references/tmote-skydatasheet.pdf. Accessed 3 Jan 2014. Mpala Research Centre and Mpala Wildlife Foundation. 2009. About Mpala. Mpala Research Centre and Mpala Wildlife Foundation. http://www.mpala.org/About_Mpala.php. Accessed 20 Aug 2016. National Geographic. 2019. Plains Zebra. National Geographic. https://www.nationalgeographic. com/animals/mammals/p/plains-zebra/. Accessed 1 June 2019. National Semiconductor. 2000. LP2982: Micropower 50 mA Ultra Low-Dropout Regulator. National Semiconductor Corporation. http://pdf1.alldatasheet.com/datasheet-pdf/view/9264/ NSC/LP2982AIBP-2.8.html. Accessed 16 Aug 2016. NEC/TOKIN. 2013. Super Capacitors. NEC/TOKIN. https://www.nec-tokin.com/english/product/ pdf_dl/supercapacitors.pdf. Accessed 19 June 2016. Nelms, R.M., and J.E. Schatz. 1992. “A Capacitor Charging Power Supply Utilizing a Ward Converter.” Transactions on Industrial Electronics (IEEE) 39 (5): 421–428. Nordic Semiconductor. 2004. nRF2401: Single Chip 2.4 GHz Transceiver. Nordic Semiconductor ASA. https://www.sparkfun.com/datasheets/RF/nRF2401rev1_1.pdf. Accessed 25 May 2016. Nordic Semiconductor. 2016. Enhanced ShockBurst User Guide. Nordic Semiconductor. https:// infocenter.nordicsemi.com/index.jsp?topic=%2Fcom.nordic.infocenter.sdk5.v11.0.0%2Fesb_ users_guide.html. Accessed 1 June 2019. Nordic VLSI. 2002. nRF903: 430 MHz-950 MHz Single Chip RF Transceiver. Nordic VLSI ASA. http://s3.amazonaws.com/zanran_storage/www.sunstare.com/ContentPages/19262394. pdf. Accessed 23 July 2016. NXP Semiconductors. 2004. 1N4148;1N4448: High-Speed Diodes. NXP Semiconductors. http:// www.nxp.com/documents/data_sheet/1N4148_1N4448.pdf. Accessed 23 Feb 2016. Panasonic. 1998. Solar Cells Technical Handbook ‘98/99. Matsushita Battery Industrial Co., Ltd. http://www.solarbotics.net/library/datasheets/sunceram.pdf. Accessed 4 Apr 2016. Panasonic. 2016a. Conductive Polymer Aluminum Solid Capacitors (OS-CON). Panasonic Corporation. https://industrial.panasonic.com/ww/products/capacitors/polymer-capacitors/oscon. Accessed 18 Aug 2016. Panasonic. 2016b. Capacitors: Electric Double Layer Capacitors (Gold). Panasonic Corporation of North America. https://na.industrial.panasonic.com/node/6045. Accessed 23 Feb 2016.

606

7 Energy Harvesting Projects for WSNs

Park, C., and P.H. Chou. 2006a. “AmbiMax: Autonomous Energy Harvesting Platform for Multi-supply Wireless Sensor Nodes.” In The 3rd Annual IEEE Communications Society on Sensor and Ad Hoc Communications and Networks (SECON), 168–177. Reston, VA: IEEE, 2006. Park, C., and P.H. Chou. 2006b. “Eco: Ultra-Wearable and Expandable Wireless Sensor Platform.” In International Workshop on Wearable and Implantable Body Sensor Networks (BSN), 162–165. Cambridge, MA: IEEE. Polastre, J., J. Hill, and D. Culler. 2004. “Versatile Low Power Media Access for Wireless Sensor Networks.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 95–107. Baltimore, MD: ACM. Polastre, J., R. Szewczyk, and D. Culler. 2005. “Telos: Enabling Ultra-Low Power Wireless Research.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN), 364–369. Los Angeles, CA: ACM/IEEE. PostgreSQL. 2013. PostgreSQL 2013-12-05 Update Release. http://www.postgresql.org. Accessed 7 Jan 2014. Praher. 2016. Operating Instruction: Solar Set SC 01. Praher. http://www.peraqua.com/uploads/ products/Product_0253/downloads/desc%20002.pdf. Accessed 18 Dec 2016. Radio-Electronics.com. 2016. PIN diode. Adrio Communications Ltd. http://www.radioelectronics.com/info/data/semicond/pin_diode/p-i-n_diode.php. Accessed 8 Oct 2016. Raghunathan, V., A. Kansal, J. Hsu, J. Friedman, and M. Srivastava. 2005. “Design Considerations for Solar Energy Harvesting Wireless Embedded Systems.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN), 457–462. Los Angeles, CA: ACM SIGBED and IEEE Signal Processing Society. Randall, J. 2006. Designing Indoor Solar Products. Chichester, West Sussex: Wiley. Randall, J.F., and J. Jacot. 2003. “Is AM1.5 Applicable in Practice? Modelling Eight Photovoltaic Materials with Respect to Light Intensity and Two Spectra.” Renewable Energy 28 (12): 1851– 1864. Rao, A., W. McIntyre, U.-K. Moon, and G.C. Temes. 2005. “Noise-Shaping Techniques Applied to Switched-Capacitor Voltage Regulators.” Journal of Solid-State Circuits (IEEE) 40 (2): 422–429. Ridley, R. 2000. Second-Stage LC Filter Design. In Switching Power Magazine (Ridley Engineering), 1–10. ROHM. 2016. Standard Voltage Detectors—Bd4835G. ROHM Semiconductor. http://www.rohm. com/web/global/products/-/product/BD4835G. Accessed 11 July 2016. Roundy, S., and P.K. Wright. 2004. “A Piezoelectric Vibration based Generator for Wireless Electronics.” Smart Materials and Structures (IOP Publishing) 13 (5): 1131–1142. Roundy, S., D. Steingart, L. Frechette, P. Wright, and J. Rabaey. 2004. Power Sources for Wireless Sensor Networks. In Wireless Sensor Networks, vol. 2920, ed. H. Karl, A. Wolisz, and A. Willig, 1–17. Berlin, Heidelberg: Springer. Safari, A., and S. Mekhilef. 2011. “Simulation and Hardware Implementation of Incremental Conductance MPPT With Direct Control Method Using Cuk Converter.” Transactions on Industrial Electronics (IEEE) 58 (4): 1154–1161. Schweber, B. 2017. Understanding the Advantages and Disadvantages of Linear Regulators. Digi-Key Electronics. https://www.digikey.com/en/articles/techzone/2017/sep/understandingthe-advantages-and-disadvantages-of-linear-regulators. Accessed 1 June 2019. Sensirion. 2016. SHT1x—Digital Humidity & Temperature Sensor (RH/T). Sensirion AG Switzerland. https://www.sensirion.com/products/digital-humidity-sensors-for-reliablemeasurements/digital-humidity-sensors-for-accurate-measurements/. Accessed 15 Jan 2016. Sensirion. 2019. Digital Humidity Sensor SHT1x (RH/T). Sensirion. https://www.sensirion.com/ en/environmental-sensors/humidity-sensors/digital-humidity-sensors-for-accuratemeasurements/. Accessed 2 June 2019. Sera, D., R. Teodorescu, and P. Rodriguez. 2007. “PV Panel Model Based on Datasheet Values.” International Symposium on Industrial Electronics (ISIE), 2392–2396. Vigo, Spain: IEEE.

References

607

Sera, D., L. Mathe, and T. Kerekes. 2013. “On the Perturb-and-Observe and Incremental Conductance MPPT Methods for PV Systems.” Journal of Photovoltaics (IEEE) 3 (3): 1070– 1078. Shibaura Institute of Technology. 2016. News & Topics. http://www.shibaura-it.ac.jp/en/. Accessed 17 June 2016. Silicon Designs. 2016. High-Accuracy Industrial-Grade MEMS Capacitive Accelerometers. Silicon Designs. http://www.silicondesigns.com. Accessed 23 Feb 2016. Silicon Solar. 2015. 4 V 100 mA OEM Solar Panels. Silicon Solar. http://www.siliconsolar.com/ oem-solar-panels-p-500964.html. Accessed 17 Jan 2016. Simjee, F., and P.H. Chou. 2006. “Everlast: Long-Life, Supercapacitor-Operated Wireless Sensor Node.” In The International Symposium on Low Power Electronics and Design (ISLPED), 197–202. Tegernsee, Germany: IEEE. Simjee, F.I., and P.H. Chou. 2008. “Efficient Charging of Supercapacitors for Extended Lifetime of Wireless Sensor Nodes.” Transactions on Power Electronics (IEEE) 23 (3): 1526–1536. Simpson, C. 2011. Linear and Switching Voltage Regulator Fundamentals. White Paper, Dallas, TX: Texas Instruments. Solar Electric Supply. 2016. Shell Solar. Solar Electric Supply, Inc. http://www. solarelectricsupply.com/solar-panels/shell/shell-solar. Accessed 19 June 2016. SolarWorld Americas. 2016. Technical Downloads. SolarWorld Americas, Inc. http://www. solarworld-usa.com/technical-downloads/datasheets. Accessed 23 Feb 2016. SparkFun. 2016. LTC4150 Coulomb Counter Hookup Guide. SparkFun Electronics. https://learn. sparkfun.com/tutorials/ltc4150-coulomb-counter-hookup-guide. Accessed 22 Mar 2016. SparkFun Electronics. 2016. The Basics of GPS. SparkFun Electronics. https://learn.sparkfun.com/ tutorials/gps-basics. Accessed 16 Aug 2016. Stanley-Marbell, P., and D. Marculescu. 2007 “An 0.9
1.2”, Low Power, Energy-Harvesting System with Custom Multi-Channel Communication Interface.” In The Conference on Design, Automation and Test in Europe (DATE), 15–20. Nice Acropolis, France: ACM/IEEE. State Electronics. 2016. Potentiometer Frequently Asked Questions. State Electronics, Inc. http:// www.potentiometers.com/potcomFAQ.cfm?FAQID=19. Accessed 20 May 2016. STMicroelectronics. 2013. ST890: 1.2 A Current Limited High-Side Power Switch with Thermal Shutdown. STMicroelectronics. http://www.st.com/web/en/resource/technical/document/ datasheet/CD00002878.pdf. Accessed 23 Feb 2016. Szewczyk, R., E. Osterweil, J. Polastre, M. Hamilton, A. Mainwaring, and D. Estrin. 2004. “Habitat Monitoring with Sensor Networks.” Communications of the ACM (ACM) 47 (6): 34– 40. Tadesse, Y., S. Zhang, and S. Priya. 2009. “Multimodal Energy Harvesting System: Piezoelectric and Electromagnetic.” Journal of Intelligent Material Systems and Structures 20 (5): 625–632. Tan, Y.K., and S.K. Panda. 2011. “Energy Harvesting From Hybrid Indoor Ambient Light and Thermal Energy Sources for Enhanced Performance of Wireless Sensor Nodes.” Transactions on Industrial Electronics (IEEE) 58 (9): 4424–4435. Taneja, J., J. Jeong, and D. Culler. 2008. “Design, Modeling, and Capacity Planning for Micro-Solar Power Sensor Networks.” In The 7th international Conference on Information Processing in Sensor Networks (IPSN), 407–418. St. Louis, MO: ACM. TAOS. 2007a. TSL230RD, TSL230ARD, TSL230BRD Programmable Light-to-Frequency Converters. Texas Advanced Optoelectronic Solutions Inc. https://www.mouser.com/ds/2/ 588/TSL230RDTSL230ARDTSL230BRD-P-519226.pdf. Accessed 30 May 2019. TAOS. 2007b. TSL250RD, TSL251RD, TSL260RD, TSL261RD: Light-to-Voltage Optical Sensors. Texas Advanced Optoelectronic Solutions Inc. https://www.mouser.com/ds/2/588/ TSL230RDTSL230ARDTSL230BRD-P-519226.pdf. Accessed 30 May 2019. Techopedia. 2016. Time to Live. Techopedia, Inc. https://www.techopedia.com/definition/27002/ time-to-live-ttl. Accessed 17 June 2016. TechTarget. 2016. T1 (T-1) definition. TechTarget. http://searchenterprisewan.techtarget.com/ definition/T1. Accessed 15 Jan 2016.

608

7 Energy Harvesting Projects for WSNs

TECTEG. 2019. Low (DT) Thermoelectric Harvesting TEG Power Module. TECTEG. https:// tecteg.com/low-dt-thermoelectric-harvesting-teg-power-module/. Accessed 2 June 2019. Texas Instruments. 2004a. MSP430x11x2, MSP430x12x2 Mixed Signal Microcontroller. Texas Instruments, Inc. http://www.ti.com/lit/ds/symlink/msp430f1122.pdf. Accessed 8 Oct 2016. Texas Instruments. 2004b. TPS61100T,TPS61103,TPS61106T, TPS611070: Dual-Outputs, Single-Cell Boost Converter. Texas Instruments, Inc. April 1, 2004. http://www.ti.com/lit/ds/ symlink/tps61106.pdf. Accessed 8 Oct 2016. Texas Instruments. 2006. TPS3103xxx-EP, TPS3106xxx-EP, TPS3110xxx-EP: Ultra-LowSupply-Current/Supply-Voltage Supervisory Circuits. Texas Instruments, Inc. http://www.ti.com/lit/ds/symlink/tps3106-ep.pdf. Accessed 19 Aug 2016. Texas Instruments. 2007. CC1000: Single Chip Very Low Power RF Transceiver. Texas Instruments. http://www.ti.com/lit/ds/symlink/cc1000.pdf Accessed 23 Dec 2013. Texas Instruments. 2008. TPS71501-EP: 50-mA, 24-V, 3.2-lA Supply Current, Low-Dropout Linear Regulator. Texas Instruments, Inc. http://www.ti.com/lit/ds/symlink/tps71501-ep.pdf. Accessed 21 Sept 2016. Texas Instruments. 2012. SimpleLink GPS CC4000 Fully integrated GPS solution. Texas Instruments, Inc. http://www.ti.com/ww/en/wireless_connectivity/gps/CC4000Brochure.pdf. Accessed 2 June 2019. Texas Instruments. 2015. TPS6107x 90% Efficient Synchronous Boost Converter With 600-mA Switch. Texas Instruments, Inc. http://www.ti.com/lit/ds/symlink/tps61070.pdf. Accessed 8 Oct 2016. Texas Instruments. 2016. MSP430F149: 16-Bit Ultra-Low-Power Microcontroller, 60 kB Flash, 2 KB RAM, 12 bit ADC, 2 USARTs, HW Multiplier. Texas Instruments, Inc. http://www.ti.com/ product/MSP430F149. Accessed 16 Aug 2016. The Free Dictionary. 2016a. Amperage. Farlex. http://www.thefreedictionary.com/amperage. Accessed 16 Aug 2016. The Free Dictionary. 2016b. Mane. Farlex. http://www.thefreedictionary.com/mane. Accessed 18 Aug 2016. Tolle, G., et al. 2005. “A Macroscope in the Redwoods.” The 3rd International Conference on Embedded Networked Sensor Systems (SenSys), 51–63. San Diego, CA: ACM. l-blox ag. 1999. l-blox GPS-MS1E: GPS Receiver Module based on SiRFstar I/LX. l-blox ag. https://www.qsl.net/dk7nt/gpsrx/gpsms1.pdf. Accessed 1 June 2019. UC Berkeley WEBS Project. 2004. nesC: A Programming Language for Deeply Networked Systems. UC Berkeley. http://nescc.sourceforge.net. Accessed 23 Mar 2014. U.S. Department of Energy. 2010. Discussion of Data for Smart Grid Metrics and Benefits: Storage System Performance Supplement. U.S. Department of Energy. https://www.smartgrid. gov/files/energy_storage_system_performance_supplement.pdf. Accessed 18 Apr 2016. Villalva, M.G., J.R. Gazoli, and E.R. Filho. 2009. “Comprehensive Approach to Modeling and Simulation of Photovoltaic Arrays.” Transactions on Power Electronics (IEEE) 24 (5): 1198– 1208. Warnell School of Forestry and Natural Resources. 2019. Dilution of Precision. University of Georgia. http://introgps.uga.edu/course/Dilution_of_Precision.html. Accessed 1 June 2019. Wellshow Technology. 2016. What is an SMA Connector? Wellshow Technology. http://www. wellshow.com/technical-support/connector-support/what-is-an-sma-connector/. Accessed 15 Jan 2016. WhatIs.com. 2005. bit error rate (BER). TechTarget. http://whatis.techtarget.com/definition/biterror-rate-BER. Accessed 12 July 2016. WhatIs.com. 2006. Backhaul Definition. TechTarget. http://whatis.techtarget.com/definition/ backhaul. Accessed 15 Jan 2016. WhatIs.com. 2011a. kelvin (K). TechTarget. http://whatis.techtarget.com/definition/kelvin-K. Accessed 26 Dec 2016. WhatIs.com. 2011b. lux (lx). TechTarget. http://whatis.techtarget.com/definition/lux-lx. Accessed 26 Dec 2016.

References

609

WhatIs. 2016. Order of Magnitude. TechTarget. http://whatis.techtarget.com/definition/order-ofmagnitude. Accessed 5 Sept 2016. Williams, J. 1989. High Efficiency Linear Regulators. Linear Technology Corporation. Accessed 10 Oct 2016. Wu, C.-H., and T.-W. Kuo. 2007. “An Efficient B-Tree Layer Implementation for Flash-Memory Storage Systems.” Transactions on Embedded Computing Systems (ACM) 6 (3). Xilinx. 2008. XC2C32A CoolRunner-II CPLD. Xilinx, Inc. http://www.xilinx.com/support/ documentation/data_sheets/ds310.pdf. Accessed 8 Oct 2016. Zhang, H.J. 2013. Basic Concepts of Linear Regulator and Switching Mode Power Supplies. White Paper, Milpitas, CA: Linear Technology. Zhang, P., C.M. Sadler, S.A. Lyon, and M. Martonosi. 2004. “Hardware Design Experiences in ZebraNet.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 227–238. Baltimore, MD: ACM.

Chapter 8

Energy Management Projects for WSNs

A winning athlete distributes his energy over the run …

8.1

Energy Management Projects

The chapter herein focuses on projects that manage energy needs and consumption based on the protocols presented in Part II of this book. Typically, the project in Sect. 8.2 presents the experience gained from building and developing a sustainable WSN for environmental monitoring. The illustration starts from the initial design up to hardware and software evolutions through an iterative process of software and hardware designs and developments, while maintaining backward design revisions for sake of hardware and software compatibility. The initial design was implemented using commercial off-the-shelf technology; this is especially important in applications where no prior data had been collected on a similar scale in the same environment. Then, for the power thirsty detection nodes, a first system evolution stage introduced a novel adaptive sensing approach based on reinforcement learning, and also included a delay-tolerant data collection. The second evolution phase proceeded to enhance the hardware of the most power thirsty nodes to reduce their energy consumption. The new platform provided a drastic reduction in the need for labor-intensive field trips to replace depleting batteries and maintenance. Such optimization led to a meaningful improvement in terms of maintenance costs, while triggering another round of software optimization. Importantly, more data are collected at low duty-cycles. The introduced project provides a model to be adopted to foster more elaborate projects that consider energy from a global view that is not bound to either energy harvesting or energy management techniques, but to both in a thoughtfully designed embracing system.

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_8

611

612

8 Energy Management Projects for WSNs

Fig. 8.1 European badger “Meles meles” (Hillman 2016)

8.2

Evolution and Sustainability of a Wildlife Monitoring Sensor Network

This project designed and built a distributed WSN designed to monitor wildlife and environmental conditions in a dense woodland environment, the Wytham Woods, Oxfordshire, UK (The University of Oxford 2016). Figure 8.1 presents the animals that are monitored, the European badgers “Meles meles”1 (Hillman 2016). As portrayed in Fig. 8.2, the planned network was made up of several components (Dyo et al. 2010): • Active RFID transmitters that are attached directly to European badgers as wearable collars. • The European badgers are monitored by a collection of fixed detection nodes containing RFID readers, and distributed throughout the woods at key locations close to known badger setts and latrines. • A bed of fixed sensor nodes that are deployed within badger foraging areas to monitor micro-climatic conditions and their effect on species migration and mobility patterns. • Zoologists carrying out routine observations and equipment maintenance act as mobile sinks and assist in the task of data collection, relieving the network from some of its communication load.

1

Badgers have white faces with two distinct black stripes running down either side. They are powerful mammals with large heads, and strong legs and claws, well suited for digging and burrowing into the earth. Badgers jaws are powerful enough to crush most bones; the badgers are of the few predators able to kill and eat hedgehogs. The head and body length are 65–80 cm, tail length 12–17 cm, and weight 8–12 kg.

8.2 Evolution and Sustainability of a Wildlife …

Detection node 1

613

Sensor node D Sensor node F Detection node 3

Sensor node B Mobile sink

Sensor node A

Sensor node E Sensor node C

Detection node 2

Legend: The network comprises badgers equipped with RFID collars, detection nodes composed of fixed RFID receivers, mobile sinks consisting of environmental sensor nodes and zoologists, and a fixed gateway.

Fig. 8.2 Heterogeneous wildlife monitoring network (Dyo et al. 2010)

• A single solar-powered 3G gateway with cellular connectivity to relay data instantaneously to the end-users. The data collected throughout the deployment have the potential to offer zoologists a deep insight into the social life of badgers and on the correlation of their activities with weather and micro-climatic variations. Zoologists and network engineers can assign priorities to different data types; a priority value reflects the tolerable delay between generating sensor data and delivering them to the user. The data of interest as generated by the network fall into three categories: • RFID readings captured by detection nodes to reflect badgers observations. • Environmental data, such as humidity and temperature, monitored at regular intervals by fixed sensor nodes. • Network health data indicating battery levels, memory usage, and any sensor errors. Section 8.2.1 describes the initial “exploratory” field-deployable prototype designed to understand the application requirements and the usage patterns. Then in Sects. 8.2.2 and 8.2.3, a focus is accorded to the gradual alterations in the initial design, based on feedback from the zoologists. Each evolution phase is evaluated in terms of maintenance cost. The system design went over two phases: • In the first phase, the system at the software level was optimized. A novel sampling approach, based on reinforcement learning, was devised for the power thirsty detection nodes. The idea was to exploit the behavior patterns of observed badgers in order to more efficiently control energy consumption. A storage management scheme was proposed to take into account data urgency and sink mobility so as to allocate sensor data to carefully selected storage nodes. It was observed that these software optimizations had a noticeable effect

614

8 Energy Management Projects for WSNs

on the maintenance costs, but the network still required long hours of hands-on human intervention. • The second phase proceeded to enhance the hardware of the most power thirsty nodes to reduce their energy consumption. The new platform provided a drastic reduction in the need for labor-intensive field trips to replace depleting batteries. Such optimization led to a dramatic improvement in terms of maintenance costs, while triggering another round of software optimization. Sampling and in-network storage are revisited in light of the new hardware capabilities. Evolving the hardware significantly impacted the performance of algorithms running on the nodes, which prompted the introduction of a more energy-efficient sampling algorithm for detecting badgers. Also, it effected the performance of the storage management scheme by altering the patterns of sink mobility. The running costs of the resulting system were reduced to such extent that it made it realistic for zoologists to envision network expansion. The know-how acquired in this project highlights the impact of maintenance costs on system design, and the evolution as well as the interplay between hardware and software optimizations. It also points out the need to take into account domain knowledge and application requirements to enable successful long-term deployments.

8.2.1

Initial System Design

As will be illustrated, the initial design of this badgers monitoring system focused on strong modularity and portability.

8.2.1.1

Sensing

Environmental Monitoring To investigate the impact of microclimate on individual badger behavior, Tmote Sky motes (Moteiv 2006) were equipped with two external SHT-71 digital temperature and humidity sensors (Sensirion 2019). One of the sensors was buried 30 cm underground to measure temperature, and the other was mounted at one-meter height. Ten of these nodes were deployed in the woods, and each took a measurement every five minutes. Suitable sensor housing was developed by trial and error to protect the sensor and also to allow it to record accurate humidity measurements. Devices were configured to either act as standalone data loggers that have very low average current consumption of 30 lA, or as normal network nodes.

8.2 Evolution and Sustainability of a Wildlife …

615

Badger Monitoring Wildlife tracking presents unique challenges by requiring animal borne tags to be small, highly reliable, and inexpensive. This incited the use of a commercial 433 MHz active RFID tag (Wavetrend 2004) over the alternative of designing a custom miniature mote platform. The datasheet of the Wavetrend Domino Tag LTG-100 is available in Appendix B. The selected tags satisfied several design requirements including lower cost, miniature size, and long lifetime. The small size of the tags was crucial as it allows tracking of much smaller animals. Notably, the selection of commercial low-cost tags allowed the team to capitalize on the advantages of the tested component and focus on ground-sensor network design, measurements, data collection, and analysis. The tag measured 40 mm
20 mm
3 mm in size, and was equipped with a 56 mm external whip antenna2 (Chen 2007); it had power from an onboard 3 V CR2450 coin cell battery (Energizer 2016) with an expected minimum lifespan of 2 years at 0.4 s transmit interval. Each RFID tag was hermetically sealed in waterproof epoxy resin3 (Johnson 2016) to protect the tag from environmental and mechanical damage (e.g., chewing by an animal). The presence of tagged animals was registered by 26 RFID detection nodes placed at setts and latrines in the core study area. The detection range of a tagged animal was up to 30 m, with the selected 433 MHz frequency, providing longer communication range and lower obstacle fading through dense vegetation. Figure 8.3 illustrates a badger detection node, and an RFID tag potted in epoxy and mounted on a collar. Each badger detection node was composed of an active RFID reader powered by an SLA battery (Sect. 2.2.4.1 in Chapter 2), a Tmote Sky mote (Moteiv 2006) powered by AA-size batteries, and a custom-designed mote extension board. For each detected tag, the reader provided the following information: tag ID, reader ID, serial counter number, received signal strength (RSSI), and a checksum. The serial counter number facilitated an estimation of the tag age and could be used for localization purposes for a tag simultaneously registered by several readers. The extension board allowed the interconnection of the mote, RFID reader, and peripheral devices to an RS232 to TTL converter, MOSFET switches, and the voltage regulators. The output voltage was in the range from 6 to 12 V, and was configurable either through potentiometers and switches onboard or from the mote via a standard I2C (I2C) interface4 (Texas Instruments 2015). The power management software on the mote, duty-cycled the peripheral devices including the

A whip antenna consists of a single straight flexible wire or rod. The bottom end of the whip is connected to the radio receiver or transmitter; it is designed for flexibility so that it does not easily break. The name is derived from the whip-like motion when disturbed. 3 Epoxy adhesives are sold in local hardware stores, and epoxy resin is used as a general purpose adhesive, such as binder in counter tops or coatings for floors, binder in cements and mortars, rigid foams, for solidifying sandy surfaces in oil drilling, and in industrial coatings. 4 The I2C bus is a powerful bus used for communication between a master (or multiple masters) and a single (or multiple) slave devices. 2

616

8 Energy Management Projects for WSNs

Fig. 8.3 Badger detection node and RFID tag potted in epoxy and collar mounted (Dyo et al. 2010)

reader, and monitored both mote and reader voltages to shut down the system when the voltage become low. It should be noted that the tracking and ground communication have different requirements in terms of communication range and antenna configuration, so decoupling the two communication systems is desirable. Specifically, ground communication requires extended communication range with preferably high bandwidth, whereas detection requires a biologically meaningful communication range, with a high degree of consistency, which requires consistent antenna orientation and receiver sensitivity of all detection nodes.

8.2.1.2

Data Collection

The initial system design distinguished between two types of data, high-volume data consisting of raw badger observations and low-volume data composed of environmental readings, summaries of badger visits, and network status reports.

8.2 Evolution and Sustainability of a Wildlife …

617

Compression and Local Storage Due to the large data volumes generated by the network, typically in excess of 400,000 observations per week, a simple delta-based compression technique5 (Suel and Memon 2002) was implemented to allow more data to be stored in the 1 MByte flash memory of the Tmote Sky. This approach is application-specific and computationally lightweight; it achieves 25% higher compression factor than standard compression methods, like gzip. Each raw reading, composed of a timestamp, tag age, and received signal strength, occupies 10 Bytes in its uncompressed form. Using this simple scheme, raw data are typically compressed by an average factor of 2.7
. This compares favorably with the LZ776 (Ziv and Lempel 1977) resource hungry gzip algorithm, which just achieves a compression factor of 2.0
on the same dataset. Thus, by reducing the volumes of data that need to be buffered within the network, it was possible to extend the memory lifetime of the RFID node almost threefold. These data might be compressed further using dictionary type compression algorithms such as S-LZW (Sadler and Martonosi 2006), but the required additional node resources would not justify the gains.

Routing Low-volume data such as network status messages are forwarded to the fixed 3G gateway node using a proactive shortest-path routing algorithm. The algorithm can be highlighted in several steps: • Every node maintains a routing table containing its distance to the gateway node. • Initially, the gateway advertises beacons with distance 0 to itself, and with increasing sequence (freshness) numbers. • The distance from a node to the gateway is evaluated taking into account the link qualities along the route. • Each node maintains a neighborhood table that shows statistics of outgoing traffic. The expected transmissions etx per message from the current node to a neighbor node N is computed as:

Delta compression and remote file synchronization techniques are concerned with efficient file transfer over a slow communication link in the case where the receiving party already has a similar file (or files). 6 LZ77 and LZ78 are two lossless data compression algorithms. Lossless compression is a class of data compression algorithms that allows the original data to be perfectly reconstructed from the compressed data. By contrast, lossy compression permits reconstruction only of an approximation of the original data, though this usually improves compression rates (and therefore reduces file sizes). The algorithms were named an IEEE Milestone in 2004. These two algorithms form the basis for many variations including LZW, LZSS, LZMA, and others. 5

618

8 Energy Management Projects for WSNs

etxðN Þ ¼ attempted txðN Þ=successful txðN Þ

ð8:1Þ

• The distance to the gateway node is defined as the sum of expected transmissions on all links along the route. To be noted, all the links along a route have an etx = 1, the distance is equal to the number of hops along the route. • Every node broadcasts its distance to the gateway every 30 min. Upon receiving an advertisement from a neighbor N, a node compares the advertised distance (advDist) to the distance in its local routing table (rtDist): If advDist þ etxðN Þ\rtDist then sets rtDist: ¼ advDist þ etxðNÞ and sets neighbor N as its next hop: • If the route quality deteriorates significantly, a node will simply select the next best available route. • Once routes have been identified, data are transmitted using the uIP (micro-IP) IPv6 networking stack (Durvy et al. 2008), which in turn uses the X-MAC protocol (Buettner et al. 2006), both being distributed with Contiki OS7 (Dunkels et al. 2004; Contiki-Developers 2016).

uIP The choice of using Contiki uIP networking stack was strongly influenced by the positive findings in (Hui and Culler 2008). The added flexibility of using the IPv6 standard allows to easily adapt the network to other tasks during further network deployment, for example, accessing and maintaining individual nodes or allowing near real-time data streaming from specific nodes within the network. Although the overhead incurred in terms of code size is considerable for the Tmote Sky platform, approximately 16 KBytes of additional flash usage, the added modularity, and flexibility of the IPv6 network permitted to easily maintain and extend the network with new functions. Data are disseminated toward storage nodes on a local hop-by-hop basis, instead of an end-to-end basis. The usual end-to-end connection used in IPv6 networks is TCP/IP, which requires an additional overhead for establishing a connection and requires the end nodes to negotiate retransmissions. To avoid this overhead, UDP connections are used to transmit data along each hop toward the storage node. Upon receiving a packet from a child node, the parent node returns a UDP ACK message to confirm reliable data transfer.

7

Contiki is an open source, highly portable, multtasking operating system for memory-efficient networked embedded systems and wireless sensor networks. Contiki is designed for microcontrollers with small amounts of memory.

8.2 Evolution and Sustainability of a Wildlife …

619

Messages are stored in onboard flash and they are only marked for deletion once an ACK message is received from the parent indicating successful safekeeping transfer. A background garbage collection routine periodically cleans up the flash memory, formatting pages where all messages have been successfully uploaded. In addition, the frequency of neighbor advertisement and solicitation messages is reduced to a validity of 24 h. This helped to dramatically reduce network overhead.

MAC Layer X-MAC (Buettner et al. 2006) is used at the MAC layer, a preamble-based protocol in which senders indicate their intent to send data by frequently transmitting short wakeup messages. Nodes periodically wakeup, and if they hear a preamble which indicates that a packet is addressed to them, they respond with an acknowledgement. This terminates the wakeup phase and the packet is sent. Nodes are configured to wakeup every 500 ms and listen for 5.8 ms, resulting in effective basic duty-cycle of 1.1%.

8.2.2

Evolution Stage 1: Improving Sensing and Data Collection

The initial system design has evolved by introducing algorithmic improvements. The main weaknesses of the initial design were the high-energy consumption of the badger detection nodes (RFID readers), and the heavy communication load around the fixed gateway. As presented in Sect. 8.2.4, about a visit a week was necessary to change batteries and keep the system running.

8.2.2.1

Adaptive Sensing

RFID readers are the major source of power consumption on detection nodes. Despite being powered by a 12 V 18 Ah battery, they only lasted for one week without a duty-cycling. Increasing the lifetime of readers is therefore critical for large-scale long-lived deployments. An obvious way to save energy is to duty-cycle the RFID reader by periodically turning it ON for a fixed duration of Ton seconds every Tinterval seconds. Yet, setting optimal parameters is not straightforward. A high-frequency sampling may be too wasteful, whereas low-frequency sampling may lose important contacts. Tuning also requires knowledge of badger activity, which may not be known in advance. An adaptive duty-cycling approach was thus devised; it dynamically adapts the parameters Ton and Tinterval taking into account the badger activity. The problem

620

8 Energy Management Projects for WSNs

was formulated in terms of reinforcement learning8 (Kaelbling et al. 1996), and a control strategy was created to adjust node duty-cycles based on animal arrival patterns (Dyo and Mascolo, Efficient node discovery in mobile wireless sensor networks 2008). The initial values of Ton and Tinterval are set to reflect the targeted duty-cycle and the hardware capabilities of the detection nodes. For example, to achieve a target duty-cycle of about 9%, Ton is set to 30 s and the initial value of Tinterval to 330 s. For efficiency sake, Ton is chosen to be significantly longer than the RFID reader boot time Tboot, which was 10 s. The approach consists of two main components, the short-term adaptation component and the long–term adaptation component. Short-term adaptation extends the awake time Ton of the RFID reader by a fixed short period of Text seconds each time a badger activity is detected (i.e., a tag is in range). The short-term adaptation exploits the temporal burstiness of badger arrivals, as the detection of a beacon is usually a good predictor of activity. The drawback of the periodic sampling technique, even in the presence of short-term adaptation, is that it assumes uniform badger activity throughout the day. However, it is rare that animals or humans remain continuously active throughout a day, but rather they follow a 24-h circadian rhythm9 (National Sleep Foundation 2019), which may vary depending on the environmental conditions. Badgers are nocturnal animals that are inactive during the day, which means that sampling during the day may be wasteful. The long-term adaptation component learns daily patterns of badger activity and adapts the interval Tinterval accordingly. A target daily budget B is defined to be the amount of seconds that a badger detection node should spend in active state per day. Each day is divided into N equal time slots. Then, each node computes the expected number of sightings E(d,t) during a day d for timeslot t and assigns to each timeslot a budget B(d,t), proportional to E(d,t): Bðd; tÞ ¼ B  E ðd; tÞ=

N X

E ðd; iÞ

ð8:2Þ

i¼1

8

Reinforcement learning is the problem faced by an agent that must learn behavior through trial and error iterations with a dynamic environment. There are two main strategies for solving reinforcement learning problems. The first is to search in the space of behaviors in order to find one that performs well in the environment. This approach has been taken by work in genetic algorithms. The second approach is to use statistical techniques and dynamic programming methods to estimate the utility of taking actions in states of the world. Reinforcement learning is thus a type of machine learning, and thereby also a branch of artificial intelligence. It allows machines and software agents to automatically determine the ideal behavior within a specific context, in order to maximize their performance. Simple reward feedback is required for the agent to learn its behavior; this is known as the reinforcement signal. 9 Circadian rhythms are physical, mental, and behavioral changes that follow a roughly 24-h cycle, responding primarily to light and darkness in an organism’s environment. They are found in most living things, including animals, plants, and many tiny microbes. The study of circadian rhythms is called chronobiology.

8.2 Evolution and Sustainability of a Wildlife …

621

This is the equivalent of “bidding” more resources in what has been a productive timeslot in previous days. B(d,t) should be in the range [Bmin,Bmax] in order to still explore all timeslots, even if they have not recently experienced any sighting, and to constrain the maximum number of times the node wakes up within a given time slot. Since in a timeslot of length T, the RFID reader is to be active only for B(d, t) seconds, B(d,t)/T = Ton/Tinterval and the node can adjust the duty-cycle in each timeslot by setting the interval Tinterval ¼ T  Ton =Bðd; tÞ between successive wakeups. On the first B(d,t) day, the budget is spread uniformly throughout all N timeslots, since there is no information about sightings. Then, the expected number of sightings E(d,t) in timeslot t of a particular day d is evaluated as follows: Eðd; tÞ ¼ a  Oðd  1; tÞ þ ð1  aÞ  Eðd  1; tÞ

ð8:3Þ

where O(d−1, t) is the actual number of sightings that were observed in the same timeslot on the previous day, and a is a weight in the range [0,1] to control how rapidly new information is incorporated into the filter. Small values of a will give more weight to past history, but will make the adaptation process slow and unable to capture sudden changes, whereas large values will make it very reactive to short-term changes and less able to capture long-term patterns.

Simulation-Based Evaluation The evaluation of the adaptive duty-cycling technique has been performed both through simulation and real deployment. Throughout the project evaluation, N = 24 one-h timeslots were used; that is T = 3600 s. Within each timeslot, the detection node turns the RFID reader ON and OFF to achieve the targeted duty-cycle using Eq. 8.2. The ON time, Ton, was selected to be 30 s, the initial interval Tinterval = 330 s, which corresponded to a budget B = 7,854 s and a duty-cycle of about 9%, and the extension time Text = 300 s. The [Bmin,Bmax] range was set to [B/120, B/24]  [65, 327]. A fixed duty-cycling algorithm was used, where a node wakes up and goes to sleep at fixed intervals of time, as a baseline. The algorithms have been implemented in Tossim 2.0.2 simulator (Levis and Lee 2003) and evaluated by replaying the real data recorded by always-ON node. WSN simulators and emulators are meticulously exposed in (Fahmy 2016). Ten simulation runs were conducted for each algorithm with random node offsets. The performance of always-ON, fixed duty-cycling, and adaptive algorithms were found to be: • The always-ON node detected all 76,707 encounters at 100% duty-cycle. • The fixed duty-cycling node detected 7,773 (10.13%) encounters at 9% duty-cycle. • The adaptive node detected 46,214 (60.24%) encounters at 5% duty-cycle, resulting in much higher encounters per duty-cycle than always-ON and fixed duty-cycling nodes.

622

8 Energy Management Projects for WSNs

Deployment-Based Evaluation In order to evaluate the proposed duty-cycling technique in a real deployment, two detection nodes were placed with the same hardware and antenna orientation next to each other. One of the nodes was always ON, whereas the other executed the proposed adaptive duty-cycling technique. Also, data were processed from the always-ON node to simulate a fixed schedule. The adaptive node was configured to work at 9% duty-cycle. The evaluation was based on 833 h of summery July deployment data from both nodes. A zoologist periodically retrieved data from the two nodes. The obtained results are found to be: • The fixed duty-cycling node captured 7,201 sightings while using 10% of the power of the always-ON node. • The adaptive duty-cycled node detected 54,568 (73%) of all sightings, while consuming approximately 8.2% of the energy.

8.2.2.2

Delay-Tolerant Data Collection

The initial design of the data collection algorithm was built on the basis that raw RFID data, which are high-volume and low-priority, are stored locally at sensor nodes. Other data had higher priority and were forwarded to the 3G gateway using a tree-based routing algorithm. This initial approach has its foundation on a work that prioritizes data traffic and takes into account routing costs to determine whether to discard data, store it locally, or forward it to the gateway (Werner-Allen et al. 2008). In the proposed approach, a further step is added and a delay-tolerant data collection scheme is conceived, which leverages the movement of zoologists and environmental scientists to efficiently collect sensor data. Not only data are prioritized based on their urgency, but also nodes are prioritized based on the frequency in which mobile sinks visit them. Hence, data are forwarded to carefully selected storage nodes, depending on both data and node priorities.

Data Priorities When data were generated, they were assigned a data priority class that represents the latency allowed until they had to be delivered to the end-user. The network generated observations data of tagged badgers as captured by the detection nodes and environmental sensor data such as temperature and humidity. Nodes also created heartbeat messages that reflected their current operational status. This included information such as remaining battery level, memory usage, and network statistics. The motivation behind the delay-tolerant networking approach is the fact that the majority of the generated data do not have strict latency constraints, but it was imperative, however, that all data get eventually collected. In order to

8.2 Evolution and Sustainability of a Wildlife …

623

maximize the battery lifetime of nodes in the network, a distributed storage and delivery method was adopted; messages are directed to different destinations based on their tolerable delay. In the proposed algorithm, three priority classes are made available; more priority classes may be added. The three priority classes are proposed to be: • Priority class 1 to represent data with urgent latency requirements (maximum of a few hours delay). These data are forwarded to the 3G gateway node for direct access by the researchers. Data of this class could either represent an unusual event or a network status report to ensure the proper network functioning throughout the deployment. • Priority class 2 is for the data with medium latency requirements (maximum of a few days delay). These data are forwarded to frequently visited storage nodes for opportunistic collection. Data of this class could be summaries of badger visits. • Priority class 3 represents data with no latency constraints (delays of weeks are acceptable). Data of this class were collected for later use, such as raw sensor data, and will remain in memory till needed through direct download. Priorities can be assigned not only to raw sensor data, but also to composite events or aggregated data. For example, raw badger information may have priority class 3, but when unusually high activity is observed around a certain sett this composite event can be assigned priority class 1, and will be forwarded to the fixed gateway for immediate delivery. Furthermore, data priorities can either be fixed or dynamic, for example, they could vary depending on the zoologists needs and the data collected from the sensor network.

Node Priorities The proposed priority-based in-network storage management approach is simple and effective as revealed in the coming arguments: • Initially, each node is assigned a priority class PN based on the expected frequency of being visited by mobile sinks for data collection. Some nodes, such as those close to roads and paths, are regularly in contact with a mobile sink and thus contribute a small delay. Other nodes that are placed in rarely visited remote locations will be subject to a large delay. • The more a node is frequently visited, the lower is the expected data delivery time, and the lower is the assigned node priority class. The 3G gateway is assigned a priority class 1 as it can offer the lowest data delivery latency. Nodes that are visited at least every three days by mobile sinks act as temporary data storage nodes of priority class 2. The remaining nodes in the network have priority class 3.

624

8 Energy Management Projects for WSNs

• In the storage management scheme, a data item of priority class PD is stored at the closest node with priority class PN, where PN  PD. Messages with data priority class 1 are directed toward the 3G enabled gateway, which allows users to access them with small delay. Data of priority class 2 are stored at the closest node that has priority class 1 or 2. Data of priority class 3 are stored locally at the node where they have generated. Noteworthy, node priorities can change dynamically due to changes in sink mobility. If a node becomes visited less often, some of the messages it stores may need to migrate to another node depending on their priorities. By consulting domain experts to classify data into priority groups, it is possible to map data to suitable storage nodes; thus, data are delivered on time and with the lowest communication cost. As a data item remains stored at a node, it gradually ages, and its remaining tolerable delay decreases; hence, it can dynamically change priority and be forwarded to another suitable storage node.

Priority and Mobility Aware Routing Once data are assigned a priority and are compressed, they are forwarded to the appropriate destination node; specifically, 3G gateway nodes of priority class 1 or storage nodes of priority class 2. As shown in Table 8.1, every node maintains a routing table containing information of each of the available priority classes. Every node periodically broadcasts its routing table information regarding each priority class; this broadcast period is set to 30 min. To be noted, a single advertisement contains routing information regarding all priority classes. Thus, the size of advertisements does not increase with the number of destination nodes, but only in proportion to the number of priority classes. Therefore, the routing overhead of building multiple trees, instead of one, was negligible. Table 8.1 Routing table (Dyo et al. 2010) Priority

Next hop

Seq. no

Dest. node

Distance

1 NA 30 NE 3 34 NF 1 2 NB Legend: • The next hop represents the neighbor to which the data of a certain priority will be forwarded • The sequence number (Seq. no) and destination node (Dest. node) fields are used to deal with loops occurring in the network. Sequence number is issued by the destination node and represents the freshness of routing information concerning that node, as in DSDV (Perkins and Bhagwat 1994) • The distance to a destination node (Distance) was evaluated taking into account the link qualities along the route, as was evaluated the distance to the gateway in Sect. 8.2.1.2

8.2 Evolution and Sustainability of a Wildlife …

625

Evaluation This section presents the results out of 20 days network deployment period with a total of 24 RFID readers. Multiple considerations and findings are thoughtfully observed: • For half of the time, data were collected using the distributed storage approach described in Sect. 8.2.2.2.1, and for the other half of the time using a centralized storage approach, as in the initial design presented in Sect. 8.2.1. The centralized approach simply forwarded all data to the 3G node; the distributed approach used three additional priority class 2 storage nodes where data were temporarily stored for opportunistic pickup. An interesting article on opportunistic data collection for disconnected wireless sensor networks by mobile mules is available in (Tseng et al. 2013). • In order to have comparable results, a fixed data generation rate is used throughout the network evaluation period. Two latency constraints are seen: – Priority class 1 data consisting of network status messages, generated at each node every 30 min, and had to be delivered to the end-user within two hours. – Priority class 2 data composed of badger activity summaries are generated at each node every 15 min with a delivery latency of three days. In the centralized approach, these data are forwarded to the fixed 3G gateway, whereas in the distributed approach, they are delivered to the nearest storage node that satisfies latency constraints. • In both centralized and distributed approaches, a high delivery ratio was achieved; specifically, 99.9% of the data were correctly transferred to the appropriate storage or 3G nodes. Furthermore, the average latency was as low as 14.1 s per hop, thus data can be sent over five hops in less than 75 s on average. • The network status messages, which contain the radio ON time at each node, permitted deriving the average radio duty-cycle of each node over the test period. Furthermore, with the centralized and distributed storage management schemes, the distribution of radio duty-cycles across the different nodes in the network was studied. It was found that the centralized approach exhibited 46% higher duty-cycle than the proposed distributed approach in the average case, and 57% in the worst-case at routing hotspots. This reveals that a careful forwarding of different priorities data to suitable storage nodes, not only reduced the average energy consumption, but also balanced the load more evenly in the network. Benefits would have been better if priority class 3 data were forwarded to the gateway thru the centralized approach.

626

8.2.3

8 Energy Management Projects for WSNs

Evolution Stage 2: Hardware Improvements

The algorithms proposed in Sect. 8.2.2 improved the usability of the initial design introduced in Sect. 8.2.1, but they were limited by hardware, i.e. the RFID detection node. By experimentation, field deployment and data gathering are imperative to a system’s successful deployment. For quick deployment, detection nodes were built using off-the-shelf components; unluckily, these components turned out to be too general for the specific needs. A new hardware became a must.

8.2.3.1

Designing a New Node

Feedback from the zoologists helped improving the system design; a comparison of the major design modifications is listed in Table 8.2. The ubiquitous Tmote Sky had enabled rapid prototype deployment, however, its limitations in terms of radio range and usable memory posed major limitations. A more modern and flexible module had to be incorporated into the design; its criteria focus was on low cost, power efficiency, and preferably being hand solderable. The Atmel ZigBit AMP microcontroller platform (Atmel 2009) is adopted as its ATmega1281V 8-bit processor (Atmel 2014) had better community support, especially in Contiki OS7 (Dunkels et al. 2004; Contiki-Developers 2016). This allowed porting the existing code rapidly from the Tmote Sky platform to the Atmel ATmega1281V platform, with minor modifications to the existing RF230 radio driver (SourceForge 2012). The radio range of the new module was improved to exceed 1 km in woodland at maximum power, a great improvement that considerably increased the span of the network. Remarkably, this is the transmission range of the radio, not the detection range of the RFID reader that remained unchanged. The drawback of transmitting at

Table 8.2 Comparative design evolution (Dyo et al. 2010) Node Processor Node RAM Node flash External flash RFID reader power Reader turn-ON time Radio range Cost per unit Mote battery Reader battery

Version 1 reader

Version 2 reader

Tmote Sky MSP430 10 KByte 48 KByte 1 MByte 900 mW 10 s 50 m USD 590 3 AA 18 Ah SLA

ZigBit AMP AVR ATmega1281V 8 KByte 128 KByte Up to 2 GB SD 96 mW 0.1 s 1 km USD 320 None 18 Ah SLA

8.2 Evolution and Sustainability of a Wildlife …

627

Fig. 8.4 Second design version (Dyo et al. 2010)

the highest power level is the increased current consumption from 17 to 50 mA. An illustration of the modified design is displayed in Fig. 8.4. The Atmel ZigBit AMP is essentially a microcontroller with an embedded radio. To satisfy application requirements, additional components are added: • External memory to the board in order to overcome the constraints that tied the initial system (Sect. 8.2.1). The board is equipped with a 4 MByte serial DataFlash chip (Atmel 2008) and a removable mini-SD memory card. This allows the addition of up to 2 GBytes of SD-based flash; larger capacities could be supported with modifications to the SD driver software to allow the use of high capacity cards. • An RTC with battery backup to allow the nodes maintain their time when changing the batteries. • An inconvenience with Tmote Sky is that the onboard sensors are not removable. In a real deployment, sensors must be placed externally in protective housing; thus with the new design, light and temperature sensors that are detachable from the main board were incorporated.

628

8 Energy Management Projects for WSNs

• A transition from the RS-485 version10 (Kugelstadt 2016) of the RFID reader to an OEM board11(Investopedia 2016). The RS-485 version was a suitable choice for the initial deployment, as it had a simple serial interface, and allowed great flexibility in daisy chaining multiple readers together. However, the high power consumption and slow turn-ON time were problematic. These issues imposed the use of a separate RFID reader and mote batteries, so that the mote would remain powered even if the reader exhausted its supply. Switching to the OEM version of the RFID reader solved performance restraining problems: – OEM RFID reader has a simple synchronous serial TTL interface and a pin that could be used to trigger an interrupt on the microcontroller when a tag was read. This permitted powering down the microcontroller while the reader was active, whereas in the previous version, the clock had to be kept for the UART. – In the Tmote Sky, the radio and the UART were multiplexed, which instigated difficulties of hardware locking to prevent concurrent access to the peripherals. In the new version, the OEM RFID reader has its own dedicated pins. Worth mentioning, the turn-ON time for the OEM RFID reader is below 100 ms, and it uses 96 mW when active. • A simpler power distribution system, with 3 V as a common rail, is adopted. A small charge pump was used to generate the 5 V required for the OEM RFID board. The nodes can be powered either from a 3 V battery or from a 12 V battery using a switching regulator.

10

RS-485 is an electrical-only standard. In contrast to complete interface standards, which define the functional, mechanical, and electrical specifications, RS-485 only defines the electrical characteristics of drivers and receivers that could be used to implement a balanced multipoint transmission line. This standard is intended to be referenced by higher-level standards. The RS-485 standard suggests that its nodes be networked in a daisy chain, also known as party line or bus topology. In this topology, the participating drivers, receivers, and transceivers connect to a main cable trunk via short network stubs. The interface bus can be designed for full-duplex or half-duplex transmission. Key features of RS-485 are: • • • • • •

Balanced interface. Multipoint operation from a single 5 V supply. −7 to +12 V bus common-mode range. Up to 32 unit loads. 10 Mbps maximum data rate (at 40 ft). 4000-ft maximum cable length (at 100 Kbps). 11 An original equipment manufacturer (OEM) has two definitions: • An OEM is a company whose products are used as components in the products of another company, referred to as the value-added reseller (VAR). The OEM generally works closely with the VAR company that sells the finished product and customizes designs based on its needs. • In the computer industry, OEM may also refer to the VAR, the company that buys products and incorporates or rebrands them into a new product under its own name.

8.2 Evolution and Sustainability of a Wildlife …

629

• A small prototyping area was included on the board to connect additional devices, such as a moisture sensor, to a node. The latest version of the detection node paid off in performance. The power consumption dropped by nearly an order of magnitude12 (WhatIs.com 2016). The storage space increased to a size that allows for 40 years of storage at the available data generation rates, as opposed to one week in the previous design. More information may thus be gathered and environmental sensors may be sampled at a much higher resolution. The communication range also augmented significantly, which allowed the network to cover a much larger area with fewer devices.

8.2.3.2

Duty-Cycling Revisited

Given that the RFID reader on the redesigned node can be powered up in 0.1 s, as opposed to the 10 s for the prior version, it was possible to modify the parameters of the learning algorithm presented in Sect. 8.2.2.1. The original Tinterval was set to 330 s, with a 9% duty-cycle. Although this saved a large amount of power which allowed the node to operate for longer time, it had the drawback of being unable to react to the presence of animals outside the normal predicted times, as the OFF time could be quite long (up to an hour). In order to address this issue, Ton was modified to be 1 s, and Tinterval to 11 s. This still resulted in a 9% duty-cycle, but the short-term adaptivity could react to the presence of unusual events, such as a badger emerging during the day. The longest time for which the reader was OFF has been reduced to less than a minute, which increased the chances of detecting animals, while still accounting for their nocturnal behavior. With the new parameters, the shorter wakeup interval resulted in both higher encounters and higher efficiency. The adaptive algorithm detected 89% of all encounters while working at 5% duty-cycle.

8.2.3.3

Data Collection Revisited

The hardware improvements introduced in Stage 2 had a dual effect on the data collection process: • Change in sink mobility patterns. Mobile sinks are the domain scientists that roam through the woods and opportunistically collect data from storage nodes;

12

In base 10, the most common numeration scheme, an increase of one order of magnitude is the same as multiplying by 10. An increase of two orders of magnitude is the equivalent of multiplying by 100, or 102. In general, an increase of n orders of magnitude is the equivalent of multiplying a quantity by 10n. As values get smaller, a decrease of one order of magnitude is the same as multiplying a quantity by 0.1. A decrease of two orders of magnitude is the equivalent of multiplying by 0.01, or 10−2. In general, a decrease of n orders of magnitude is the equivalent of multiplying a quantity by 10−n.

630

8 Energy Management Projects for WSNs

some of them are zoologists visiting the network for maintenance purposes. The visits of the zoologists were reduced because hardware optimizations made the change of batteries less frequent. With a fewer mobile sink visits, one would expect an increase in the data propagated over multiple hops through the fixed network, and thus an increase in the average and worst-case communication cost. • Change in communication range. The effect of reduced sink mobility was, however, offset by the significant increase in the communication range of fixed nodes. Interestingly, the hardware optimizations introduced in the second version drastically increased the communication range of sensor and badger detection nodes from 50 m to 1 km. As a result, all nodes had one-hop connectivity to the fixed 3G gateway, and no longer needed to make use of mobile sinks. Hence, in the second evolution stage, the hierarchy of nodes based on priorities collapsed, and it became less efficient to use mobile sinks for collecting data and relieving the fixed network. This reveals that the benefits of software-level optimizations, such as the priority-based delay-tolerant data collection, are tightly dependent on the hardware used. Clearly, the priority-based delay-tolerant scheme proposed in Sect. 8.2.2.2 yielded significant benefits in the first version of the system, but proved of little use to the second version.

8.2.4

Network Maintenance Costs

This section accounts for the evolution of the proposed system in terms of the costs involved, the baseline is the approximate cost of conventional VHF tracking (Kenward 2001). This involves tagging animals with VHF tags that emit periodic radio signals. The distinction between RFID and VHF tags is first to be highlighted: • VHF tags are analog devices achieving individual identification by frequency separation and limiting the number of IDs available. While the active RFIDs are digitally encoded, which allows more IDs in a given band without the need for a receiver to scan multiple channels. • The VHF tags can be picked up by receivers carried by field-workers at a range of tens to thousands of meters depending on environmental conditions. Using triangulation that requires at least two persons on the ground, the approximate location of the animal can be found. The RFIDs transmit at much lower power than VHF tags, which increases battery life, while limiting the transmission range to 30 m, so giving a more precise location estimate for tagged animals. VHF tracking has been a popular method since the late 1960s because it was, and still is in many circumstances, the only way of tracking wild animals.

8.2 Evolution and Sustainability of a Wildlife …

631

Noteworthy, the data collection methods by VHF and RFID are rather different, although they collect the same information (i.e. the location of a specific animal). Functionally, several points are to be recognized: • RFIDs log an animal about twice per second when it is nearby a detection node. • For VHF tracking, the more animals tracked by VHF, the more human trackers are required on the ground, which might disturb the animals being tracked. • The RFID system proposed throughout this project offers continuous automatic detection (presence/absence) of the animals at specific locations with minimal interference. • From previously tracking studies of badgers, using VHF, at least one person is needed to work for about 10 h a night to track one animal. Assuming enough people to work for 28 days, this would result in 280 h per person, costing USD 2,030 using a 7.25 USD/hour wage. Clearly infeasible in the long run, because one person can only track one animal at a time. It is also impossible to provide continuous tracking (i.e. 24/7) without considerable costs and man-hour overhead, not to mention the fact that the more people there are in the woods, the more the animals are disturbed. Importantly, there are several other methods of animal tracking, such as the GPS and ARGOS13 (CLS 2016) satellite-based systems. They are inappropriate though because of ARGOS inferior spatial resolution, and unreliable GPS performance in woodland. Furthermore, ARGOS tags cost over USD 1,500 each, while a badger-sized GPS tag lasts for only a few months. The RFID tags proposed in this project costed about USD 60 each and lasted for two years. An added USD 300 cost in the detection node includes the sensor motes and RFID readers. Totally, the network deployment including 74 RFID tags and 26 detection nodes, sums to 74  60 + 74  300 = USD 12,240, while buying 74 ARGOS tags would cost approximately USD 111,000. Table 8.3 shows the summary of the costs involved in maintaining the proposed system. From the table, several details are focused on: • The number of man-hours needed and the battery costs for each stage are considered. The total cost includes the price of monthly upkeep of the system.

13

Argos is a pioneer satellite-based system, which has been operating since 1978. The Argos system collects data from platform terminal transmitters, PTTs, and distributes sensor and location data to the final users. The Argos data collection and location system (DCS) is a data collection and relay program that provides global coverage and platform location. It helps the scientific community to better monitor and understand the environment, but also enables industry to comply with environmental protection regulations implemented by various governments. To meet system use requirements, all programs using Argos have to be related in some way or other to environmental protection, awareness or study, or to protecting human life. Applications for which there is a clear governmental interest are also approved. Argos was established under an agreement between the French space agency (CNES), the National Oceanic and Atmospheric Administration (NOAA, USA), the National Aeronautics and Space Administration (NASA, USA). Today, Argos is operated and managed worldwide, by CLS Group, a CNES subsidiary.

632

8 Energy Management Projects for WSNs

Table 8.3 Compared average cost to maintain each stage for four weeks Visits (man-hours) Stage 1 29.7 (HW only) Stage 1 10.8 (HW & SW) Stage 2 2.7 (HW only) Stage 2 1.35 (HW & SW) Based on Dyo et al. (2010)

Battery cost (USD)

Total cost (USD)

Detections per animal

Cost per detection per animal (USD)

155.32

370.645

56,107

0.0066

131.2

40,958

0.0032

52.9 1.04

20.615

56,107

0.00037

0.52

10.3

56,107

0.00018

Also included, how many animal detections were recorded in a month and how much each of these recordings did cost. Two main sources of costs were considered, namely maintenance visits to the woods by the zoologists and the costs involved in battery consumption and charging. Developing the new software and hardware for each stage also adds to the total cost, however, this was excluded from the evaluation. • Stage 1 (Sect. 8.2.2) is the initial 26-hardware nodes deployment. Kept logs recorded how much money was spent on batteries and how much time was spent in the woods. Each detection node is made up of an RFID reader and a Tmote Sky. Tmote Sky is powered by AA-size batteries, while the reader is powered by an 18 Ah 12 V SLA battery. An expenditure of USD 147 was spent on AA batteries and about USD 8.32 (four times a month, using 0.4 kWh for 20 cents/ kWh) for recharging the reader batteries on all 26 detection nodes. From the logs, about 29.7 h per month were spent in the woods costing USD 215.325 (for USD 7.25 hourly wage). An overall sum of USD 370.645 was thus incurred. From the database, it was possible to collate the total number of active tags per month during the deployment, as well as the number of detections per month; thus on average, one animal generated 56,107 records per month, giving a single detection cost of around 0.66 cents. The bottleneck was in the 1 MByte storage on the detection node, without compression; this, depending on activity, fills up within a week. However, using the proposed data compression technique (Sect. 8.2.1.2.1), it was possible to double the lifetime of the nodes, thus requiring only two field visits per month for a total of 10.8 h at a USD 78.3 cost (for USD 7.25 hourly wage). The adaptive duty-cycling approach allowed the battery cost to be reduced to about USD 52.9. A total sum of USD 131.2 per month is reached. Slightly fewer records were generated, but a single record cost was 0.32 cent, which is less than the 0.66 cents cost in Stage 1 with hardware only.

8.2 Evolution and Sustainability of a Wildlife …

633

• In stage 2 (Sect. 8.2.3), the new hardware increased radically the lifetime of the detection node, while yielding the same number of sightings as in Stage 1. In the first deployment of Stage 2, the hardware was revised for testing, without any software enhancement, such as duty–cycling either the radio or the reader. A node lasted for two months on the same battery, and due to its extensive memory capacity, it did not require data download. Since the new hardware used one large rechargeable car battery, this refuted the need to purchase AA batteries for the motes. The charging costs of the car batteries amounted to 0.2 cents/kWh  0.4 kWh  26  0.5 (once in two months) = USD 1.04 per month. On the average, one visit lasted about 5.4 h every two months, averaging to 2.7 h per month (for USD 7.25 hourly wage). As it was needed to visit the nodes once in two months, the monthly cost was 2.7 h  7.25 + 1.04 = USD 20.615. Accordingly, the cost of a single detection was reduced to 0.037 cents. Interestingly, the cost of getting to the woods or tagging the animals is higher than the maintenance cost. The introduction of the enhanced software in Stage 2 further extended the lifetime of the new hardware. A twofold increase in the lifetime of the node was realized; hence, only one visit in every four months became required. This resulted in a maintenance cost of USD 10.3 per month; the cost of a single detection thus became negligible. In the period from March 14, 2009 to September 19, 2009, over 29 million records were collected.

8.2.5

Gained Experience

This project went from an initial design up to hardware and software evolutions through an iterative process of software and hardware designs and developments, while maintaining backward design revisions for sake of hardware and software compatibility. The valuable experience conveyable to similar projects may be recapped in multiple items: • Network maintenance should not be an afterthought, but a key consideration in the original design of the system. Otherwise, maintaining a sensor network would end up being far more expensive than building it. • Before delving into algorithmic improvements and strenuous testing of new software, it is important to carefully consider hardware limitations. Sometimes, it is more cost-efficient to replace the hardware platform than to design and test new software for an existing platform. • The benefits of software optimization, such as improving sampling, storage, and data collection algorithms, largely depend on the hardware. An algorithmic improvement that yields significant benefits on one platform may be less efficient or even inapplicable to another.

634

8 Energy Management Projects for WSNs

• Engineering sustainable sensor networks is an iterative process that alternates between hardware and software modifications. These changes must be performed in a controlled manner so that they do not disrupt the data collection process. • No initial deployment will be perfect, thus the best approach for long-term monitoring systems is to design a prototype that can be rapidly deployed using commercial off-the-shelf technology. This is especially important in applications where no prior data had been collected on a similar scale in the same environment. Although the initial prototype might suffer from several practical issues, it is possible to get the system working in the field to collect suitable data that help understanding how the system design could be improved. These observations guide the evolution of the project, allowing a meaningful reduction of the system maintenance cost by increasing the runtime of devices. Moreover, no simulation or laboratory testing is capable of addressing typical real deployment issues. Practically, failures are common, and some failures such as animals interfering with equipment are unquantifiable until the system is actually deployed. • In the evolution of a system, a choice has to be made, whether to improve it gradually or to switch over to a new system entirely; the decisions are basically influenced by the needs of the application. In this project, new improvements were slowly incorporated; testing them went for months in the field, so as to gather continuous data without cavities that could significantly reduce their biological significance. Other applications can tolerate interruptions, which allows to focus efforts on designing and deploying a new and improved version; implying thus step changes in capability and functionality, while upgrading all components of the system simultaneously. • The design of smart protocols and algorithms to reduce message overhead or energy consumption is only useful if it complies with the requirements of the eventual system users. Interactions with the users are not only useful to ensure a system working as expected, but also to provide interesting ideas for optimizations. By understanding the needs of the domain scientists and zoologists in this project, it was possible to tailor the system design, extending thus the lifetime of devices in the network, while still satisfying application requirements. • The results and findings in this project provide an important insight into the workings of a long-lived outdoor sensor deployment.

8.3

Conclusion for Brightness

Starting by the events this chapter has witnessed. In sports, scoring early and late, Liverpool won sixth Champions League title by overcoming Tottenham Hotspur in Madrid, on Saturday, June 1, 2019. Jürgen Klopp and Liverpool claimed the prize

8.3 Conclusion for Brightness

635

they had coveted for so long. Mohamed Salah second-minute penalty and a late Divock Origi goal settled what was not a classic at the Metropolitano Stadium, as Liverpool made up for their defeat in 2018 final against Real Madrid and the deflation of missing out on the Premier League title to Manchester City. In politics, on Monday, June 3, 2019, President Donald Trump and first lady Melania Trump were in the United Kingdom for a three-day state visit. They were welcomed at Buckingham Palace by Queen Elizabeth II, Prince Charles and Charles wife Camilla, the Duchess of Cornwall. On Tuesday, June 4th, the President met with outgoing British Prime Minister Theresa May. He travelled on Wednesday, June 5th, to Portsmouth, on the south coast of England, to participate in events marking the 75th anniversary of the D-Day landings. Later on the same week, Trump visited Ireland and France. Back to science and technology. The project presented all along this chapter stressed on the need to factor in the maintenance costs while designing the system, to look carefully at software and hardware interactions, to rapidly deploy the initial prototype, and to interact continuously with domain scientists. The full implementation considered the energy management schemes meticulously offered in Part II of this book and provided a live environment monitoring application. The deployment of WSNs in a variety of real-world applications has turned from scientific vision to seen and felt reality. A multitude of systems have already been deployed, ranging from glacier monitoring (Beutel et al. 2009) to real-time environmental and wildlife tracking (Mainwaring et al. 2002) and (Zhang et al. 2004). Such systems have enabled the collection of spatio-temporal data at unprecedented granularities, and have revolutionized the way in which scientists perform field experiments. At the same time, with the outset of new sensor deployments, the need has come to maintain WSNs over prolonged deployment periods. Low effort maintenance and self-reconfiguration are definitely the idealistic selling points of WSNs. Network maintenance may involve a number of tasks, such as changing batteries, replacing faulty nodes, and collecting data from special-purpose storage or gateway nodes. When the maintenance costs exceed user expectations and budget, there is a need to develop the system and make it sustainable. A complete energy-aware WSN must involve energy harvesting as well as energy management techniques. Efforts are sought in such embodying direction at the research and implementation levels.

8.4

Exercises

1. Review the energy management techniques presented in Part II of this book and are applied in the project presented in Sect. 8.2. 2. Recheck the calculations made in Table 8.3. 3. Find and study more energy monitoring projects.

636

8 Energy Management Projects for WSNs

References Atmel. 2008, January 1. AT45DB041B: 4-Megabit 2.5-Volt or 2.7-Volt DataFlash. (Atmel Corporation) Retrieved August 16, 2016, from http://www.atmel.com/images/doc3443.pdf. Atmel. 2014, January 1. Atmel ATmega640/V-1280/V-1281/V-2560/V-2561/V: 8-bit Atmel Microcontroller with 16/32/64 KB In-System Programmable Flash. (Atmel Corporation) Retrieved September 19, 2016, from http://www.atmel.com/Images/Atmel-2549-8-bit-AVRMicrocontroller-ATmega640-1280-1281-2560-2561_datasheet.pdf. Atmel. 2009, June 1. ZigBit 2.4 GHz Amplified Wireless Modules Datasheet. (Atmel Corporation) Retrieved September 19, 2016, from http://www.atmel.com/images/doc8228.pdf. Beutel, J., et al. 2009. “PermaDAQ: A Scientific Instrument for Precision Sensing and Data Recovery in Environmental Extremes.” In The 8th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN), 265–276. San Fransisco, CA: ACM/IEEE. Buettner, M., G.V., Yee, E., Anderson, and R. Han. 2006. “X-MAC: A Short Preamble MAC Protocol for Duty-Cycled Wireless Sensor Networks.” In The 4th International Conference on Embedded Networked Sensor Systems (SenSys), 307–320. Boulder, CO: ACM. Chen, Z. 2007. Antennas for Portable Devices. Chichester, West Sussex, England: Wiley. CLS. 2016, January 1. Why Choose Argos? (CLS Group, a CNES subsidiary) Retrieved September 19, 2016, from http://www.argos-system.org/argos/why-choose-argos/. Contiki-Developers. 2016, January 1. Contiki: The Open Source OS for the Internet of Things. (Contiki-Developers) Retrieved September 14, 2016, from http://www.contiki-os.org/index. html. Dunkels, A., B., Gronvall, and T., Voigt. 2004. “Contiki-A Lightweight and Flexible Operating System for Tiny Networked Sensors.” In The 29th Annual IEEE International Conference on Local Computer Networks (LCN), 455–462. Tampa, FL: IEEE. Durvy, M., et al. 2008. “Making Sensor Networks IPv6 Ready.” In The 6th ACM Conference on Embedded Network Sensor Systems (SenSys), 421–422. Raleigh, NC: ACM. Dyo, V., and C. Mascolo. 2008. “Efficient Node Discovery in Mobile Wireless Sensor Networks.” In Distributed Computing in Sensor Systems, ed. S. Nikoletseas, B. Chlebus, D. Johnson, and B. Krishnamachari, 478–485. Berlin, Heidelberg, Germany: Springer. Dyo, V., et al. 2010. “Evolution and Sustainability of a Wildlife Monitoring Sensor Network.” In The 8th ACM Conference on Embedded Networked Sensor Systems (SenSys), 127–140. Zurich, Switzerland: ACM. Energizer. 2016, January 1. CR2450 Datasheet. (I. Energizer Holdings, Producer) Retrieved September 6, 2016, from http://data.energizer.com/PDFs/cr2450.pdf. Fahmy, H.M.A. 2016. “Simulators and Emulators for WSNs.” In Wireless Sensor Networks: Concepts, Applications, Experimentation and Analysis, ed. H.M.A. Fahmy, 381–491. Singapore: Springer. Hillman, P. 2016, June 6. European Badger (Meles meles). Retrieved June 2, 2019, from https:// petehillmansnaturephotography.wordpress.com/2016/06/16/european-badger/europeanbadger-meles-meles-08/. Hui, J., and D. Culler. 2008. “IP is Dead, Long Live IP for Wireless Sensor Networks.” In The 6th ACM Conference on Embedded Network Sensor Systems (SenSys), 15–28. Raleigh, NC: ACM. Investopedia. 2016, January 1. Original Equipment Manufacturer—OEM. (Investopedia, LLC) Retrieved September 17, 2016, from http://www.investopedia.com/terms/o/oem.asp. Johnson, T. 2016. January 1. Epoxy Resin. (About.com) Retrieved September 6, 2016, from http:// composite.about.com/od/Resins/a/Epoxy-Resin.htm. Kaelbling, L., M., Littman, and A., Moore. 1996, “Reinforcement Learning: A Survey.” Journal of Artificial Intelligence Research 4: 237–285. Kenward, R. 2001. A Manual for Wildlife Radio Tagging. London, UK: Academic Press. Kugelstadt, T. 2016, October 1. The RS-485 Design Guide. (Texas Instruments, Inc.) Retrieved June 19, 2019, from http://www.ti.com/lit/an/slla272c/slla272c.pdf.

References

637

Levis, P., and N. Lee. 2003, September 17. TOSSIM: A Simulator for TinyOS Networks. Retrieved April 14, 2015, from http://www.tinyos.net/tinyos-1.x/doc/nido.pdf. Mainwaring, A., D. Culler, J. Polastre, R. Szewczyk, and J. Anderson. 2002. “Wireless Sensor Networks for Habitat Monitoring.” In The 1st ACM International Workshop on Wireless Sensor Networks and Applications (WSNA), 88–97. Atlanta, GA: ACM. Moteiv. 2006, November 13. Tmote Sky Data Sheet: Ultra Low Power IEEE 802.15.4 Compliant Wireless Sensor Module. Retrieved June 2, 2019, from Moteiv: https://insense.cs.st-andrews. ac.uk/files/2013/04/tmote-sky-datasheet.pdf. National Sleep Foundation. 2019, January 1. What is Circadian Rhythm? (National Sleep Foundation) Retrieved June 3, 2019, from https://www.sleepfoundation.org/articles/whatcircadian-rhythm. Perkins, C., and P. Bhagwat. 1994. “Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers.” In The Conference on Communications Architectures, Protocols and Applications (SIGCOMM), 234–244. London, UK: ACM SIGCOMM. Sadler, C., and M. Martonosi. 2006. “Data Compression Algorithms for Energy-Constrained Devices in Delay Tolerant Networks.” In The 4th International Conference on Embedded Networked Sensor Systems (SenSys), 265–278. Boulder, CO: ACM. Sensirion. 2019, January 1. Digital Humidity Sensor SHT7x (RH/T). (Sensirion) Retrieved June 2, 2019, from https://www.soselectronic.com/productdata/45/89/4/45894/SHT7x.pdf. SourceForge. 2012, July 24. Contiki 2.6: RF230 Hardware Level Drivers. (SourceForge) Retrieved September 14, 2016, from http://contiki.sourceforge.net/docs/2.6/a01750.html. Suel, T., and N. Memon. 2002. Algorithms for Delta Compression and Remote File Synchronization. Academic Press. Book Chapter. Texas Instruments. 2015, June 1. Understanding the I2C Bus. (Texas Instruments, Inc.) Retrieved September 6, 2016, from http://www.ti.com/lit/an/slva704/slva704.pdf. The University of Oxford. 2016, January 1. Wytham Woods. (The University of Oxford) Retrieved September 5, 2016, from http://www.wytham.ox.ac.uk. Tseng, Y.-C., F.-J. Wu, and W.-T. Lai. 2013. “Opportunistic Data Collection for Disconnected Wireless Sensor Networks by Mobile Mules.” Ad Hoc Networks 11 (3): 1150–1164. Wavetrend. 2004, March 18. Tag Domino: L-TG100 / L-TG100 MS. Retrieved September 6, 2016, from http://www.sourcesecurity.com/datasheets/wavetrend-domino-tag-l-tg100.1/co-1816-ga/ L-TG100.pdf. Werner-Allen, G., S., Dawson-Haggerty, and M., Welsh. 2008. “Lance: Optimizing High-Resolution Signal Collection in Wireless Sensor Networks.” In The 6th ACM Conference on Embedded Network Sensor Systems (SenSys), 169–182. Raleigh, NC: ACM. WhatIs.com. 2016, January 1. Order of Magnitude. (TechTarget) Retrieved September 5, 2016, from http://whatis.techtarget.com/definition/order-of-magnitude. Zhang, P., C.M. Sadler, S.A. Lyon, and M. Martonosi. 2004. “Hardware Design Experiences in ZebraNet.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 227–238. Baltimore, MD: ACM. Ziv, J., and A. Lempel. 1977. “A Universal Algorithm for Sequential Data Compression.” Transactions on Information Theory 23 (3): 337–343.

Chapter 9

WSNs Energy Testbeds

Go on working as if lifetime is forever …

9.1

Functionalities

One of the key characteristics that define WSNs is the focus on energy consumption and the efforts to reduce it as much as possible, in order to deploy large-scale networks built out of small, autonomous devices operating without human assistance. Typical WSN deployment scenarios involve a multiplicity of nodes powered by small batteries, such that the average energy consumption remains confined to tens of lW, which keeps repairing or dispensing of non-repairable nodes within economic feasibility. Typically, sensor nodes consume energy in the order of mW when processing and tens of mW when communicating. The focus of the sensor network community on energy efficiency has produced a string of novel MAC, routing, and data-aggregation protocols. WSN power consumption using such protocols has mainly been assessed through simulations, i.e., by counting the fraction of time spent in sending, receiving, and computing, and then multiplying it by numerals taken from datasheets or isolated single-node power measurements. Common approaches for achieving energy efficiency in the newly introduced protocols include duty-cycling the radio at the MAC layer (Chap. 4) and sleep/ awake scheduling at the routing/clustering layer (Chap. 5). Such approaches can be very effective with researchers claiming energy savings of a factor of ten to a hundred, or even more, over conventional protocols like IEEE 802.11 (CSMA/CA). The real gains, however, remain yet to be determined as most results are based on simulations using coarse energy–consumption models. In the case of MAC protocols, for instance, statistics are maintained about what percentage of time the

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_9

639

640

9 WSNs Energy Testbeds

radio is used in the states of transmitting, receiving, and listening for incoming traffic. These statistics are used to compute the energy a node consumes through consulting the datasheets for the current drawn in each of these radio states. A slight improvement is to measure the power consumption of each state using an expensive oscilloscope hooked up to a node performing a Ping–Pong test where the node is alternating between states. For sake of easiness or practicality, WSN research mainly depended on simulation results; only 10% of the papers mention the term testbed, with an even smaller fraction actually using a testbed for experimental evaluation (Haratcherev et al. 2008). As WSN technology is maturing, the number of experimental testbeds rapidly expanded, as well as the number of nodes within. Nowadays, testbeds with more than 100 nodes are readily available; namely, MoteLab (Werner-Allen et al. 2005), Kansei (Arora et al. 2006), TWIST (Handziski et al. 2006), provide researchers with a sure grip on the particularly unpredictable wireless channel. In that respect, the deployment support network approach (Dyer et al. 2007) offers the unique possibility to exercise and debug protocols in real-world conditions by attaching wireless monitoring nodes instead of running cables to each sensor node. However, when it comes to determining the energy efficiency of a protocol, few testbeds provide support to actually measuring the power consumed by each node. An in-depth elaborate comparative presentation of WSN testbeds is presented in (Fahmy 2016b). With the tendency toward experimentation on testbeds, as motivated by the lack of realism regarding the simulation models of wireless channels, some testbeds like MoteLab (Werner-Allen et al. 2005) include one node connected to a digital multimeter for accurately measuring the node power consumption. This single probe provides valuable insight into the actual behavior of an application, but it is hard to expand the scope to the full testbed due to the costs of the measuring equipment. Some extrapolation, however, is possible by feeding the measured power consumption levels into Eq. 9.1 while actually recording at each node the time spend in each state. Energy ¼

X

Pstate j  tstate j

ð9:1Þ

state j

where Pstate j is the power consumed in state j, and tstate j is the time spent in state j. A large-scale study is required to determine the validity of this approach, which implies the need for low cost, yet accurate equipment to measure the power consumption of an individual node (Polastre et al. 2004; Jiang et al. 2007). In the coming section, a focus is accorded to PowerBench, a testbed that provides a benchmarking for power consumption in WSNs.

9.2 Typical WSNs Energy Testbed

9.2 9.2.1

641

Typical WSNs Energy Testbed PowerBench: A Scalable Testbed Infrastructure for Benchmarking Power Consumption

PowerBench developed at Delft University of Technology, The Netherlands, is a scalable testbed infrastructure for benchmarking power consumption (Haratcherev et al. 2008). It is centered on a low-cost interface board capturing the power consumption of a target TNOde in the testbed, by means of a shunt resistor and a 30 lA resolution A/D converter. The TNOde is a MICA2 clone (Crossbow 2002) and (Hill and Culler 2002). Up to eight interface boards are connected to a modified Linksys NSLU2 device (Linksys 2008), an embedded Linux platform, sampling the output of the A/D converters in parallel at a rate of 5 kHz. The samples are time-stamped and sent out over an Ethernet backbone to a central host storing the power data of the complete testbed. The configuration consists of 24 nodes, distributed over four rooms with six nodes per room, generating a continuous feed of 180 kbyte/s. After each experimental run, the power traces can be graphically displayed for comprehensive analysis, or processed into a set of per-node statistics (average power consumption, etc.). Preliminary experience with the PowerBench testbed included the eye-catching observation, that basic three-level (three-state) timer-based estimations, can match true measured power consumption values within a few percent. The level of detail, i.e., the number of states would vary from elementary two states (RX/TX), to practical three states (RX/TX/SLEEP)1 (ARMmbed 2016), up to an elaborate six or more states when taking radio state transitions into account. Eventually, this approach is adopted into mainstream network simulators like ns-2 (ISI 2011) and GloMoSim/Qualnet (SCALABLE Network Technologies 2014), as well as WSN specific simulators like Prowler (Simon et al. 2003), PowerTOSSIM (Shnayder et al. 2004), and AEON (Landsiedel et al. 2005). More on network and WSN simulators are available in (Fahmy 2016a). Besides estimating the power consumption of the radio, both PowerTOSSIM and AEON also consider the other hardware components in a sensor node, such as, CPU, sensors, and LEDs. The resulting level of accuracy in estimating power consumption was determined to be within 5% in AEON and 10% in PowerTOSSIM for a number of TinyOS applications running on a single node (Haratcherev et al. 2008). AEON has the edge as it

1

The difference between IDLE and SLEEP modes is in the amount of power the chip uses. In IDLE mode, the IC is still ON but is not executing any instructions. In SLEEP mode, the IC basically shuts itself down by turning OFF the clock as well as all instructions execution. It can only be brought out of SLEEP mode by a watchdog timer (WDT) interrupt or another external interrupt. A WDT is a hardware timer often used to automatically reset an embedded device that hangs because of a software or hardware fault. Some systems may also refer to it as a computer operating properly (COP) timer. Many microcontrollers have watchdog timer hardware.

642

9 WSNs Energy Testbeds

is emulating at the instruction level, while PowerTOSSIM performs a discrete event simulation at the basic block level.

9.2.1.1

PowerBench Design

The primary objective of designing the PowerBench infrastructure was to measure the power consumption of all 24 nodes in the testbed. Thus, several requirements must be met: • Achieving a low-cost solution. TNOdes, which are similar to the familiar MICA2 nodes, are used in combination with a separate programmer board powering a TNOde and offering convenient wired access by means of a USB interface for programming and serial I/O. The idea was to redesign the programmer to provide the additional functionality of measuring the power consumption of the TNOde it hosts. The target stood at keeping low the costs of the additional components, such as A/D converter, roughly at 25% of the total price of a programmer. • Realizing minimum accuracy and time resolution of the power samples as derived from the objective to study energy-efficient MAC protocols. An effective carrier sense operation with the CC1000 radio takes about 0.5 ms (Polastre et al. 2004), so a sampling rate in the order of 5–10 kHz would suffice to observe these shortest events. To limit the measurement errors to 5%, knowing that the base current drawn by an active TNOde is about 5 mA (CPU idle, radio OFF), the sampling resolution must be below 250 lA. • Viewing the power data after it has been captured by the testbed. Such offline processing requires accurate time stamping of batches of raw power samples. Some local processing would reduce the delay between measurement and time stamping. Also, it is essential that the stream of power data can be aligned with the serial output of the TNOde such that internal protocol events can be used to elucidate particular power profiles. The finest granularity of events is in the order of milliseconds, which translates to time synchronization in the order of 100 ls. Figure 9.1 illustrates the basic architecture of the PowerBench testbed. The central component is a pair of Linksys NSLU2 devices, a low-cost embedded Linux platform, accessible by Ethernet. One of them controls the application running on the TNOdes in the testbed; it can be instructed to install specific program images, start/stop execution, and to handle serial I/O. The second Linksys device is dedicated to capturing the raw ADC samples produced by the modified programmer board. The basic building block of the PowerBench infrastructure, the pair of Linksys NSLU2 devices, can host up to eight TNOdes and must be replicated to support larger configurations. PowerBench testbed was installed into the ceilings of four rooms; it consists of 24 nodes and eight Linksys devices (one pair per room).

9.2 Typical WSNs Energy Testbed

643 Ethernet

Linksys NSLU2 Control

USB hubs Linksys NSLU2 A/D sampling

4 4

TNOdes

Fig. 9.1 PowerBench architecture (Haratcherev et al. 2008)

Experiments can be started from any PC connected to the Ethernet backbone, with the power data being streamed back to the PC by the Linksys devices. Besides hardware, the PowerBench infrastructure includes four categories of software programs to make it all work. There are programs for controlling the execution of the applications run on the testbed, for sampling the A/D converters, and a set of tools to process the recorded power consumption traces. Also, an extensive debugging interface using serial I/O was implemented.

9.2.1.2

Experimentation and Outcomes

PowerBench experimentation was carried on two different MAC protocols; specifically, B-MAC (Polastre et al. 2004) and Crankshaft (Halkes and Langendoen 2007). For the B-MAC protocol different polling frequencies were tested (always ON, 12, 5.5 Hz), to check the impact of the radio switching. After calibrating the testbed, the following experiment was conducted to determine the accuracy of the estimations derived from tracing the radio states:

644

9 WSNs Energy Testbeds

• The core CC1000ControlM module (nesdoc 2016) from the TinyOS protocol stack was modified to intercept all state changes of the radio; at every change an internal, high-accuracy timer was read, and the amount of time spent in the current state (RX/TX/SLEEP) updated. This corresponds to the simple three-state model commonly used to estimate the energy consumption of MAC protocols. • The resulting statistics are periodically output over the serial link. • The measured time in each radio state is multiplied by a combined figure for CPU and radio power consumption for that state. This is because the CPU at each node consumes more energy during transmission and reception than during sleep, for two reasons: – The radio on the nodes is a byte-based radio, which means that each byte received or transmitted causes an interrupt on the CPU. – To receive that interrupt the CPU cannot go to its lowest power mode, but instead, it has to stay in idle mode. Interestingly, out of PowerBench experimentation twofold knowledge was acquired: • Comparing the traditional method of estimating energy consumption, as based on counting the time spent in each radio state, with the true measurements provided by PowerBench. It was found that the estimates could be perfectly tuned (calibrated) such that the resulting errors reside mostly below 2%, provided the CPU is used sparingly. • A graphical display of the power traces is an effective means to study and debug protocol behavior; in particular, internode-related timing issues can be easily viewed from the state changes (IDLE/COMPUTE/RX/TX) embodied in the power data.

9.3

Conclusion for Brilliance

This chapter celebrated Queen Elizabeth II 93rd birthday, she was born on April 21, 1926. But every year, she also has a public birthday celebration during the month of June. Why the second date? It all comes down to the weather; a weekend in the summer is the only time for a parade. This year all red parade, also known as Trooping the Color, took place on Saturday, June 8, 2019. The festivities begun at 10:30 am local time and were attended by the royal family. Testbeds are representative of WSNs, they support the diversity of their hardware and software constituents, they are deployed in the same conditions and would be environment, they make use of the protocols to be used at a larger scale. Testbeds are intended to safeguard would be implemented WSNs from malfunctions that may not be seen in theoretical simulations. Malfunctions may be in

9.3 Conclusion for Brilliance

645

inconvenient hardware, buggy software, and deployment prone to energy depletion and radio interferences. By momentarily tolerating faults that cannot be accepted in everyday actual WSNs, testbeds find the curing solutions. In the literature many testbeds are reported, not all are typically implemented, not all are available now. Knowledge is to be acquired from those who got it by researching, trying, and experimenting; this chapter considers PowerBench with the authentic information it provided. Pioneering testbeds continue to offer models in concepts, implementation, and applications. Some of the testbeds are built for general use, while others are meant for typical applications such as visual surveillance and energy benchmarking.

9.4

Exercises

1. Dig the literature for more WSN energy testbeds. 2. Differentiate between WSN testbeds and WSN energy testbeds.

References ARMmbed. 2016. WatchDog Timer. mbed. https://developer.mbed.org/cookbook/WatchDogTimer. Accessed 20 July 2016. Arora, A., E. Ertin, R. Ramnath, M. Nesterenko, and W. Leal. 2006. “Kansei: A High-Fidelity Sensing Testbed.” Internet Computing (IEEE) 10 (2): 35–47. Crossbow. MICA2. 2002. http://www.eol.ucar.edu/isf/facilities/isa/internal/CrossBow/DataSheets/ mica2.pdf. Accessed 3 Feb 2014. Dyer, M. et al. 2007. “Deployment Support Network: A Toolkit for the Development of WSNs.” In The 4th European Conference on Wireless Sensor Networks (EWSN), 195–211. Delft, The Netherlands: Springer. . Fahmy, Hossam M.A. 2016a. “Simulators and Emulators for WSNs.” In Wireless Sensor Networks: Concepts, Applications, Experimentation and Analysis, 381–491. Springer. Fahmy, Hossam M.A. 2016b. “Testbeds for WSNs.” In Wireless Sensor Networks: Concepts, Applications, Experimentation and Analysis, 251–379. Springer. Halkes, G.P., and K.G. Langendoen. 2007. “Crankshaft: An Energy-Efficient MAC-Protocol for Dense Wireless Sensor Networks.” In The 4th European Conference on Wireless Sensor Networks (EWSN), 228–244. Delft, The Netherlands: Springer. Handziski, V., A. Köpke, A, Willig, and A. Wolisz. 2006. “TWIST: A Scalable and Reconfigurable Testbed for Wireless Indoor Experiments with Sensor Networks.” In The 2nd International Workshop on Multi-hop Ad hoc Networks: From Theory to Reality (REALMAN), 63–70. Florence, Italy: ACM. Haratcherev, I., G. Halkes, T. Parker, O. Visser, and K. Langendoen. 2008. “PowerBench: A Scalable Testbed Infrastructure for Benchmarking Power Consumption.” In The International Workshop on Sensor Network Engineering (IWSNE). Santorini Island, Greece: ACM/IEEE. Hill, J.L., and D.E. Culler. 2002. “Mica: A Wireless Platform for Deeply Embedded Networks.” IEEE Micro (IEEE) 22 (6): 12–24. ISI. ns-2. ISI. 2011. http://nsnam.isi.edu/nsnam/index.php/User_Information. Accessed 3 May 2015.

646

9 WSNs Energy Testbeds

Jiang, X., P. Dutta, D. Culler, and I. Stoica. 2007. “Micro Power Meter for Energy Monitoring of Wireless Sensor Networks at Scale.” In The 6th International Conference on Information Processing in Sensor Networks (IPSN), 186–195. Cambridge, MA: ACM/IEEE. Landsiedel, O., K. Wehrle, and S. Götz. 2005 “Accurate Prediction of Power Consumption in Sensor Networks.” In The 2nd IEEE Workshop on Embedded Networked Sensors (EmNetS-II). Sydney, Queensland, Australia: IEEE. Linksys. 2008. Network Storage Link for USB 2.0 Disk Drives. Cisco Systems. http://support.you. gr/catalog/33/3330F2E4FC0C764CB803BE02D5F4B99F.pdf. Accessed 10 July 2014. nesdoc. 2016. Component: CC1000ControlM. http://socs.acadiau.ca/*shussain/wsn/apps/tos1/ mica2/tos.platform.mica2.CC1000ControlM.nc.html. Accessed 17 July 2016. Polastre, J., J. Hill, and D. Culler. 2004. “Versatile Low Power Media Access for Wireless Sensor Networks.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 95–107. Baltimore, MD: ACM. SCALABLE Network Technologies. 2014. Qualnet. SCALABLE Network Technologies, Inc. http://web.scalable-networks.com/content/qualnet Accessed 8 May 2015. Shnayder, V., M. Hempstead, B.-R. Chen, G.W. Allen, and M. Welsh. 2004. “Simulating the Power Consumption of Large-Scale Sensor Network Applications.” In The 2nd International Conference on Embedded Networked Sensor Systems (SenSys), 188–200. Baltimore, MD: ACM. Simon, G., P. Völgyesi, M. Maróti, and A. Lédeczi. 2003. “Simulation-Based Optimization of Communication Protocols for Large-Scale Wireless Sensor Networks.” In IEEE Aerospace Conference. Big Sky, MT: IEEE. Werner-Allen, G., P. Swieskowski, and M. Welsh. 2005. “MoteLab: A Wireless Sensor Network Testbed.” In The 4th International Symposium on Information Processing in Sensor Networks (IPSN). Los Angeles, CA: ACM/IEEE, 2005.

Part IV

Ignition

Chapter 10

Last Flare

Watch out every step… It will be needed later.

WSN is a wealthy field of research and implementation; it is more and more challenging when energy comes out as a must care concern. Throughout this book, the concepts of energy and energy harvesting, the energy management perspectives, and the energy harvesting and management projects and testbeds, are presented in full details as made available in the literature. All definitions and terminologies are stressed out. Neatly drawn figures assist in viewing and imagining the offered topics. The exercises at every chapter end are built to emphasize must-know concepts and knowledge, and to highlight the needed research work in specific topics, as well as to foster implementing WSN applications that make use of energy harvesting and management techniques. References at the finale of every chapter were checked and double-checked for accuracy. Three parts form the backbone of this book. Part I (Concepts and Energy Harvesting) includes Chaps. 1 and 2, for a review of WSNs concepts and in-depth presentation of the energy harvesting techniques. Part II (Energy Management Perspectives) embodies Chaps. 3, 4, 5, and 6, for a thorough analysis of the three perspectives on energy management; specifically, duty-cycling, data-driven, and mobility-based approaches. Part III (Harvesting and Management Projects and Testbeds) containing Chaps. 7, 8, and 9 brings practice to theory through energy harvesting and management projects and testbeds. This last chapter ignites the launch into the wide realm of study, research, and implementation of energy-focused protocols and techniques for energy harvesting and management. A longer WSN lifetime is the prime target. WSNs are not just theories, a broad spectrum of WSN technologies and industry products are available; also, a wide diversity of manufacturers engages in the market with astounding innovations. The more WSNs are getting ubiquitous and pervasive, the more energy-related research and implementation are becoming vital. Energy is the life of WSNs.

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8_10

649

650

10

Last Flare

Energy projects that include coordinated energy harvesting and energy management techniques need more activities to promote reliable diversified WSN applications. This book did expose at the most possible extent all needed knowledge; it is a helper and mentor at more than a level. It is for senior undergraduates willing to understand energy issues in WSNs and build their graduation projects. Also, it is intended for graduate students making a thesis and in need for specific knowledge on WSNs and the related energy harvesting and management techniques. Moreover, it is targeting researchers and practitioners interested in features and applications of WSNs, and on the available energy harvesting and management projects and testbeds. When I go to bed for a sleep, a popping idea wakes me up, restless nights, lengthy, in hot and cold. When the book becomes a fact another turmoil takes my nights, what could it be? Another book? A call for rest? Is author satisfaction a concern that boosts more endurance for the painful writing devotion? There should be an end… This book is one more hop forward…

Index

A ACC. See Active congestion control Accuracy, 502 Acknowledgement, 138, 139, 164, 199, 206, 207, 210, 503, 545, 619 ACK, 29 Acoustic energy harvesting. See Energy harvesting mechanisms Active congestion control, 30 Adaptive model selection for time-series prediction in WSNs (AMS). See Time-series forecasting approaches Adaptive sampling for energy conservation in WSNs for snow monitoring applications. See Adaptive sampling Adaptive sampling. See Energy-efficient data acquisition adaptive sampling for energy conservation in WSNs for snow monitoring applications, 342 event-sensitive autonomous adaptive sensing and low-cost monitoring in networked sensing systems (e-Sampling), 349 Adaptive self-configuring sensor networks topologies. See Connectivity-driven protocols A Hybrid MAC for WSNs (Z-MAC). See Hybrid MAC protocols Airflow of respiration. See Biochemical energy harvesting Algorithmic approaches. See Data prediction protocols

buddy, 331–335, 339, 378, 385 energy-efficient data collection in distributed sensor environments (EEDC), 316 A lightweight medium access protocol. See TDMA-based MAC protocols AmbiMax. See Energy harvesting projects An adaptive energy-efficient MAC protocol for WSNs (T-MAC). See Contention-based MAC protocols An efficient lossless compression algorithm for tiny nodes of monitoring WSNs (LEC). See Data compression protocols Application layer, 21, 25, 30, 34, 35, 50, 110, 133, 134, 169, 247 Approximate data collection in sensor networks using probabilistic models (Ken). See Stochastic approaches Asynchronous schemes. See Sleep/Wakeup protocols asynchronous wakeup for ad hoc networks (AWP), 156 random asynchronous wakeup protocol for sensor networks (RAW), 160 Asynchronous wakeup for ad hoc networks (AWP). See Asynchronous schemes B Biochemical energy harvesting. See Energy harvesting mechanisms airflow of respiration, 77 chemical energy sources, 78 physical energy sources, 76

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8

651

652 thermal gradient, 76 Buddy. See Algorithmic approaches C Characteristics and classifications of the harvestable energy sources. See Energy harvesting concepts and components Chemical energy sources. See Biochemical energy harvesting Clock synchronization, 30, 31, 35, 125, 142, 155, 159, 166, 230, 231, 236, 238, 240, 241, 243, 251 Cluster-based data aggregation protocols. See In-network processing protocols Conditions for energy-neutral operation. See Energy harvesting concepts and components Congestion, 29, 30, 33, 35, 225, 238, 239, 242, 350, 524 Congestion control, 29, 30, 33, 35 end-to-end, 20, 21, 34 hop-by-hop, 30 Connectivity-driven protocols. See Topology control protocols adaptive self-configuring sensor networks topologies (ASCENT), 120 degree-dependent energy management algorithm (DDEMA), 130 Naps, 123–126, 130, 248 Span, 117–120, 124, 125, 248, 324 uncoordinated power saving mechanisms with latency considerations, 127 Contention-based MAC protocols. See MAC protocols with low duty-cycle an adaptive energy-efficient MAC protocol for WSNs (T-MAC), 197 medium access control with coordinated adaptive sleeping for WSNs (S-MAC), 184 versatile low power media access for sensor networks (B-MAC), 218 Controlled sink mobility for prolonging WSNs lifetime (GMRE). See Controlled sink mobility protocols Controlled sink mobility protocols. See Mobile sink protocols controlled sink mobility for prolonging WSNs lifetime (GMRE), 421 maximizing the lifetime of WSNs with mobile sink in delay-tolerant applications (DT-MSM), 437 Cross-layer protocols, 32

Index D Data aggregation, 14, 24, 25, 31, 32, 34, 35, 151, 166, 178, 185, 202, 260–274, 378, 383–385, 388, 482 fusion, 30, 31, 35, 355–357 Data compression protocols. See Data reduction protocols an efficient lossless compression algorithm for tiny nodes of monitoring WSNs (LEC), 276 Data-driven approach taxonomy, 259 data reduction protocols, 260 energy-efficient data acquisition, 340, 376 Data link layer, 24, 27, 35 Data prediction protocols. See Data reduction protocols algorithmic approaches, 316 stochastic approaches, 286 time-series forecasting approaches, 298 Data reduction protocols. See Data-driven approach taxonomy data compression protocols, 274 data prediction protocols, 285, 339 in-network processing protocols, 261, 274 Degree-dependent energy management algorithm (DDEMA). See Connectivity-driven protocols Derivative-based prediction (DBP). See Model-based active sampling Downstream, 179, 440 Duty-cycling approach taxonomy, 109 power management protocols, 133 topology control protocols, 111, 132 E Electromagnetic transducers. See Energy harvesting from motion and vibration Electrostatic transducers. See Energy harvesting from motion and vibration End-to-end. See Congestion control Energy-aware routing to maximize lifetime in WSNs with mobile sink. See Uncontrolled sink mobility protocols Energy conservation approaches, 103 data-driven techniques, 106 duty-cycling techniques, 105 mobility-based techniques, 106 Energy constraints. See Energy harvesting Energy consumption, 103–106 Energy depletion. See Packet loss Energy efficiency, 104, 168, 173, 243, 400, 451

Index Energy-efficient data acquisition. See Data-driven approach taxonomy adaptive sampling, 342 model-based active sampling, 360 multi-level and cooperative sampling, 351 Energy-efficient data collection in distributed sensor environments (EEDC). See Algorithmic approaches Energy harvesting, 41, 43, 48, 60, 62, 64, 69–73, 80, 82, 84, 649 energy constraints, 41 energy harvesting concepts and components, 43 energy harvesting mechanisms, 62 MEMS for energy harvesters fabrication, 85 Energy harvesting architectures. See Energy harvesting concepts and components Energy harvesting concepts and components. See Energy harvesting characteristics and classifications of the harvestable energy sources, 58 conditions for energy-neutral operation, 56 energy harvesting architectures, 43 energy harvesting versus battery operated systems, 48 harvesting theory, 55 multi-supply and autonomous energy harvesting, 60 power and energy differentiated, 44 storage technologies, 50 Energy harvesting from hybrid indoor ambient light and thermal energy sources. See Energy harvesting projects Energy harvesting from motion and vibration. See Energy harvesting mechanisms electromagnetic transducers, 66 electrostatic transducers, 64 mechanisms for converting motion and vibration to electricity compared, 67 piezoelectric transducers, 65 Energy harvesting from temperature differences. See Energy harvesting mechanisms pyroelectric energy harvesting, 70 thermoelectric energy harvesting, 69 Energy harvesting mechanisms. See Energy harvesting acoustic energy harvesting, 80 biochemical energy harvesting, 73 energy harvesting from motion and vibration, 64 energy harvesting from temperature differences, 69

653 hybrid energy harvesting, 82, 84 photovoltaic energy harvesting, 64 wind energy harvesting, 71 wireless energy harvesting, 71 Energy harvesting projects, 489, 490 AmbiMax, 55, 61, 84, 550–554, 556, 557, 590, 601 energy harvesting from hybrid indoor ambient light and thermal energy sources, 580 Everlast, 64, 536–539, 544–546, 548, 549, 567, 601 Fleck1, 64, 524, 525, 601 Heliomote, 514, 526, 533–536, 552, 562, 567, 601 micro-solar power sensor networks, 569 Prometheus, 507, 508, 511–514, 536, 552, 556, 557, 567, 601 Solar Biscuit, 64, 514, 517, 601 Sunflower, 558–563, 565–567, 569, 601 ZebraNet, 404, 451, 488–499, 501–504, 506, 595 Energy harvesting versus battery operated systems. See Energy harvesting concepts and components Energy management project wildlife monitoring sensor network, 612 Energy testbeds PowerBench, 640–644 Environmental applications environmental monitoring, 514, 515, 601 Event-sensitive autonomous adaptive sensing and low-cost monitoring in networked sensing systems (e-Sampling). See Adaptive sampling Everlast. See Energy harvesting projects Exploiting mobility for energy-efficient data collection in WSNs (MULEs). See Mobile relay protocols Exploiting sink mobility for maximizing sensor networks lifetime. See Uncontrolled sink mobility protocols Extending the lifetime of WSNs through mobile relays. See Mobile relay protocols F Fault-tolerance. See Performance metrics of WSNs, 7 Fleck1. See Energy harvesting projects Flooding, 17, 28, 29, 430, 490, 493, 503 Flow-aware medium access. See TDMA-based MAC protocols

654 G Geographical adaptive fidelity (GAF). See Location-driven protocols Geographic random forwarding (GeRaF). See Location-driven protocols Gossiping, 28 H Harvesting theory. See Energy harvesting concepts and components Heliomote. See Energy harvesting projects Hop-by-hop, 30 Hybrid energy harvesting for Indoor WSNs. See Hybrid energy harvesting Hybrid energy harvesting methodologies for indoor WSNs. See Hybrid energy harvesting Hybrid energy harvesting. See Energy harvesting mechanisms hybrid energy harvesting for indoor WSNs, 82 hybrid energy harvesting methodologies for indoor WSNs, 84 limitations of single source energy harvesting for indoor WSNs, 82 Hybrid MAC protocols. See MAC protocols with low duty-cycle a hybrid MAC for WSNs (Z-MAC), 231 Hybrid tree/cluster-based data aggregation protocols. See In-network processing protocols Hybrid tree/multipath-based data aggregation protocols. See In-network processing protocols I IEEE 802., 21–23, 28, 118, 141, 155, 159, 167, 172, 183, 184, 186, 190, 193, 194, 218, 219, 245, 315, 335, 563, 573, 639 Implosion, 28, 29, 237 Inductive coupling energy harvesting. See Wireless energy harvesting Industrial, Scientific and Medical, 21 In-network processing protocols. See Data reduction protocols cluster-based data aggregation protocols, 266 hybrid tree/cluster-based data aggregation protocols, 269 hybrid tree/multipath-based data aggregation protocols, 272 multipath-based data aggregation protocols, 270

Index tree-based data aggregation protocols, 262 International Telecommunications Union, 21 L Latency. See Performance metrics of WSNs Limitations of single source energy harvesting for indoor WSNs. See Hybrid energy harvesting Location-driven protocols. See Topology control protocols geographic random forwarding (GeRaF), 113 geographical adaptive fidelity (GAF), 111 M MAC protocols with low duty-cycle. See Power management protocols appraisal of MAC protocols with low duty-cycle, 244 contention-based MAC protocols, 183, 248 hybrid MAC protocols, 230, 248 TDMA-based MAC protocols, 167 Maximizing the lifetime of WSNs with mobile sink in delay-tolerant applications (DT-MSM). See Controlled sink mobility protocols Mechanisms for converting motion and vibration to electricity compared. See Energy harvesting from motion and vibration Medium access control with coordinated adaptive sleeping for WSNs (S-MAC). See Contention-based MAC protocols MEMS for energy harvesters fabrication. See Energy harvesting Metric, 19, 21 Micro-solar power sensor networks. See Energy harvesting projects Mobile ad hoc networks, 7 Mobile relay protocols. See Mobility-based approach taxonomy exploiting mobility for energy-efficient data collection in WSNs (MULEs), 448 extending the lifetime of WSNs through mobile relays, 464 Mobile sink protocols. See Mobility-based approach taxonomy controlled sink mobility protocols, 421 uncontrolled sink mobility protocols, 407 Mobile WSNs, 17 Mobility-based approach taxonomy, 405 mobile relay protocols, 448 mobile sink protocols, 406

Index Model-based active sampling. See Energy-efficient data acquisition derivative-based prediction (DBP), 370 model-driven data acquisition in sensor networks (BBQ), 361 Model-driven data acquisition in sensor networks. See Model-based active sampling Multi-camera coordination and control in surveillance systems. See Multi-level and cooperative sampling Multi-level and cooperative sampling. See Energy-efficient data acquisition multi-camera coordination and control in surveillance systems, 356 multiscale approach for structural health monitoring, 362 Multimedia WSNs, 16 Multipath-based data aggregation protocols. See In-network processing protocols Multiscale approach for structural health monitoring. See Multi-level and cooperative sampling Multi-supply and autonomous energy harvesting. See Energy harvesting concepts and components N Naps. See Connectivity-driven protocols Network layer, 25, 27, 28, 34, 35, 299, 376 Network lifetime. See Performance metrics of WSNs Nickel Cadmium (NiCd), 51 O On-demand schemes. See Sleep/Wakeup protocols pipelined tone wakeup (PTW), 140 sparse topology and energy management (STEM), 135 Optimal wakeup scheduling of data gathering trees for WSNs. See Scheduled rendezvous schemes P Packet loss, 30, 113, 120, 160, 171, 209, 210, 215, 236, 272, 273, 276, 356, 357, 360, 361–378 bit error, 173, 336, 524 energy depletion, 18, 42, 106, 400, 407, 412, 420, 432, 433, 467, 478, 600, 645

655 interference, 138, 239 node failure, 27, 111, 121, 177, 230, 265 Performance metrics, 118, 165, 238, 345, 420, 448, 451, 452, 454, 458, 459, 470, 480, 600 fairness, 30 quality of service, 25, 28, 34, 35 reliability, 29 Performance metrics of WSNs, 19 accuracy, 502 fairness, 30, 104, 168, 184, 185, 219, 226, 239, 242, 243, 248, 251 fault-tolerance, 7, 9, 20 latency, 7, 20, 28, 104, 113, 115, 116, 118, 125, 128, 135, 136, 138–142, 156, 158, 160, 161, 164–166, 173, 175, 181, 183–186, 190–193, 195–197, 205–208, 211–219, 224–226, 228, 229, 239, 243–248, 262, 266, 276, 317, 320, 326, 356–358, 360, 378, 383, 387, 423, 431–435, 448, 449, 451, 452, 455, 458–460, 463, 482, 496, 622, 623, 625 network lifetime, 19 quality of service, 16, 17, 25, 28, 34, 35, 262, 383 reliability, 9, 17, 20, 29, 42, 65, 67, 134, 182, 183, 206, 210, 219, 335, 356, 399, 464, 501, 504, 515, 516, 573, 599, 600 scalability, 20, 104, 215, 451 success rate, 21 Photovoltaic energy harvesting. See Energy harvesting mechanisms Physical energy sources. See Biochemical energy harvesting Physical layer, 24, 26, 32, 34, 35, 171, 172, 174, 176, 179, 180, 185, 186, 248, 470 Piezoelectric transducers. See Energy harvesting from motion and vibration Pipelined tone wakeup (PTW). See On-demand schemes Positioning, 30, 31, 33, 35, 117, 119 Power and energy differentiated. See Energy harvesting concepts and components PowerBench. See Energy testbeds Power management protocols. See Duty-cycling approach taxonomy MAC protocols with low duty-cycle, 166 sleep/wakeup protocols, 134, 165 Prometheus. See Energy harvesting projects Pyroelectric energy harvesting. See Energy harvesting from temperature differences

656 R Random asynchronous wakeup protocol for sensor networks (RAW). See Asynchronous schemes Reliability. See Performance metrics of WSNs RF energy harvesting. See Wireless energy harvesting S Scalability, 20, 104, 215, 451 Scheduled rendezvous schemes. See Sleep/Wakeup protocols optimal wakeup scheduling of data gathering trees for WSNs, 151 wakeup scheduling patterns in WSNs, 143 Sensor management protocol, 32 Sensor protocols for information, 29 Simulators GloMoSim, 641 ns-2, 206, 214, 238–241, 431, 641 Prowler, 641 Simulators and emulators OMNeT++, 176, 197, 202 Sleep/Wakeup protocols. See Power management protocols appraisal of sleep/wakeup protocols, 165 asynchronous schemes, 155 on-demand schemes, 135 scheduled rendezvous schemes, 142 SMP. See Sensor management protocol Solar Biscuit. See Energy harvesting projects Span. See Connectivity-driven protocols Sparse topology and energy management (STEM). See On-demand schemes SPIN. See Sensor protocols for information Stochastic approaches. See Data prediction protocols approximate data collection in sensor networks using probabilistic models (Ken), 286 Storage technologies. See Energy harvesting concepts and components Success rate, 21 Sunflower. See Energy harvesting projects T TDMA-based MAC protocols. See MAC protocols with low duty-cycle a lightweight medium access protocol for WSNs (L-MAC), 172 flow-aware medium access (FLAMA), 176 traffic-adaptive medium access protocol (TRAMA), 168

Index Terrestrial WSNs, 13 Testbeds, 181, 641 Kansei, 640 MoteLab, 640 PowerBench, 640–644 Trio, 64, 576 TWIST, 640 Thermal gradient. See Biochemical energy harvesting Thermoelectric energy harvesting. See Energy harvesting from temperature differences Time division multiple access, 167 Time-series forecasting approaches. See Data prediction protocols adaptive model selection for time-series prediction in WSNs (AMS), 309 time-series forecasting for approximate query answering in sensor networks (PAQ), 299 Time-series forecasting for approximate query answering in sensor networks (PAQ). See Time-series forecasting approaches Topology, 109–111, 117 Topology control protocols. See Duty-cycling approach taxonomy connectivity-driven protocols, 117 location-driven protocols, 111 Traffic bursty, 324, 327 dense, 14, 17, 18, 35 downstream, 30, 440, 571 upstream, 177 Traffic-adaptive medium access protocol. See TDMA-based MAC protocols Transport control protocol, 29 Transport layer, 6, 21, 25, 34, 35 Tree-based data aggregation protocols. See In-network processing protocols Trickle, 52 Types of wireless sensor networks mobile WSNs, 17 multimedia WSNs, 16 terrestrial WSNs, 13 underground WSNs, 14 underwater ASNs, 15 U Uncontrolled sink mobility protocols. See Mobile sink protocols energy-aware routing to maximize lifetime in wsns with mobile sink, 414 exploiting sink mobility for maximizing sensor networks lifetime, 407

Index Uncoordinated power saving mechanisms with latency considerations. See Connectivity-driven protocols Underground WSNs, 14 Underwater ASNs, 15 V Versatile low power media access for sensor networks (B-MAC). See Contention-based MAC protocols W Wakeup scheduling patterns in WSNs. See Scheduled rendezvous schemes Wildlife monitoring sensor network. See Energy management projects

657 Wind energy harvesting. See Energy harvesting mechanisms Wireless energy harvesting. See Energy harvesting mechanisms inductive coupling energy harvesting, 72 RF energy harvesting, 71 Wireless mesh networks, 8–10, 91, 505, 573 Wireless sensor nodes, 11, 550 WSNs standards, 21 6LoWPAN, 573 IEEE 802.15.3, 22 IEEE 802.15.4, 22, 219, 245, 315, 563, 573 ZigBee, 21 Z ZebraNet. See Energy harvesting projects

Index of Abbreviations and Acronyms

A Active Congestion Control (ACC), 30 Adaptive Election Algorithm (AEA), 170 Adaptive Fidelity Energy-Conserving Algorithm (AFECA), 125 Adaptive Self-Configuring Sensor Networks Topologies (ASCENT), 120 Ad Hoc on-Demand Distance Vector (AODV), 268 Amplitude Shift Keying (ASK), 193 Analog to Digital Converter (ADC), 11, 61 Asynchronous Wakeup Protocol (AWP), 156 Automatic Repeat reQuest (ARQ), 210 Autonomous Intelligent Network and Systems (AINS), 526 AutoRegressive Moving-Average (ARMA), 302 Average Minimum Reachability Power (AMPR), 269 B Berkeley-MAC (B-MAC), 218 Bit Error Rate (BER), 524 British Thermal Unit (BTU), 45 C Carrier Sense Multiple Access with Collision Avoidance (CSMA/CD), 266 Center at Nearest Source (CNS), 264 Centre National d’Etudes Spatiales (CNES), 631 Clear Channel Assessment (CCA), 218 Clear to Send (CTS), 189 Cluster-based Energy Conservation (CEC), 332

Clustered Diffusion with Dynamic Data Aggregation (CLUDDA), 270 Complex Programmable Logic Device (CPLD), 559 Computer Operating Properly (COP), 641 Connected Dominating Set (CDS), 125 Control/Charger (CC), 546 CUmulative SUM (CUSUM), 342 D Damage-Sensitive Feature (DSF), 358 Data Collection and location System (DCS), 631 Data Send (DS), 200 Data Success Ratio (DSR), 455 2-Degree-of-Freedom (2-DOF), 80 Delay-Bounded Medium Access Control (DBMAC), 266 Delay Tolerant Mobile Sink Model (DT-MSM), 437 Delay Tolerant Network (DTN), 399 Depth of Discharge (DOD), 51 Derivative-Based Prediction (DBP), 370 Digitally Controlled Oscillator (DCO), 565 Digital Signal Processing (DSP), 61 Digital to Analog Converter (DAC), 61 Dilution of Precision (DOP), 502 Distributed Coordination Function (DCF), 245 Dynamic Probabilistic Model (DPM), 298 E Electric Double Layer Capacitor (EDLC), 50 Electromagnetic Energy Harvester (EM-EH), 87

© Springer Nature Switzerland AG 2020 H. M. A. Fahmy, Wireless Sensor Networks, Signals and Communication Technology, https://doi.org/10.1007/978-3-030-29700-8

659

660 ElectroMagnetic Interference (EMI), 534 Electronic Grade Silicon (EGS), 558 Energy-Aware Distributed Aggregation Tree (EADAT), 266 Energy-Efficient Data Collection (EEDC), 316 Energy Harvesting (EH), 550, 587 Enhanced Directed Diffusion (EDD), 265 Equivalent Series Resistance (ESR), 537, 541 Explicit Contention Notification (ECN), 235 EYES-Medium Access Protocol for WSNs (EMACS), 173 F Fast Fourier Transform (FFT), 344 Fast Path Algorithm (FPA), 148 Flame Retardant (FR), 560 FloodNet Adaptive Routing (FAR), 351 FLow-Aware Medium Access (FLAMA), 176 Frequency-to-Voltage (FV), 553 Future Request to Send (FRTS), 200 G General Purpose Input/Output (GPIO), 61 GNU Linear Programming Kit (GLPK), 443 Graphics interchange file (GIF), 274 Greedy Incremental Tree (GIT), 263 Greedy Maximum Residual Energy (GMRE), 423 H High Contention Level (HCL), 235 Highly Dynamic Destination-Sequenced Distance-Vector routing (DSDV), 624 Hybrid Energy-Efficient Distributed clustering (HEED), 268 Hybrid Energy Harvesting (HEH), 84 I Identification (ID), 24 Implantable Biomedical Device (IMD), 73 Information Sciences Institute (ISI), 182 Insulation-Displacement Connector (IDC), 574 Inter-Integrated Circuit (I2C), 609 J Joint Photographic Experts Group (JPEG), 275 Joint Test Action Group (JTAG), 563 L Lightweight Medium Access Protocol (LMAC), 172 Lithium-ion(Li-ion), 51 Lithium Polymer (LiPo), 53 Location-based Clustering Scheme (LCS), 270

Index of Abbreviations and Acronyms Low Contention Level (LCL), 235 Low-Energy Adaptive Clustering Hierarchy (LEACH), 167, 267 Low Power Listening (LPL), 219 M Macro-Sensor Electro-Mechanical System (MSEMS), 558 Maximum Power Point (MPP), 570 Maximum Power Point Tracking (MPPT), 61 Maximum Slot Number (MSN), 233 Mean Relative Error (MRE), 345 Media Access Protocol for Wireless LAN (MACAW), 209 Micro Electro Mechanical System (MEMS), 73 Micro-Energy Harvester (MEH), 73 Micro-scale Thermoelectric Generator (l-TEG), 77 Minimal Dominating Set (MDS), 124 Mixed Integer Linear Programming (MILP), 426 Mobile Adhoc Network (MANET), 7 Mobile Data Collector (MDC), 402 Mobile Relay (MR), 403 Mobile Sink Model (MSM), 438 Mobile Sink (MS), 402 Mobile Ubiquitous LAN Extension (MULE), 448 More to Send (MTS), 206 Multi-Camera Coordination and Control (MC3), 352 Multiple Shortest Path Routing (MSPR), 419 N National Oceanic and Atmospheric Administration (NOAA), 631 Neighbor Protocol (NP), 169 Network Allocation Vector (NAV), 186 Nickel Metal Hydrid (NiMH), 51 Node Activation Multiple Access (NAMA), 172 O Office of Naval Research (ONR), 526 ON-OFF Keying (OOK), 136 Optimal Wakeup Frequency Assignment (OWFA), 155 Original Equipment Manufacturer (OEM), 628 P Pan-Tilt-Zoom (PTZ), 352 PhotoVoltaic (PV), 580 Piecewise Linear Approximation (PLA), 374 PLA, 380

Index of Abbreviations and Acronyms Platform Terminal Transmitter (PTT), 631 POlynomial Regression (POR), 375 Position-based Aggregator Node ELection (PANEL), 269 Power Aware Computing and Communications (PAC/C), 526 Power Aware Multi-Access protocol with Signaling (PAMAS), 245 Power-Efficient Gathering in Sensor Information Systems (PEGASIS), 265 Power Save Mechanism (PSM), 155 Power-Saving Mode (PSM), 335 PREdiction-based MONitoring (PREMON), 332 Primary Receiver (PRX), 545 Primary Transmitter (PRT), 545 Printed Circuit Board (PCB), 534 Probability Density Function (PDF), 291 Programmable Logic Controller (PLC), 370 P-type Metal Oxide Semiconductor Field Effect Transistor (P-MOSFET), 513 Pulse Width Modulation (PWM), 550 R Radio Frequency IDentification (RFID), 615 Radio Trigger Wakeup with Addressing Capabilities (RTWAC), 142 Random Movement (RM), 423 Real-Time Clock (RTC), 627 Recursive Least Square (RLS), 313 Reference Broadcast Synchronization (RBS), 210 Region of Interest (ROI), 352 Request to Send (RTS), 188 Reservoir Capacitor Array (RCA), 546 Reservoir Capacitor (RC), 550 Restricted Input Network Activation Scheme (RINAS), 357 Run-Length Encoding (RLE), 285 S Schedule Exchange Protocol (SEP), 169 Sealed Lead Acid (SLA), 51 Sensor-MAC (S-MAC), 168 Sensor Management Protocol (SMP), 32 Sensor Protocols for Information (SPIN), 29 Sequential Lossless Entropy Compression (S-LEC), 285 Shared Wireless Infostation Model (SWIM), 404 SHM, 357

661 Signal to Noise Ratio (SNR), 26 Silicon-On-Insulator (SOI), 87 Similarity-based Adaptive Framework (SAF), 307 Solar Biscuit (SB), 514 Standard International (SI), 581 Standard Testing Condition (STC), 84 Stateless Non-deterministic Geographic Forwarding (SNGF), 161 Static Sink Model (SSM), 438 Structural Health Monitoring (SHM), 350 SubMiniature version A (SMA), 574 Switched-Capacitor (SC), 540 Switched Mode Power Supply (SMPS), 529 T Thermal Energy Harvesting (TEH), 584 Thermoelectric Power Generator (TEG), 69 Thin Dual-in-line Flat Package (TDFN), 530 Time Division Multiple Access (TDMA), 167 Timeout-MAC (T-MAC), 197 Time to Live (TTL), 520 Tiny AGgregation Service (TAG), 264 Topology Management by Priority Ordering (TMPO), 124 Total Solar Radiation (TSR), 574, 575 Traffic-Adaptive Medium Access Protocol (TRAMA), 168 U Underwater Acoustic Sensor Network (UASN), 15 University of California Los Angeles (UCLA), 182 University of Southern California (USC), 182 V Value-Added Reseller (VAR), 628 Vibration-based Energy Harvester (VEH), 85 W Wakeup Schedule Function (WSF), 157 WatchDog Timer (WDT), 641 Wireless Impedance Device (WID), 360 Wireless LAN (WLAN), 9 Wireless Mesh Network (WMN), 10 Wireless Personal Area Network (WPAN), 21 Wireless Sensor Network (WSN), 10 WSN with Mobile Element (WSN-MEs), 400