Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021 (Lecture Notes in Mechanical Engineering) 9819919878, 9789819919871

Citation preview

Lecture Notes in Mechanical Engineering

Esko Juuso Diego Galar   Editors

Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021

Lecture Notes in Mechanical Engineering Series Editors Fakher Chaari, National School of Engineers, University of Sfax, Sfax, Tunisia Francesco Gherardini , Dipartimento di Ingegneria “Enzo Ferrari”, Università di Modena e Reggio Emilia, Modena, Italy Vitalii Ivanov, Department of Manufacturing Engineering, Machines and Tools, Sumy State University, Sumy, Ukraine Editorial Board Francisco Cavas-Martínez , Departamento de Estructuras, Construcción y Expresión Gráfica Universidad Politécnica de Cartagena, Cartagena, Murcia, Spain Francesca di Mare, Institute of Energy Technology, Ruhr-Universität Bochum, Bochum, Nordrhein-Westfalen, Germany Mohamed Haddar, National School of Engineers of Sfax (ENIS), Sfax, Tunisia Young W. Kwon, Department of Manufacturing Engineering and Aerospace Engineering, Graduate School of Engineering and Applied Science, Monterey, CA, USA Justyna Trojanowska, Poznan University of Technology, Poznan, Poland Jinyang Xu, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China

Lecture Notes in Mechanical Engineering (LNME) publishes the latest developments in Mechanical Engineering—quickly, informally and with high quality. Original research reported in proceedings and post-proceedings represents the core of LNME. Volumes published in LNME embrace all aspects, subfields and new challenges of mechanical engineering. To submit a proposal or request further information, please contact the Springer Editor of your location: Europe, USA, Africa: Leontina Di Cecco at [email protected] China: Ella Zhang at [email protected] India: Priya Vyas at [email protected] Rest of Asia, Australia, New Zealand: Swati Meherishi at [email protected] Topics in the series include: • • • • • • • • • • • • • • • • •

Engineering Design Machinery and Machine Elements Mechanical Structures and Stress Analysis Automotive Engineering Engine Technology Aerospace Technology and Astronautics Nanotechnology and Microengineering Control, Robotics, Mechatronics MEMS Theoretical and Applied Mechanics Dynamical Systems, Control Fluid Mechanics Engineering Thermodynamics, Heat and Mass Transfer Manufacturing Precision Engineering, Instrumentation, Measurement Materials Engineering Tribology and Surface Technology

Indexed by SCOPUS, EI Compendex, and INSPEC All books published in the series are evaluated by Web of Science for the Conference Proceedings Citation Index (CPCI) To submit a proposal for a monograph, please check our Springer Tracts in Mechanical Engineering at https://link.springer.com/bookseries/11693

Esko Juuso · Diego Galar Editors

Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021

Editors Esko Juuso Control Engineering, Environmental and Chemical Engineering University of Oulu Oulu, Finland

Diego Galar Division of Operation and Maintenance Engineering Luleå University of Technology Luleå, Sweden

ISSN 2195-4356 ISSN 2195-4364 (electronic) Lecture Notes in Mechanical Engineering ISBN 978-981-99-1987-1 ISBN 978-981-99-1988-8 (eBook) https://doi.org/10.1007/978-981-99-1988-8 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 This work is subject to copyright. All rights are solely and exclusively licensed 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 Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Preface

The Fifth Conference on Maintenance, Condition Monitoring and Diagnostics, MCMD 2021, continued the series of seminars earlier organized by the University of Oulu and POHTO—The Institute for Management and Technological Training. In the first three seminars, MCMD 2000, MCMD 2005 and MCMD 2010, all the lecturers were personally invited to contribute to the occasion. The seminar series transformed to a combined conference in 2015. The MCMD 2015 was combined with the Conference on Maintenance Performance Measurement and Management, MPMM 2015. University of Oulu, Finnish Maintenance Society (Promaint) and Expomark Oy organized MCMD 2021 in cooperation with International Society for Condition Monitoring (ISCM) and Finnish Society of Automation. MCMD 2021 was a conference of ISCM, which has been active in the previous two events. Both Promaint and Finnish Society of Automation are member societies of ISCM. Together, these societies have almost 3,000 members, including more than 500 companies. Societies build cooperation networks with national and international groups. The conference was organized during the COVID-19 pandemic times. This had a strong effect on paper submissions which led to postponing to February 2021. Originally, the conference was planned to be a physical event, then a hybrid event, but finally a completely virtual event was the only alternative since the pandemic situation became even worse. Virtual conferences were also quite new at that time. Organizing a successful meeting during these challenging times was a great achievement. The MCMD 2021 combined plenary lectures, four technical sessions and a panel discussion on the Future of Condition Monitoring. The plenary lectures were given by three distinguished specialists of condition monitoring: • Condition Monitoring Techniques Based on Numerical Simulation Models— Opportunities and Limitations, Prof. Dr.-Ing. habil. Jens Strackeljan, Rektor, Otto-von-Guericke-Universität Magdeburg, Rektorat, Germany • Improving Process Stability by Advanced Analytics, VP Engineering, Kari Ojala, Asset and Risk Management, SSAB Europe Oy

v

vi

Preface

• Health Management for Aircraft Gearboxes, Prof. Andrew Starr, Centre for Lifecycle Engineering & Management, Cranfield University, UK These proceedings include twelve articles selected and revised from 14 papers presented at the conference. We would like to thank the Management Committee of ISCM for accepting this conference as a conference of ISCM. We are grateful to the Program Committee and reviewers of the conference for their substantial work in referencing papers and providing revising instructions. The Finnish Maintenance Society, Promaint, is acknowledged for organizing the event together with Expomark Oy in cooperation with the University of Oulu, ISCM and Finnish Society of Automation. We also wish to thank our publisher, especially Swati Meherishi, Editorial Director, Applied Sciences and Engineering at Springer Nature, for their support and collaboration in preparing the publication in the Lecture Notes in Mechanical Engineering (LNME). On behalf of the Organizing Committee, we would like to thank all of you who participated in MCMD 2021 and contributed to its success. Oulu, Finland Luleå, Sweden February 2023

Esko Juuso Diego Galar

In Memoriam

Professor Sulo Lahdelma started the MCMD conference series in 2000 and chaired it until 2015. With his work, he contributed to developing this research area at University of Oulu to a high international level. His methodology based on generalizations has provided new solutions for diagnostics, especially for slowly rotating machines and process devices. He brought the generalized time derivatives and integrals with fractal, real and complex orders to the diagnostics and design. These methodologies combine time and frequency domains. The same methodologies operate in a very wide area of applications. He worked actively in the International Society of Condition Monitoring (ISCM) from its foundation, and he was a member of the management committee until he passed away in July 2021. His severe illness didn’t stop his research and publishing activities: his last article was published in August 2021. Professor Lahdelma participated in the MCMD 2021 conference. The audience had an opportunity to hear greetings from the father of the conference series. His work continues to encourage new researcher generations through publications which are available for us via web search engines. On behalf of the conference team, we thank Prof. Lahdelma for introducing this conference and taking care of it for so many years. Esko Juuso February 2023 Diego Galar

vii

Contents

Fault Detection Data Sharing Concept For Electric Car Services: Fleet Level Optimization and Emission Reduction Based on Monitored Data . . . . . . Lasse Metso and Ari Happonen

3

Unsupervised Deep Neural Network Considering the Uncertainties Effect in Pipeline Condition Monitoring Using Guided Ultrasonic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yon Kong Chen, Norhisham Bakhary, Khairul Hazman Padil, and Mohd Fairuz Shamsudin

15

Detecting Gas Injection Problems in Vacuum Tank Degassing Using Measurements of Multiple Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . Juhani Nissilä, Mika Pylvänäinen, Jouni Laurila, Seppo Ollila, Ville-Valtteri Visuri, and Toni Liedes Feature Assessment for a Hybrid Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Antonio Gálvez, Dammika Seneviratne, Diego Galar, and Esko Juuso

31

43

Fatigue and Wear Prediction Operability Forecasting Combining Neural Network and Survival Analysis with an Application to Hot Corrosion in Turbofan . . . . . . . . . . . . Raphaël Langhendries and Jérôme Lacaille

61

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Housam Mohammad, Pavel Mazal, Frantisek Vlasic, and Baraah Maya

73

Fatigue Risk Analysis with Intelligent Digital Twins Based on Condition Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Esko K. Juuso

83

ix

x

Contents

Low-End Hardware in Stress Monitoring of CNC Machines . . . . . . . . . . . Konsta Karioja, Kristóf Lajber, and Esko Juuso

97

Predictive Maintenance Real Order Derivatives and Spectral Norms in Fault Detection . . . . . . . . 111 Esko K. Juuso, Konsta Karioja, and Juhani Nissilä Systematic Methodology for Generating Natural Spall Faults on Rolling Element Bearings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Agusmian Partogi Ompusunggu Physics-Inspired Feature Engineering for Condition Monitoring of Alternating Current-Powered Solenoid-Operated Valves . . . . . . . . . . . . 139 Agusmian Partogi Ompusunggu and Erik Hostens Advanced Prognostics for a Centrifugal Fan and Multistage Centrifugal Pump Using a Hybrid Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Marc Vila-Forteza, Alberto Jimenez-Cortadi, Alberto Diez-Olivan, Dammika Seneviratne, and Diego Galar-Pascual

About the Editors

Esko Juuso Senior Research Fellow Esko Juuso has D.Sc. (Tech.) on Control and Systems Engineering in Department of Process and Environmental Engineering at the University of Oulu. He is Adjunct Professor on Computational Intelligence and ISCM Fellow. He has an M.Sc. (Tech.) degree in Technical Physics (Material Physics) from the University of Oulu, Finland. He is the developer of the linguistic equation (LE) approach and the nonlinear scaling methodology used in various applications. His research interests are in modelling and control of industrial processes with a special emphasis on combining intelligent control, fault diagnosis and performance monitoring into smart adaptive systems (SAS) and cyber physical systems (CPS). He has done research visits to the UK (University of Manchester Institute of Science and Technology - UMIST), Spain (Plataforma Solar de Almeriá - PSA), Germany (University of Dortmund) and France (Institut Français de Méchanique Avancé IFMA). He has been a member of Esprit Working Groups, SiE - Simulation in Europe and FALCON - Fuzzy Algorithms for Control and has been active in European Networks of Excellence (ERUDIT and EUNITE), including the vice-chairman of Production Industries Committee of EUNITE NoE. Dr. Diego Galar He is Full Professor of Condition Monitoring in the Division of Operation and Maintenance Engineering at LTU, Luleå University of Technology where he is coordinating several H2020 projects related to different aspects of cyber physical systems, Industry 4.0, IoT or Industrial AI and Big Data. He was also involved in the SKF UTC centre located in Lulea focused on SMART bearings and also actively involved in national projects with the Swedish industry or funded by Swedish national agencies like Vinnova. He has authored more than 500 journal and conference papers and books and technical reports in the field of maintenance he also works as member of editorial boards and scientific committees, chairing international journals and conferences and actively participating in national and international committees for standardization and R&D in the topics of reliability and maintenance.

xi

Fault Detection

Data Sharing Concept For Electric Car Services: Fleet Level Optimization and Emission Reduction Based on Monitored Data Lasse Metso

and Ari Happonen

Abstract The study showcases a literature review on common sensors used in electric cars and proposes a novel service concept for electric car maintenance and tune up servicing, based on the data sharing concept. Modern cars and electric vehicles are monitored through sensors and control systems utilizing data analysis and vehicle operations automating algorithms. The data sharing concept for electric car services is based on the data, collected from the sensors and systems and then utilized by partners in service network, using cloud based data sharing models. The vehicle status with easy and direct access to services are available when needed and service companies have access to service instructions and training material supported by the car manufacturer. New car manufacturers or brands would have easier access to new markets with independent car service operators. Data stored in cloud allows low cost access for all accepted stakeholders, e.g. car owners, car service providers and car manufacturers. Keywords Electric vehicle · Data sharing · Sensor · Monitoring · Service business · Artificial intelligence · Autonomous driving · Data platform · Cloud service · Digitalization · Digital transformation

L. Metso (B) · A. Happonen LUT University, Lappeenranta, Finland e-mail: [email protected] A. Happonen e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_1

3

4

L. Metso and A. Happonen

1 Introduction The number of electric cars and demand towards new options where to choose an electric vehicle (EV) is growing, because of globally ongoing emission reduction actions. These actions include e.g. near future openly discussed politics on emissions matters, corporate and private level cost saving aims per driven kilometers, new novel products entering electric vehicle markets and overall attitude and opinions have changed towards carbon neutral mobility. On the political and legal side, many countries, e.g. Norway, Denmark, Netherlands, France, United Kingdom, and Germany have set restrictions on using combustion engine cars in city centres and/ or limited their use during rush hours in cities, e.g. Berlin, Hamburg, Copenhagen, Rome, Athens, Paris and Madrid [1, 2]. But still, the current EV market share is low compared to other options as the technology is still developing and only a limited number of EV options and total production capacity are available at the moment. Additionally, the EVs are rather expensive compared to combustion engine cars, and it might be difficult for new EV brand owners to enter the market and be trusted by any brand loyal customers. The clear market advantages for traditional car brands are their loyal customer base and highly visible service network, which they have created with their partners in previous decades. New brands are missing service networks, but with a fast and easy service concept for new EV brands, they could have a way to enter new markets more easily. These sorts of platforms support independent car service operators e.g. with spare parts (lists), maintenance manuals, updated maintenance procedure specifications, car history data and training materials with high efficiency. This is important especially as the EV service is different than cars with combustions engines, meaning that a set of new skills will be needed. Positively for EVs, based on recent studies, EVs need less maintenance than combustion engine cars do [3, 4]. But it also means, there is less available commercial experience available for EV maintenance and challenges in creating new maintenance networks. Key in the data sharing concept is to enable offering of the training material and service instructions to independent service companies. Also, the vehicle owners should be offered base maintenance information, about their car(s), in the online platform. Owner should have more availability for direct access to routine service reminders, details on most nearby qualified service points and up-to-date status details of the vehicle, even remotely. This study focuses on vehicle condition monitoring for electric cars as a continuation of previous data sharing service platform studies [5, 6]. In general, different fleets [7] use fleet unit condition and asset monitoring in industry, both for fixed assets and vehicles. The asset maintenance is based on sensors and data the vehicles generate in unit level and the knowledge and information the organization collects on the fleet level. For example, the real-time sensor data from cars could be used for general road safety action. E.g. sharing the sensor data for icy and dangerous driving conditions for nearby vehicles. In the fleet level, corporations could seek out best performing drivers and offer internal driving advice, guided by the best performers.

Data Sharing Concept For Electric Car Services: Fleet Level Optimization …

5

Through analysis of carbon dioxide emissions per driven kilometers (e.g. 100 km), corporations can find the most ecologic drivers and have them to instruct and educate other drivers in the company fleet and improve organizational level asset management through user education [8].

2 Methods The current research is a literature review. Relevant conference and journal papers are selected based on set filtering rules including the age of publications (to accommodate the need to focus on the current area specific literature) and the reputation of publishing channels (e.g. focusing on Scopus, Elsevier, IEEE, WoS etc., databases, which have a long tradition to publish high quality academic research on focus area specific context).

3 Sensors and Monitoring in EVs 3.1 Autonomous Driving Autonomous driving has multiple improvement goals [9], compared with the traditional manual driving activity, goals like accident reduction. Statistically speaking, human errors are the main reason for traffic accidents. When a driver is tired or intoxicated, they can end up making human errors, which lead to accidents. Utilization of different kinds of autonomous driving systems can reduce the amount of accidents [10]. Autonomous driving is divided by the SAE (Society of Automotive Engineers) standard into six levels. Level 0 there is no automation and the highest level is 5, where all driving related tasks are automated. SAE levels are presented in Table 1 [11–13]. Autonomous vehicles are called self-driving cars, driverless cars or robocars. Speed control [14] can be done using GPS and traffic signals to optimize traffic speed [15]. If a driver fails to monitor the roadway when driving, a level 1 adaptive cruise control (ACC) system keeps the speed but the driver is responsible for the steering. However, the crash avoidance features, including intervention-type active safety systems, may be included in vehicles equipped with driving automation systems at any level. Automated Driving System (ADS) in vehicles at levels 3–5 perform the complete Dynamic Driving Task (DDT). The crash avoidance capability is part of ADS functionality [13]. Both passive and active sensors are needed in autonomous driving for vehicle self-localization. Passive sensors collect signals, the radiation of light from the environment. Passive sensors are for example Global Navigation Satellite System (GNSS) receivers and vision sensors. An example of an active sensor is a Light Detection

6

L. Metso and A. Happonen

Table 1 SAE levels of driving automation [11–13] SAE level

Name

Description

0

No automation The human driver has control and is responsible for monitoring environment and decisions

1

Driver assistance

Steering, braking and acceleration are assisted but the human driver performs dynamic driving. The human driver has control and is responsible for monitoring environment and decisions

2

Partial automation

The driver’s assistance takes control of steering, braking and acceleration. The human driver performs other driving aspects of dynamic driving. The human driver has control and is responsible for monitoring environment and decisions

3

Conditional automation

All driving tasks are performed by an automated driving system. The human driver is responsible of driving and taking control when needed. The driver holds the hand on the steering wheel at all times

4

High automation

An automated driving system takes all control, monitoring and decision of driving. Even if the human driver does not want to take control the system will continue with automated mode

5

Full automation

An automated driving system performs all driving tasks. The human driver has no access to take control of driving

and Ranging (LIDAR) sensor [16, 17]. Passive sensors can be used to improve the global positioning. Camera as a vision-based lane detection and speedometer together with the Inertial Measurement Unit (IMU) can reduce positioning errors [17]. GNSS and inertial sensor integration are solutions for vehicle self-localization [18]. In selflocalization following sensors are needed: laser scanner, camera, digital map, vehicle dynamics and GPS [19]. Levels 3–5 autonomous driving needs a good communication environment. The vehicles are equipped with sensors to collect data, and the collected data is used to support autonomous driving. Autonomous vehicles need diagnostics, e.g. On-Board Diagnostics (OBD) to perform fault diagnosis. Because of the vast amount of data, neural networks are used in real-time diagnostics. Vehicle-to-vehicle communication enables the traffic control and improves the safety in autonomous driving [20]. Autonomous vehicles are integrated systems with functions such as environmental perception, path plan and automatic in-traffic navigation. Intelligent vehicle encompasses pattern recognition, sensor technology, automotive electronics, electrical, computer, machinery and other subjects [21]. According to Elon Musk, Tesla Incorporation will release a full self-driving feature in early 2021, still in the near end of the year (October 2021) this is not the case [9]. So far Tesla cars have an Autopilot which is actually a driver-assistance feature set. This feature set allows a car to park themselves, change lanes, automatically reduce the driving speed and/or apply the needed amount of brakes, and identify stop signs and traffic lights but it really does not make their cars to be true autonomous vehicles [22]. Overall, many car manufacturers have achieved the level 2 in autonomous

Data Sharing Concept For Electric Car Services: Fleet Level Optimization …

7

driving. For real progress, there is e.g. Mercedes-Benz solution, with the new Sclass, which includes autonomous driving at the third level (with some limitations). This feature works only in German autobahns and the speed maximum is limited to 60 km/h (based on current legislation [23]), so it has mostly practical value for e.g. traffic jam situations. Current road legislation does not take into account fully autonomous driving, e.g. EU lacks a mutual type-approval procedure for autonomous driving projects, especially for legislation and cooperation specific matters to agree on the research of the autonomous driving matters [24]. However, several countries have their own attempts to help the development of autonomous driving, through creation of their own national legislations. For example, Germany has modified their existing Road Traffic Act by defining the requirements for highly and fully automated vehicles. In the same manner, France had allowed the testing of autonomous vehicles on public roads from 2019. Additionally, Spain made modifications to national laws offering a legal framework for the development of autonomous driving and Italy has its first law for testing autonomous driving [25]. Weather conditions are a challenge for autonomous driving systems. For example, rainy and foggy conditions are known to affect the functionality of cameras and lidars. Sensors can even end up sending wrong or misleading information to the control systems of the vehicle. For example, fog, raindrops and snowflakes can cause lidar to “detect” obstacles in near range and give data based false alarms. The worst condition for lidar is fog, which is the biggest cause for false alarms. At subzero temperatures the lenses of the cameras can be directly blocked through ice formation and the device can face some mechanical misfunctions or electronics can get water condensation on them. Fog, raindrops and snowflakes can also change the focus points where the camera electronics focus the optics of the system. GPS does not get directly highly influenced by rain or snow, but it can get weaker level and quality signal information from low horizon satellites, e.g. because of additional water content in trees, which will partly block or suppress satellite signals in addition to the problems e.g. tall buildings, mountains or similar structures will do [26]. However, the space weather condition can have a wide range of unexpected effects on GPS signals [27]. For example, the condition in the solar system, especially so collect solar storms. Radar systems use a Millimeter-wave (mm-wave) spectrum, in which rain, snow, mist and hail can have an impact, which reduces the performance of the radar (i.e. less range, worse signal quality and ghosting) [28].

3.2 Safety The systems and sensors, which are in use for autonomous driving can also be used for active safety systems, e.g. automated emergency braking and lane keeping assistance. However, the safety systems can also include solutions which ensure the safe and device protective use of the vehicle. As an example, temperature monitoring sensors for batteries in EVs can prevent over-heating and fire in batteries [29]. In addition to physical, also virtual thermal sensors have been developed [30]. Sensors for driving

8

L. Metso and A. Happonen

control are antilock brake system (ABS), traction control (TCS) and vehicle stability control (VSC) [31]. In the case of an accident, crash sensors activate the switch controls to disconnect the battery [32] and to execute an emergency call. Also, in combustion engine vehicles temperature control is used to prevent over-heating [14]. A technique for added vehicle safety is to utilize multiple sensors in case of any miss-function and to be able to offer backup data for example in case of physical sensor device damages. Additionally, lidars, cameras (visible, infrared etc. light), radars and ultrasonic can be used together to ensure safe autonomous driving in every situation [14, 33]. Different sensor outputs can be used to check the possible validity of other sensors and also to support sensor calibration activities and self-testing processes. Wireless Vehicle-to-vehicle (V2V) communicated [32] information about the speed and position of other cars can avoid crashes, traffic jams and may decrease the amount of pollution.

3.3 Condition Monitoring In modern vehicles, all collected signals can be followed on the vehicle Controller Area Network (CAN). Here typically monitored parameters are engine coolant temperature, engine speed, oil pressure, transmission oil temperature and cooling fan function [35]. Vehicles are complex and need an effective maintenance strategy, where proper corrective maintenance is applied, when a fault has been detected. Preventive maintenance is based on beforehand defined services, that can be based on kilometers driven or a time period. In predictive maintenance, the condition of the vehicle is monitored, and the data is analyzed case by case, based on the service action needs [36]. Bearing failures in electric motors have been reported as one of the main failures in motors and sensors are needed to prevent failures [37] and fault diagnosis sensors are used [38]. In electric cars, control of vehicle operation is based on sensors such as: voltage, steering wheel torque angle, accelerator pedal position, wheel speed etc. This control ensures the continued usability of the car. The sensors are connected and controlled with an Engine Control Unit (ECU). CAN allow the communication of diagnostic protocols from which the most important protocols are ODB2 (OnBoard Diagnostics II) and Unified Diagnostic Service (UDS). System’s condition is evaluated in diagnostic and prognostic processes [36, 39]. A monitoring system takes the baseline measures and records operating parameters. The monitoring system can identify deviations in parameters. The monitoring system can notice statistical differences from normal operating conditions. That may indicate a potential problem and a need for service, for example, low tire pressure [40]. Different kinds of sensors are placed in a car to detect service and maintenance needs. The monitored activity of sensors, such as mechanical triggers, light sensors, motion sensors, magnetic sensors and radio frequency identification tags, can be recorded [41].

Data Sharing Concept For Electric Car Services: Fleet Level Optimization …

9

4 Data Sharing Concept The data sharing concept is based on the idea of data sharing with artificial intelligence (AI) utilizing real-time data from cars. New cars are equipped with sensors which are monitoring the status of the car, and when needed can even take a connection to cloud service. In a similar manner, with the shared database, all known (customer accepted) history and car related fault data and cars sensory based event data could be made available, and AI could then efficiently predict the change in the status or analyze any issues in car condition. The concept is illustrated in Fig. 1 [5, 6]. Sensor based data collection and measuring of specific events is an important success factor for EVs [42]. Mainly the concept needs monitored data but also safety related data like different weather database information, which can be used for current and predicted near future local weather information as third party data sources. The main idea of the service platform is to offer EV owners the real-time data to check the status of their cars as well as to offer service providers a view of the car service history and current status. The data sharing platform can answer the following questions: • • • • •

What services does the EV need right now and in the near future? Which spare parts are needed? When the service is available to be booked? How much does the service cost? Where is the nearest service?

Both car owners and car service providers can have benefits from the data sharing service platform. Car owners will be more aware of the condition of their EVs and they have easy access to the best service offers and booking to the nearest car service

Fig. 1 Data sharing concept (adapted from [5])

10

L. Metso and A. Happonen

hub, at any time. Car owner also has remote access to the car details and status information, through the service platform. The car service providers have access to service instructions and spare part lists through the platform. Service providers can optimize their spare part inventories, as they know the general details on what spare parts are in need, based on booked service orders and sensor data provided in the platform. The car manufacturers have access to new markets while they can use independent car services instead of creating their own service network. The manufacturer can give training videos or other training material on the data sharing concept from which independent car services can learn those and execute the car service. Stakeholders of the data sharing concept are car owners, car services, car manufacturers, spare part suppliers and of course the data sharing concept provider. They all can get benefits by using data sharing concept. Another beneficiary of sensors in cars is the traffic while cars are sharing traffic information. V2V enables wireless information exchange and the conditions of roads can be disseminated to each other, even to those who are not members of the data sharing concept [34]. Collaboration with Bosch and Microsoft has created a base for the vehicle system and the connection from the vehicle to the cloud. Their software platform enables over-the-air updates and allows developer tools for car manufacturers and suppliers to develop their own software [43]. This enables the development of the data sharing concept as well.

5 Discussion and Summary Autonomous driving increases the number of electronic devices in use, sensors in cars and technology complexity in general. These are basically practical phenomena our society has nowadays experienced through digital transformation when we are applying technology to design changes [44] and EVs and data sharing platforms to change how business models work [45, 46] and digitalization is boosting the amount of technology in use in different industries in general [47]. These changes will demand e.g. new kinds of sensors, both passive and active in nature to support the higher levels of self-driving capabilities of future EVs [11–13]. For example, the automated driving system in levels 3-4 needs a lidar sensor, (camera) vision system, good accuracy on devices’ global positioning data accordance measures, detailed map and geological information and inertial measurement units, whereas level 5 most likely needs constant data connectivity for the non-line of sight additional data gathering and sharing and so on. In addition to that, EVs will need special internal operations supportive sensors. For example, to ensure the safe use of the electric car, the temperature of the batteries needs to be monitored and the condition of the car is checked constantly by the vehicle sensors. All in all, one vehicle is quite limited with its capability to collect data from its surroundings (in a reasonable price range). For this, it looks like it would make more sense to further study all time online connectivity and fleet level data collection and sharing to go around one vehicle data

Data Sharing Concept For Electric Car Services: Fleet Level Optimization …

11

analysis, collection and processing limits. Also, the EVs have to move and drive as energy efficiently as possible, as we assume that the current intercity emissions limitations, like in Berlin, Hamburg, Copenhagen, Rome, Athens, Paris and Madrid [1, 2], will formulate in future to be energy usage limitations for EVs to reduce the amount of used energy for people to move inside cities. For that sort of overall optimization, new sorts of data analysis algorithms and support devices will be needed and the online connectivity to affiliation data analyzing platforms will come more common place and probably are generally demanded defacto part of best of the best EVS all in all. Here we assume, that the holistic system is built on top of the previously mentioned sensors, diagnostics and preventive failure analysis capability of the backend fleet level vehicle condition monitoring solutions [36–38], and at least 5G cellular level connectivity in intra cities areas. For data connectivity, we will basically, need more vehicle-to-infrastructure communication technologies [48], more data fusion and processing [49] and also novel advancements to driving strategies [50] for more sustainable self-driving vehicle fleets in the near future. And finally, from a global sustainability point of view, the real and actual emissions data, specific for a given vehicle type, its trim level and usage patterns in specific usage regions need to be recorded. With the data, companies can utilize e.g. University collaboration resources for sustainability related business model innovations [51]. For example, with real fleet level information of fuel/electricity consumption, we can truly start to optimize the emissions our vehicle fleets generate, instead of basing the sustainability affecting decisions on theoretical numbers, measured in laboratories. The data sharing concept, discussed in this study, offers the car owner access to details of the car as well as gives them access to car service information. The owner can easily control and make decisions related to the expenses used for the car services as well as ask service companies to bid to be the one they use for the services. The service companies would have access to the vehicle service history and service instructions and spare part lists. The car manufacturers and new brand owners could use independent car services and they do not have to create their own service network to enter a new market area. The big question is, how will this sort of change affect to the companies, who have to change their organizational cultures, as a result of new business model development needs [52], when the world is moving away from the ICE age towards the EV age? Acknowledgement Authors would like to express our honest thanks for colleague Matti Rissanen for proof reading our work, our sincerest appreciation for LUT School of Engineering Science for providing the financial support to study clean transportation and specially field of electronic vehicles and gratitude for South-East Finland–Russia CBC programme for supporting AWARE project, funded by the European Union, the Russian Federation and the Republic of Finland, supporting our sustainability studies.

12

L. Metso and A. Happonen

References 1. Coren M (2018) Nine countries say they’ll ban internal combustion engines. So far, it’s just words. Quartz 2. Mohareb EA, Kennedy CA (2014) Scenarios of technology adoption towards low-carbon cities. Energy Policy 66:685–693 3. De Clerck Q, van Lier T, Messagie M, Macharis C, Van Mierlo J, Vanhaverbeke L (2018) Total cost for society: a persona-based analysis of electric and conventional vehicles. Transp Res Pt D Transp Environ 64:90–110 4. Gass V, Schmidt J, Schmid E (2014) Analysis of alternative policy instruments to promote electric vehicles in Austria. Renew Energy 61:96-101 5. Metso L, Happonen A, Rissanen M, Efvengren K, Ojanen V, Kärri T (2018) Data openness based data sharing concept for future electric car maintenance services. In: 32nd International congress and exhibition on condition monitoring and diagnostic engineering management (COMADEM2019). Huddersfield, United Kingdom, pp 429–436. https://doi.org/10.1007/9783-030-57745-2_36 6. Metso L, Happonen A, Ojanen V, Rissanen M, Kärri T (2019) Business model design elements for electric car service based on digital data enabled sharing platform, In: 23rd Cambridge International Manufacturing Symposium. University of Cambridge, p 6. https://doi.org/10. 17863/CAM.45886 7. Kinnunen S-K, Happonen A, Marttonen-Arola S, Kärri T (2019) Traditional and extended fleets in literature and practice: definition and untapped potential. Int J Strat Eng Asset Manag 3(3):239–261. https://doi.org/10.1504/IJSEAM.2019.108467 8. Kortelainen H, Happonen A, Kinnunen S-K (2016) Fleet service generation—challenges in corporate asset management. In: Lecture notes in mechanical engineering. Springer, pp 373– 380. https://doi.org/10.1007/978-3-319-27064-7_35 9. Zhou H, Laval J, Zhou A, Wang Y, Wu E, Qing Z, Peeta S (2021) Review of learning-based longitudinal motion planning for autonomous vehicles: research gaps between self-driving and traffic congestion. Transp Res Rec J Transp Res Board. https://doi.org/10.1177/036119812110 35764 10. Dorothy JG (2015) Autonomous and automated and connected cars-oh my! first generation autonomous cars in the legal ecosystem. Minnesota J Law Sci Technol 16 11. Barabás I, Todoru¸t A, Cordo¸s N, Molea A (2017) Current challenges in autonomous driving. In: IOP conference series: materials science and engineering, vol 252(1). p 012096 12. Smith BW, Svensson J (2015) Automated and autonomous driving: regulation under uncertainty. Corporate Partnership Board Report, OECD, International Transport Forum. www.int ernationaltransportforum.org 13. SAE: taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles. SAE International, Standard J3016_201609, 2016-09-30 14. Ananth C, Sudalai C, Jebadurai NE, Saranya SS, Archana T (2014) Intelligent sensor network for vehicle maintenance system. Int J Emerg Trends Eng Dev (IJETED) 3(4) 15. Madhusudhanan AK (2019) A method to improve an electric vehicle’s range: efficient cruise control. Eur J Control 48:83–96 16. Tang J, Chen Y, Niu X, Wang L, Chen L, Liu J, Shi C, Hyyppä J (2015) LiDAR scan matching aided inertial navigation system in GNSS-denied environments. Sensors 7:16710–16728 17. Gu Y, Hsu LT, Kamijo S (2015) Passive sensor integration for vehicle self-localization in urban traffic environment. Sensors 15(12):30199–30220 18. Gu Y, Hsu LT, Kamijo S (2015) GNSS/onboard inertial sensor integration with the aid of 3-D building map for lane-level vehicle self-localization in urban canyon. IEEE Trans Vehic Technol 65(6):4274–4287 19. Schindler A (2013) Vehicle self-localization with high-precision digital maps. In: 2013 IEEE intelligent vehicles symposium workshops (IV Workshops). IEEE, pp 134–139 20. Jeong Y, Son S, Lee B (2019) The lightweight autonomous vehicle self-diagnosis (LAVS) using machine learning based on sensors and multi-protocol IoT gateway. Sensors 19(11):2534

Data Sharing Concept For Electric Car Services: Fleet Level Optimization …

13

21. Lv X, Chen Y (2009) A design of autonomous tracing in intelligent vehicle based on infrared photoelectric sensor. In: 2009 International conference on information engineering and computer science. IEEE, pp. 1–4 22. Levin T (2020) Elon Musk says Tesla will release its ‘full self-driving’ feature as a subscription in early 2021. Business Insider Australia 23. Taylor E (2020) Mercedes bets on evolution as Tesla touts revolution in automated driving. Reuter Technology News 24. Pattinson JA, Chen H (2020) A barrier to innovation: Europe’s ad-hoc cross-border framework for testing prototype autonomous vehicles. Int Rev Law Comp Technol 34(1):108–122 25. Autovista Group: The state of autonomous legislation in Europe (2019) [on-line document] [cited 24.2.2021]. https://autovistagroup.com/news-and-insights/state-autonomous-leg islation-europe 26. Jahkola O, Happonen A, Knutas A, Ikonen J (2017) What should application developers understand about mobile phone position data. In CompSysTech’17. ACM, pp 171-178. https://doi. org/10.1145/3134302.3134346 27. Lehtinen M, Happonen A, Ikonen J (2008) Accuracy and time to first fix using consumergrade GPS receivers. In: 16th International conference on software, telecommunications and computer networks. Split, Croatia, pp 334–340. https://doi.org/10.1109/SOFTCOM.2008.466 9506 28. Zang S, Ding M, Smith D, Tyler P, Rakotoarivelo T, Kaafar MA (2019) The impact of adverse weather conditions on autonomous vehicles: how rain, snow, fog, and hail affect the performance of a self-driving car. IEEE Vehic Technol Magaz 14(2):103–111 29. Sun P, Bisschop R, Niu H, Huang X (2020) A review of battery fires in electric vehicles. Fire Technol 1–50 30. Xiao Y (2014) Model-based virtual thermal sensors for lithium-ion battery in EV applications. IEEE Trans Indust Electr 62(5):3112–3122 31. Nguyen BM, Viehweider A, Fujimoto H, Hori Y (2013) Basic design of electric vehicle motion control system using single antenna GPS receiver. IFAC Proc Vol 46(21):612–619 32. Kumar LS, Kumar RN, Huzefa AC (2017) Collision warning with auto brake & complete high voltage disconnect during crash—a real-life safety perspective. In: 2017 IEEE transportation electrification conference (ITEC-India). IEEE, 1–4. 33. Buyval A, Gabdullin A, Mustafin R, Shimchik I (2018) Realtime vehicle and pedestrian tracking for didi udacity self-driving car challenge. In: 2018 IEEE international conference on robotics and automation (ICRA). IEEE, pp 2064–2069 34. NHTSA (2020) National Highway Traffic safety Administration, [on-line document] [cited 28.12.2020]. https://www.nhtsa.gov/technology-innovation/vehicle-vehicle-communication 35. Rögnvaldsson T, Nowaczyk S, Byttner S, Prytz R, Svensson M (2018) Self-monitoring for maintenance of vehicle fleets. Data Mining Knowl Discov 32(2):344–384 36. Shafi U, Safi A, Shahid AR, Ziauddin S, Saleem MQ (2018) Vehicle remote health monitoring and prognostic maintenance system. J Adv Transp 37. He F, Xie G, Luo J (2020) Electrical bearing failures in electric vehicles. Friction 8(1):4–28 38. Gómez-Peñate S, López-Estrada FR, Valencia-Palomo G, Osornio-Ríos R, Zepeda-Hernández JA, Rios-Rojas C, Camas-Anzueto J (2018) Sensor fault diagnosis observer for an electric vehicle modeled as a Takagi-Sugeno system. J Sens 39. Shi G, Dong P, Sun HQ, Liu Y, Cheng YJ, Xu XY (2017) Adaptive control of the shifting process in automatic transmissions. Int J Automot Technol 18(1):179–194 40. McClellan S (2010) INTHINC Inc, System and method for detecting vehicle maintenance requirements. U.S. Patent Application 12/379,157 41. Boss GJ, Finn PG, Hamilton IRA, O’Connell BM, Seaman JW, Walker KR (2012) Recording, storing, and retrieving vehicle maintenance records. U.S. Patent 8,095,265. International Business Machines Corp. 42. Jenkins J (2019) A closer look at current sensors in EVs. Charged electric vehicles magazine 43. Bosch press release: Bosch teams up with Microsoft to develop software-defined vehicle platform for seamless integration between cars and cloud (2021) [on-line document] [cited

14

44.

45.

46.

47.

48.

49.

50.

51.

52.

L. Metso and A. Happonen 24.2.2021]. https://www.bosch-presse.de/pressportal/de/en/bosch-teams-up-with-microsoftto-develop-software-defined-vehicle-platform-for-seamless-integration-between-cars-andcloud-224832.html Piili H, Widmaier T, Happonen A, Juhanko J, Salminen A, Kuosmanen P, Nyrhilä O (2013) Digital design process and additive manufacturing of a configurable product. Adv Sci Lett 19(3):926–931. https://doi.org/10.1166/asl.2013.4827 Eskelinen T, Räsänen T, Santti U, Happonen A, Kajanus M (2017) Designing a business model for environmental monitoring services using fast MCDS innovation support tools. Technol Innov Manag Rev 7(11):36–46. https://doi.org/10.22215/timreview/1119 Kortelainen H, Happonen A, Hanski J (2019) From asset provider to knowledge company– transformation in the digital era. In: Lecture Notes in Mechanical Engineering. pp 333–341. https://doi.org/10.1007/978-3-319-95711-1_33 Ghoreishi M, Happonen A (2020) Key enablers for deploying artificial intelligence for circular economy embracing sustainable product design: three case studies. AIP Conf Proc 2233(1):1– 19. https://doi.org/10.1063/5.0001339 Keane J (2020) Can self-driving cars lead to sustainability? Autonomous smart sensors, perception and planning algorithms, and data processing efficiency. Contemp Read Law Soc Just 12(1):9–15. https://doi.org/10.22381/CRLSJ12120201 Velasco-Hernandez G, Yeong DJ, Barry J, Walsh J (2020) Autonomous driving architectures, perception and data fusion: a review. In: IEEE 16th International conference on intelligent computer communication and processing (ICCP). pp 315–321. https://doi.org/10.1109/ICC P51029.2020.9266268 Zhao C, Li L, Pei X, Li Z, Wang F-Y, Wu X (2021) A comparative study of state-of-the-art driving strategies for autonomous vehicles. Accid Anal Prevent 150(105937):1–10. https://doi. org/10.1016/j.aap.2020.105937 Happonen A, Santti U, Auvinen H, Räsänen T, Eskelinen T (2020) Digital age business model innovation for sustainability. In: University industry collaboration model, E3S web of conferences, vol 211. pp 1-11. https://doi.org/10.1051/e3sconf/202021104005 Santti U, Eskelinen T, Rajahonka M, Villman R, Happonen A (2017) Effects of business model development projects on organizational culture: a multiple case study of SMEs. Technol Innov Manag Rev 7(8):15–26. https://doi.org/10.22215/timreview/1096

Unsupervised Deep Neural Network Considering the Uncertainties Effect in Pipeline Condition Monitoring Using Guided Ultrasonic Waves Yon Kong Chen , Norhisham Bakhary , Khairul Hazman Padil , and Mohd Fairuz Shamsudin

Abstract Pipeline condition monitoring is vital to ensure the safety of petrochemical pipeline systems. The Guided Ultrasonic Waves (GUW) method has been increasingly applied in pipeline condition monitoring as it provides reliable and accurate pipeline damage information. However, the existence of uncertainties from measurement and the finite element model may lead to unreliable or false damage identification. Attempts to deal with uncertainties, however, are only limited to large damage cases, while for small damage cases, the uncertainties may produce a greater amplitude than the reflected wave, resulting in unidentified reflected waves. Previous attempts employing unsupervised learning to compensate for uncertainties did not retain sufficient damage information for further damage localisation and quantification. Therefore, an unsupervised autoencoder deep neural network is proposed in this study to consider the effect of uncertainties by performing denoising while keeping the important features in the damage signal. The compensated and denoised output of the autoencoder is used to perform damage localisation and damage quantification using the proposed novel damage index based on the Residual Reliability Criterion (RRC). From the results obtained, the proposed method is able to perform uncertainty Y. K. Chen (B) · N. Bakhary · K. H. Padil Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia e-mail: [email protected] N. Bakhary e-mail: [email protected] K. H. Padil e-mail: [email protected] M. F. Shamsudin Faculty of Mechanical Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia e-mail: [email protected] N. Bakhary Institute of Noise and Vibration, Universiti Teknologi Malaysia, City Campus, Jalan Semarak, 54100 Kuala Lumpur, Malaysia © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_2

15

16

Y. K. Chen et al.

compensation and retain sufficient damage information for further damage localisation and damage quantification using the RRC damage index. It has shown consistent outcomes when the level of uncertainties is increased, even for low damage severity cases. Keywords Unsupervised deep neural network · Guided ultrasonic wave · Pipeline condition monitoring · Residual reliability criterion

1 Introduction Safety assurance of the pipeline requires monitoring of the damage occurrence. As the structure experiences continuous loading and vibrations, the material will experience deterioration (i.e. corrosion, cracks, creep) which leads to volume reduction of the structure which will affect the integrity of the structure. Moreover, exposure to extreme operating conditions and an ageing structure may reduce the reliability of the structure thus jeopardising the safety of workers. Therefore, the development of an efficient monitoring system with the ability to detect the occurrence and intensity of damage in the pipelines is vital. One of the popular methods used to monitor the pipeline integrity is to use Guided Ultrasonic Waves (GUW). GUW pipeline condition monitoring has been a popular approach due to the advantages which include; (i) long travelling distances with fast speed without substantial attenuation (ii) capability to examine entire crosssections even for coated and insulated structure inspection [1] (iii) the use of nonlinear ultrasonic waves for the detection of small defects [2]. A GUW is an elastic wave that travels in solid materials and the wave propagation characteristics depend on the boundary conditions of the structure [3]. These elastic waves that travel in the pipeline reflect when the materials have changes in thickness. The changes of material thickness portray the occurrence of damage in the pipeline system. However, the application of GUW faces a drawback caused by the effect of uncertainties. The inevitable uncertainties are from random noises due to measurement errors and modelling errors in the finite element model [4]. These errors jeopardise the accuracy of the pipeline condition monitoring results, especially when the damage severity is small. As the reflected guided wave from the small damage is small in amplitude, the reflected wave could submerge within the error due to uncertainties. A statistical learning-based compensation technique including a supervised, semisupervised and unsupervised neural network performs pattern recognition, clustering, and information extraction to compensate for the uncertainty effect during pipeline condition monitoring [5–7]. Supervised neural networks and semi-supervised neural networks perform pattern recognition to discriminate between damaged and healthy pipelines under variations of operational noises in operating pipelines. For example, [8] applied a supervised method based on a sparse representation of GUW signals which can discriminate between damaged and healthy pipelines under variations of operational noises in operating pipelines [9]. Employed a Convolutional Neural

Unsupervised Deep Neural Network Considering the Uncertainties …

17

Network (CNN) to identify patterns regardless of scale and location which was reported as a better defect classification technique even in a noisy environment. However, the supervised and semi-supervised neural networks require prior analysis of the damaged pipe state as input labels, which are generally not available on site. Moreover, these labels which define various operating, environmental or damage conditions become impracticable, as this information requires manual inspection and a labelling process of each observation which led to an expensive, time consuming and impractical technique [10]. Therefore, unsupervised learning that works with unlabelled data to describe more complex underlying distributions of structural health monitoring data was chosen by several researchers due to the fact that no prior information relating to the damage condition is required to train the model. Unsupervised learning methods such as the k-means algorithm [11] and self-organising map [12] were used for clustering signals into two groups (damaged and healthy) based on the means of the signals. However, this clustering technique was unable to localise damage due to the requirement of a high number of clustering groups where the number of groups is unknown before the identification process. Furthermore, for online damage detection strategies, a big leap of means with obvious differences between signals is required to detect damage, however, damage often occurs gradually and grows constantly without significant signal changes [13]. On the other hand, the decomposition method based on unsupervised learning such as singular value decomposition (SVD) [11], independent component analysis (ICA) [14] and principal component analysis (PCA) [12] were also used to eliminate the uncertainties caused by environmental operating conditions (EOC) while retaining important damage features of the signals. PCA and ICA decompose the signals into multiple components, maximising the statistical independence only between the components, while SVD enforces the orthonormal of both the components and the weight matrices [15]. Several applications of decomposition unsupervised learning have been successfully used for damage detection in pipes operating with different environmental changes [1, 11, 12, 14, 16]. However, the methods only provide the information of damage existence, thus more effort is required to localise and quantify the damage. Due to the reasons above, an unsupervised learning autoencoder is chosen in this study due to its capability in compressing the original input vector that preserves the important features that could be applied and extended to a deeper network architecture from a complex problem [17, 18]. Successfully enhanced the performance of CNN in classification accuracy by adopting an autoencoder for dimensionality reduction and denoising in ultrasonic signals. However, the application of an autoencoder in uncertainty compensation in a GUW pipeline for condition monitoring is quite limited. This study proposes a novel unsupervised learning-based decomposition method autoencoder for compensation of the uncertainty effect. A damage index named the Residual Reliability Criterion (RRC) is proposed for damage localisation and quantification using the damage sensitive features extracted from processed GUW signals. The proposed damage index extracts damage features by discriminating a

18

Y. K. Chen et al.

damaged pipe structure from a healthy pipe structure and requires only a single known condition (undamaged features) to train a relationship learning and decisionmaking model. The efficiency of the proposed method is proven through a numerical model of stainless-steel pipe structure. The results show that the proposed method can retain sufficient damage information for damage localisation and quantification with great confidence despite the effect of uncertainties.

2 Methodology A deep autoencoder model is used to compensate for the uncertainties from the raw GUW signals. Figure 1 shows the steps required to develop the model. Numerical simulation is used to simulate the GUW signals for each damage case. The effect of uncertainties on the GUW signals is assumed to be normally distributed zero means random variables. The GUW signals considering the uncertainty effect are used as an input of the autoencoder to retain important damage features. Training and testing of the deep autoencoder are performed to prevent overfitting and underfitting of the system. The processed GUW signals are then implemented into the proposed RRC damage index to calculate the deviation of the damaged case with a known undamaged case. The deviation calculated using the RRC damage index allows damage localisation and quantification to take place. A straight 8-inch schedule 20 pipes with elements such as excitation point, receiver point and the damage is modelled to portray the actual application of GUW in pipeline damage identification. The numerical modelling technique for torsional guided wave excitation is based on the technique adopted by [19]. The total length of the pipeline is 12 m which comprises 10 m for the damaged region and 2 m for excitation and receiver location setting. In the damaged region, the damage is modelled in different locations and for different severities for data collection. The pipe structure of the numerical model is shown in Fig. 2.

3 Uncertainty Consideration The uncertainties in this study consist of both the effects of noise in the measurement data and uncertainties in the FEM as per a study conducted by [20]. It is assumed to be normally distributed (standard deviation = 1) with zero means random variables. Therefore, the features with uncertainties are equal to true feature values plus the random variations. The GUW features with uncertainties are derived as follows. Di = D + D X i

(1)

Unsupervised Deep Neural Network Considering the Uncertainties …

Fig. 1 Flow chart for proposed pipeline condition monitoring development

Fig. 2 Numerical model of pipe structure using Abaqus FEM software

19

20

Y. K. Chen et al.

where Di is the features of damaged state with i level of uncertainties, i is the level of uncertainties represented by the coefficient of variation of random error in percentage, D is the features of damaged state generated from FEM or measured data and X i is the zero-mean normally distributed random variables with i level of uncertainties.

4 Damage Identification Using RRC Damage Index Damage identification using an unsupervised learning algorithm requires features of a single known condition to learn a decision-making model or relationship. Most of the damage localisation using unsupervised learning adopts comparisons of a measured signal with a known undamaged baseline signal. Deviation calculated from the identified undamaged baseline signal indicates damage occurrence in the pipe structure. This can be achieved by adopting the residual reliability criterion (RRC) damage index. RRC is a residual-based damage index developed based on the theory of the reliability index. The basic principle of RRC is to calculate the relative error between the indices of healthy and damaged conditions [21]. A simple form of RRC can be written as follows. R RC basic =

D−H H

(2)

where D and H denote the features of damaged and healthy states. However, it is possible to encounter negative values based on the calculated RRC because the features might be negative. To tackle this issue, both the numerator and denominator are squared to develop the RRC in the form such as follows. R RC =

D 2 + H 2 − 2D H H2

(3)

The time at peak RRC value indicates the time of the reflected wave caused by defects t2 and will be used to perform damage localisation. The GUW system excites excitation signals equally and simultaneously along the pipe wall. The signals reflect when passing through the area with volume reduction, which indicates a defect. Damage localisation can be performed by identifying the arrival time of the wave caused by reflection due to defects. According to [22], the damage location can be calculated as follows. lx =

t 2 − t1 l t3 − t1

(4)

where l is the length of pipe, l x is the identified distance between sensors and the peak of the reflected wave due to defects. t1 , t2 , t3 indicates the time of the peak excitation wave, reflected wave due to defects and reflected wave due to the end boundary. The

Unsupervised Deep Neural Network Considering the Uncertainties …

21

relative error, r for measuring the qualitative performance of the damage localisation can be defined as follows.    l x − l1   × 100%  (5) r = l1  where l1 is the distance between sensors and actual defects.

5 Deep Autoencoder The autoencoder has served as a signal denoising feature and for representation learning in the signal processing field. An autoencoder allows good reconstruction of its input and retains much of the information by minimising the reconstruction error amounts while maximising a lower bound on the mutual information between input and learnt representation [23]. A traditional autoencoder architecture consists of an encoder layer which transforms an input vector into a hidden representation and a decoder layer which reconstructs the hidden representation back to a vector whose size is the original dimensional [24]. The traditional autoencoder can be improvised to a denoising autoencoder by forcing the model to learn the reconstruction of noisy data by giving it noise-free data as shown in Fig. 3. This can be done by inputting noisy data as the input vector xnoise into the model. Then, the encoder maps the input vector into a hidden representation and the decoder reconstructs the noise-free version of xnoise from the hidden representation. Mean squared error is then calculated between the output of the decoder and noise-free data x for fine-tuning purposes through iteration of the model in the effort of minimising this mean squared error. In this study, the output of the decoder, x is inputted into the RRC Eq. (3) as D features for damage identification. A denoising autoencoder has been successfully applied by [18] in performance enhancement of a deep neural network in flaw classification using the ultrasonic signal. It also outperforms PCA in total training time, producing lower reconstruction errors and is more effective for nonlinear dimensionality reduction [25, 26]. 

Fig. 3 Denoising autoencoder architecture

22

Y. K. Chen et al.

The deep autoencoder in this study consists of 3 hidden layers in encoder layers and 3 hidden layers in decoder layers. The rectified linear unit (ReLU) activation function is applied within the hidden layers of the encoder and decoder. ReLU activation function is a linear function that will directly output the input if the input is positive, while output the input as zero if the input is anything other than positive. This activation function is suitable for developing a multilayer perceptron which overcomes the vanishing gradient problem and allows the neural network to learn faster and perform better. Furthermore, the hyperbolic tangent function (Tanh) is applied at the final layer of the decoder. Tanh function is a nonlinear activation function which is known to have better predictive performance and is easier to train over traditional sigmoid activation function [27]. Fine-tuning is performed using the ‘Adam’ optimizer once the mapping weight coefficients and bias parameters of all hidden layers are obtained to improve the training efficiency and achieve better accuracy of the deep neural network. Sensitivity study of deep autoencoder is performed to identify the suitable architecture of autoencoder that is sensitive to the damage features and uncertainty compensation. Parameters such as different levels of uncertainties, number of hidden layers, number of neurons difference between layers and number of epochs are tested. Root mean square error (RSME) and relative error, r are used as the indicators of the deep autoencoder performance. Lower RSME score indicates smaller difference between the output and the input of the neural network while lower relative error indicates better damage localisation performance of the deep autoencoder. Table 1 shows the outcome of the sensitivity study of the deep autoencoder for this study. Table 1 Sensitivity study outcomes based on different parameters Parameters

Findings

RSME score

relative error, r (%)

Number of hidden layers

3 numbers of hidden layers at both encoder and decoder

2.13e-07

0.12

Number of neuron difference

500 differences between layers: Input layer (3000 nodes), 1st hidden layer (2500 nodes), 2nd hidden layer (2000 nodes), 3rd hidden layer (1500 nodes), bottleneck layer (1000 nodes)

1.72e-07

0.18

Number of epoch

70 number of epoch

2.13e-07

0.12

Level of uncertainties as training sets

20% uncertainty level does not underfit and overfit

1.64e-07

0.12

Unsupervised Deep Neural Network Considering the Uncertainties …

23

6 Damage Identification Using the RRC Damage Index 6.1 Damage Localisation Using the RRC Damage Index Performance of damage localisation using the RRC damage index is conducted by introducing a similar through cut damage at different locations within the damaged region. The intensity of the damage is similar throughout the damage location from the damage at the 1st meter to damage at the 10th meter. Relativity error, r from Eq. (4) is used to evaluate the accuracy of the damage localisation outcome qualitatively. Table 2 shows the damage localisation results of different damage locations by identifying the peak of the RRC damage index. l 2 indicates the position of the actual damage from the edge of pipe as shown in Fig. 2. The RRC max indicates the highest deviation of the damaged GUW signals compared to a healthy GUW signal. From the results, the accuracy of the damage localisation shows inefficiency for damage located near the boundary of the pipe structure, for example, damage localisation within one meter from the end boundary shows 100% error. This is because the reflected wave caused by defects is submerged within the reflected wave due to the boundary. However, in the actual pipeline system, the pipeline would be connected continuously and the reflected wave due to the end boundary would be absent. Table 2 Damage localisation using the RRC index

l2 (m)

RRC max

r (%)

1.00

0.17

100.00

1.50

0.45

87.91

2.00

0.80

49.81

2.50

1.33

20.84

3.00

1.57

0.12

3.50

1.11

7.67

4.00

0.94

7.72

4.50

1.41

1.83

5.00

1.50

2.93

5.50

1.01

2.02

6.00

1.08

7.92

6.50

1.55

3.98

7.00

1.49

0.46

7.50

0.88

5.74

8.00

1.00

3.40

8.50

1.69

6.33

9.00

1.66

0.43

9.50

0.82

4.86

10.00

2.02

10.63

24

Y. K. Chen et al.

Fig. 4 Chart of the RRC damage index of different severity cases

6.2 Damage Quantification Using the RRC Damage Index Damage quantification is performed by comparing damage severity cases of 100%, 75%, 50% and 25% thickness reduction at similar damage locations. Figure 4 shows that the maximum value of the RRC increase when the damage is more severe. This indicates that the maximum RRC value could become an indication of different levels of damage severity. A similar trend can be observed in the study done by [19], as the amplitude of the reflected damage wave increased when the damage severity increased.

6.3 Effect of Uncertainties on Damage Identification Using the RRC Damage Index The uncertainty effect on damage localisation and damage quantification based on the RRC damage index strategy is conducted. Cases of different levels of severity are applied with different levels of uncertainty. Due to the randomness of uncertainty, the results calculated are the mean of 200 signals for each case. Based on the results in Table 3, when the level of uncertainty increases, the maximum RRC value increases while the accuracy of damage localisation decreases at all levels of damage severity. For example, the maximum RRC value for 100% damage severity increases from 1.577 to 12.493 as the uncertainty level increases from 0 to 50%. It can be observed that damage localisation is still possible with the presence of the uncertainty effect. However, the relative error reaches almost 30% for 25% thickness reduction cases when the uncertainty level reaches 45%. False damage identification occurs more often for small damage cases with a high level of uncertainty. This is because the reflected wave due to a small damage case is small in amplitude and often submerged within the effect of uncertainties.

Unsupervised Deep Neural Network Considering the Uncertainties …

25

Table 3 Damage identification with different level of severities and uncertainties, values in brackets indicate standard deviation of the results Damage severity, percentage of thickness reduction at damage (%) Level of uncertainties (%)

100%

75%

50%

25%

RRC max

r (%)

RRC max

r (%)

RRC max

r (%)

RRC max

r (%)

0

1.577 (0.00)

0.116 (0.00)

0.797 (0.00)

2.239 (0.00)

0.661 (0.00)

2.074 (0.00)

0.278 (0.00)

18.427 (0.00)

5

1.980 (0.18)

1.624 (4.55)

1.105 (0.10)

8.840 (9.07)

0.899 (0.10)

2.285 (1.17)

0.440 (0.04)

19.810 (13.23)

10

2.650 (0.35)

2.710 (6.70)

1.567 (0.21)

10.552 (9.89)

1.224 (0.22)

3.001 (3.32)

0.692 (0.13)

19.774 (14.78)

15

3.522 (0.61)

5.295 (7.38)

2.176 (0.39)

13.900 (10.58)

1.605 (0.33)

4.814 (4.85)

1.043 (0.18)

21.588 (15.27)

20

4.338 (0.81)

6.369 (8.83)

2.656 (0.58)

13.504 (11.10)

2.058 (0.51)

5.055 (5.35)

1.364 (0.28)

21.993 (17.12)

25

5.356 (1.16)

8.985 (10.36)

3.591 (0.80)

15.259 (14.37)

2.756 (0.72)

5.191 (7.19)

1.922 (0.46)

27.570 (17.16)

30

6.698 (1.68)

7.997 (9.21)

4.412 (1.13)

16.147 (13.30)

3.315 (1.05)

8.067 (7.06)

2.356 (0.53)

22.311 (18.11)

35

7.929 (2.07)

8.090 (10.44)

5.434 (1.25)

16.723 (14.26)

4.192 (1.26)

7.392 (10.10)

3.061 (0.73)

25.639 (18.42)

40

9.179 (2.75)

9.776 (11.57)

6.423 (1.80)

18.190 (14.64)

4.886 (1.55)

7.393 (11.01)

3.691 (1.13)

27.531 (17.86)

45

10.731 (3.36)

10.205 (10.32)

7.434 (1.98)

16.083 (14.63)

6.086 (1.84)

8.245 (9.13)

4.566 (1.42)

29.186 (18.73)

50

12.493 (3.48)

10.196 (11.80)

8.815 (2.14)

16.320 (15.93)

6.525 (2.22)

8.727 (12.47)

5.342 (1.42)

26.385 (19.43)

On the other hand, damage quantification using the maximum RRC damage index becomes irrelevant when the uncertainty effect is present. Presence of uncertainties increases the amplitudes of the RRC index features. As it is impossible to control the level of uncertainties during the field measurement, the classification of damage severity becomes irrelevant for GUW signals with the presence of uncertainties.

7 Uncertainty Compensation Using a Deep Autoencoder Figure 5 shows the outcome of the uncertainty compensation using deep autoencoder for 100% damage severities. The plot shows RRC damage features of (a) 0% uncertainty, (b) 50% uncertainties without uncertainty compensation and (c) 50% uncertainties after uncertainty compensation. Based on the outcome, minimum

26

Y. K. Chen et al.

Fig. 5 Plot of RRC damage features with a 0% uncertainties, b 50% uncertainties without uncertainty compensation and c 50% uncertainties after uncertainty compensation

effect of uncertainty is observed when deep autoencoder is applied even for high level of uncertainties. The number of training data for uncertainty compensation using deep autoencoder is 200 signals. Table 4 shows the outcome of the damage identification by using the output of 200 testing data from the deep autoencoder. Damage localisation using the output from the deep autoencoder has shown a consistent result when the level of uncertainties increases. On the other hand, the maximum RRC value shows a slight increase when the level of uncertainties increases. This indicates that the output from the deep autoencoder is able to compensate for the uncertainty effect and make damage localisation and quantification eligible. In short, a deep autoencoder with suitable training sets and architecture is capable of performing relationship learning and generating features with sufficient damage information for damage localisation and quantification using the RRC damage index strategy.

Unsupervised Deep Neural Network Considering the Uncertainties …

27

Table 4 Damage identification using output from deep autoencoder Damage severity, percentage of thickness reduction at damage (%) Level of 100% uncertainties RRC max r (%) (%)

75%

50%

25%

RRC max r (%)

RRC max r (%)

RRC max r (%)

0

1.609

0.116 0.816

2.239 0.684

2.273 0.292

17.897

5

1.609

0.116 0.815

2.239 0.684

2.273 0.292

17.897

10

1.610

0.116 0.815

2.239 0.684

2.273 0.292

17.897

15

1.609

0.116 0.815

2.239 0.685

2.273 0.292

17.897

20

1.611

0.115 0.816

2.239 0.685

2.264 0.293

17.940

25

1.613

0.114 0.815

2.243 0.685

2.264 0.293

17.930

30

1.614

0.114 0.815

2.243 0.686

2.255 0.292

18.040

35

1.614

0.114 0.815

2.234 0.685

2.236 0.293

17.971

40

1.610

0.117 0.815

2.222 0.686

2.235 0.293

18.064

45

1.612

0.118 0.817

2.224 0.687

2.239 0.295

18.096

50

1.613

0.109 0.817

2.255 0.687

2.228 0.296

18.087

8 Conclusions Damage location and severity can be detected using features of the RRC damage index which requires only one healthy GUW signal case as the database. Damage localisation can be performed by finding the peak of RRC features. On the other hand, damage quantification can be performed by identifying the RRC damage index value. However, during field measurement, the effect of inevitable uncertainty affects the accuracy of damage localisation and makes damage quantification impossible. Uncertainties have shown to increase the amplitude of the maximum RRC damage index features and lead to false damage localisation for small severity cases with a high level of uncertainties. The deep autoencoder is trained with GUW signals with uncertainty effect by providing GUW signals without the effect of uncertainties. The output from the deep autoencoder is adopted as the features for damage identification using the RRC damage index. The sensitivity study on the deep autoencoder architecture has shown greater performance when a higher number of hidden layers is used. The results show that the output from the deep autoencoder can retain sufficient damage features and prove its capability to improve the damage localisation using the RRC damage index strategy, and make damage quantification possible with great confidence despite the effect of uncertainties. This approach can be implemented in online pipeline condition monitoring along with permanently installed GUW sensors. The clean GUW signal without uncertainty effect can be obtained by numerical simulation or before pipeline operation. Then, the deep autoencoder is trained to perform uncertainty compensation by forcing the model to learn the reconstruction of noisy data (field measurements) by giving its

28

Y. K. Chen et al.

noise-free data (numerical generated). The results of this study using numerical simulation data show good performance of deep autoencoder in performing uncertainty compensation, further study using field measurement data is required to prove the suitability of the proposed approach in real world application. Acknowledgements The authors would like to thank the Ministry of Higher Education, Malaysia, and Universiti Teknologi Malaysia (UTM) for their financial support through the HICoE Grant (4J224) and Transdisciplinary Research Grant (07G46).

References 1. Eybpoosh M, Berges M, Noh HY (2017) An energy-based sparse representation of ultrasonic guided-waves for online damage detection of pipelines under varying environmental and operational conditions. Mech Syst Signal Process 82:260–278. https://doi.org/10.1016/j.ymssp. 2016.05.022 2. Marcantonio V, Monarca D, Colantoni A, Cecchini M (2019) Ultrasonic waves for materials evaluation in fatigue, thermal and corrosion damage: A review. Mech Syst Signal Process 120:32–42. https://doi.org/10.1016/j.ymssp.2018.10.012 3. Yeung C, Ng CT (2019) Time-domain spectral finite element method for analysis of torsional guided waves scattering and mode conversion by cracks in pipes. Mech Syst Signal Process 128:305–317. https://doi.org/10.1016/j.ymssp.2019.04.013 4. Abdulkareem M, Bakhary N, Vafaei M et al (2018) Non-probabilistic wavelet method to consider uncertainties in structural damage detection. J Sound Vib 433:77–98. https://doi. org/10.1016/j.jsv.2018.07.011 5. Ahn B, Kim J, Choi B (2019) Artificial intelligence-based machine learning considering flow and temperature of the pipeline for leak early detection using acoustic emission. Eng Fract Mech 210:381–392. https://doi.org/10.1016/j.engfracmech.2018.03.010 6. Hoang ND, Tran VD (2019) Image Processing-Based detection of pipe corrosion using texture analysis and metaheuristic-optimized machine learning approach. Comput Intell Neurosci. https://doi.org/10.1155/2019/8097213 7. Liu H, Liu Z, Taylor B, Dong H (2019) Matching pipeline In-line inspection data for corrosion characterization. NDT E Int 101:44–52. https://doi.org/10.1016/j.ndteint.2018.10.004 8. Eybpoosh M, Berges M (2015) Noh HY (2015) Nonlinear feature extraction methods for removing temperature effects in multi-mode guided-waves in pipes. Struct Heal Monit Insp Adv Mater Aerospace, Civ Infrastruct 9437:94371W. https://doi.org/10.1117/12.2084436 9. Modarres C, Astorga N, Droguett EL, Meruane V (2018) Convolutional neural networks for automated damage recognition and damage type identification. Struct Control Heal Monit 25:1–17. https://doi.org/10.1002/stc.2230 10. Bull LA, Worden K, Dervilis N (2020) Towards semi-supervised and probabilistic classification in structural health monitoring. Mech Syst Signal Process 140:106653. https://doi.org/10.1016/ j.ymssp.2020.106653 11. Liu C, Harley JB, Bergés M et al (2015) Robust ultrasonic damage detection under complex environmental conditions using singular value decomposition. Ultrasonics 58:75–86. https:// doi.org/10.1016/j.ultras.2014.12.005 12. Torres-Arredondo MA, Sierra-Pérez J, Cabanes G (2016) An optimal baseline selection methodology for data-driven damage detection and temperature compensation in acoustoultrasonics. Smart Mater Struct 25:1–15. https://doi.org/10.1088/0964-1726/25/5/055034 13. El MM, Mourot G, Yaacoubi S, Maquin D (2018) Damage detection and localization in pipeline using sparse estimation of ultrasonic guided waves signals. IFAC-PapersOnLine 51:941–948. https://doi.org/10.1016/j.ifacol.2018.09.688

Unsupervised Deep Neural Network Considering the Uncertainties …

29

14. Dobson J, Cawley P (2016) Independent component analysis for improved defect detection in guided wave monitoring. Proc IEEE 104:1620–1631. https://doi.org/10.1109/JPROC.2015. 2451218 15. Liu C, Dobson J, Cawley P (2017) Efficient generation of receiver operating characteristics for the evaluation of damage detection in practical structural health monitoring applications. Proc R Soc A Math Phys Eng Sci 473. https://doi.org/10.1098/rspa.2016.0736 16. El Mountassir M, Yaacoubi S, Dahmene F (2020) Reducing false alarms in guided waves structural health monitoring of pipelines: Review synthesis and debate. Int J Press Vessel Pip 188:104210. https://doi.org/10.1016/j.ijpvp.2020.104210 17. Pathirage CSN, Li J, Li L et al (2018) Structural damage identification based on autoencoder neural networks and deep learning. Eng Struct 172:13–28. https://doi.org/10.1016/j.engstruct. 2018.05.109 18. Munir N, Park J, Kim HJ et al (2020) Performance enhancement of convolutional neural network for ultrasonic flaw classification by adopting autoencoder. NDT E Int 111:102218. https://doi.org/10.1016/j.ndteint.2020.102218 19. Niu X, Duan W, Chen H-P, Marques HR (2019) Excitation and propagation of torsional T(0,1) mode for guided wave testing of pipeline integrity. Measurement 131:341–348. https://doi.org/ 10.1016/j.measurement.2018.08.021 20. Bakhary N, Hao H, Deeks AJ (2007) Damage detection using artificial neural network with consideration of uncertainties. Eng Struct 29:2806–2815. https://doi.org/10.1016/j.engstruct. 2007.01.013 21. Entezami A, Shariatmadar H (2017). An unsupervised learning approach by novel damage indices in structural health monitoring for damage localization and quantification. https://doi. org/10.1177/1475921717693572 22. Wu J, Wang Y, Zhang W et al (2017) Defect detection of pipes using Lyapunov dimension of Duffing oscillator based on ultrasonic guided waves. Mech Syst Signal Process 82:130–147. https://doi.org/10.1016/j.ymssp.2016.05.012 23. Vincent P, Larochelle H, Lajoie I et al (2010) Stacked denoising autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion. J Mach Learn Res 11:3371–3408 24. Nan Jiang, Wenge Rong, Baolin Peng, et al (2015) An empirical analysis of different sparse penalties for autoencoder in unsupervised feature learning. In: 2015 International Joint Conference on Neural Networks (IJCNN). IEEE, pp 1–8 25. Hinton GE (2006) Reducing the dimensionality of data with neural networks. Science (80- ) 313:504–507. https://doi.org/10.1126/science.1127647 26. Ozdagli AI, Koutsoukos X (2019) Machine learning based novelty detection using modal analysis. Comput Civ Infrastruct Eng 1–22. https://doi.org/10.1111/mice.12511 27. LeCun YA, Bottou L, Orr GB, Müller K-R (2012) Efficient BackProp. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). pp 9–48

Detecting Gas Injection Problems in Vacuum Tank Degassing Using Measurements of Multiple Variables Juhani Nissilä , Mika Pylvänäinen , Jouni Laurila , Seppo Ollila , Ville-Valtteri Visuri , and Toni Liedes

Abstract By monitoring the simultaneous time series measurements of ladle vibration, vacuum tank pressure and argon gas flow rates of two gas channels, we can detect multiple different fault states in the vacuum tank degassing process. These problems are typically caused by leaking or blockages in the gas lines or in the two porous plugs at the bottom of each ladle. We group the stirrings into the following categories: normal operation, problem in one of the gas lines, poor stirring efficiency and operator actions related to it and also temporary increase in vacuum tank pressure. Keywords Ladle metallurgy · Vacuum tank degassing · Vibration monitoring

J. Nissilä (B) · M. Pylvänäinen · J. Laurila · T. Liedes Intelligent Machines and Systems, University of Oulu, PO Box 4200, FI-90014 Oulu, Finland e-mail: [email protected] M. Pylvänäinen e-mail: [email protected] J. Laurila e-mail: [email protected] T. Liedes e-mail: [email protected] S. Ollila Technical Development, SSAB Europe Oy, PO Box 93, 92101 Raahe, Finland e-mail: [email protected] V.-V. Visuri Process Metallurgy Research Unit, University of Oulu, PO Box 4300, FI-90014 Oulu, Finland e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_3

31

32

J. Nissilä et al.

1 Introduction In secondary metallurgy of steelmaking, vacuum tank degassing (VTD) is a process where steel is refined under vacuum conditions using argon-blowing. The goal is to achieve a homogeneous composition and high-purity steel. In previous research work [1], a vibration monitoring system was developed for SSAB Europe, Raahe, Finland. It was observed that especially smoothed root mean square (RMS) values of vertical velocity measurements from the support beam of the vacuum tank correlate well with the argon gas flow rate when the process is functioning normally. Similar result was also found for example in [2], where a linear correlation between the mean of a filtered vibration index and argon gas flow rate was observed. They also studied the use of back pressure measurements from the gas lines and found that it also had a linear correlation with the argon gas flow rate when the gas flow operated normally. Vibrations measured from the ladle external wall were also studied in [3], where filtered acceleration RMS values were used as a vibration index. The theoretical justification for the linear relation between the gas flow rate and vibration can be based on the stirring power equation (found for example in [4, p 273] or in [5, p 252] without the second term).     pin 1 V˙G RT ln + R(T − Tin ) , ε˙ = m Vm pout

(1)

where ε˙ is the stirring power [W kg−1 ], V˙G is the volumetric gas flow rate [Nm3 s−1 ] (that is normal cubic meters per second), V m is the molar volume [m3 mol−1 ], m is the mass of the melt [kg], R is the universal gas constant [J mol−1 K−1 ], T is the melt temperature [K], T in is the inlet temperature [K], pin is the inlet pressure [Pa], and pout is the outlet pressure [Pa]. Equation (1) suggests that the stirring power and the volumetric gas flow rate have a linear relationship. In [6] it was concluded that stirring power and bath circulation speed had a strong linear relationship with ladle vibration. Vibrations measured from ladle water models [6, 7] have also been studied. The results found using the water model were also verified on a 160-tonne VTD [8], though they then found that although there is a linear correlation between the vibration and flow rate during any single stirring, the correlation is not visible if the data from several stirrings are grouped together, i.e., the slope of the fit is different between stirrings. This same phenomenon was also observed in [1] and was the reason that the correlations were studied from each stirring separately. It was also observed in [2], and thus was the reason why the linear correlation was observed only between the mean vibration levels and flow rate. The aim of this research is to study how pressure, gas flow rate and vibration measurements can reveal those stirrings where there have been problems in the gas injection. These problems are typically leaking or blockages in the gas lines or in the two porous plugs at the bottom of each ladle.

Detecting Gas Injection Problems in Vacuum Tank Degassing Using …

33

2 Materials and Methods Vertical and horizontal vibration was recorded from the support beam of the vacuum tank with two accelerometers. The gas flow rate information was measured by gas flow controllers which output the actual flow rate as a 4—20 mA current signal. Two accelerometers were also installed on the gas lines. Sampling frequency for data collection was 25.6 kHz. In this study, we utilize the vertical vibration measurement from the support beam together with the gas flow rate measurements. In addition, the vacuum tank pressure was collected from the automation system. The measurement system is described in more detail in [1]. Since we will look at some extreme cases from a group of hundreds of stirrings, it is helpful to have a sense of the variation that is present in this group. In [1] a ladle responsiveness value (LRV) was defined as the ratio of the cumulated vibration and the cumulated argon gas during a stirring’s vacuum phase, that is LRV = Cumulated Vibration/Cumulated Argon Gas Volume.

(2)

The idea behind the LRV is simple: the less argon gas one needs to use to generate sufficient stirring and thus vibration, the more successful the stirring was and LRV is big in this case. A more detailed formula (4) for the LRV is presented later in Sect. 3.5. The distribution of the LRVs calculated from a total of 273 stirrings and grouped by their ladles is shown in Fig. 1. This Fig. was originally published in [1] and is licensed under Creative Commons Attribution 4.0 International License (CC BY 4.0). Some of the outlier stirrings in Fig. 1 will appear later as we analyze the simultaneous time series measurements from different fault cases. In all of the figures that follow which contain simultaneous measurements, the units for one grid spacing (gray dashed lines) in the Y-direction is 200 Nl/min (normal litre/min) for gas flow rates, 0.39 kPa for absolute pressure and 0.23 mm/s for the vibration indicator. Outside of the actual vacuum period of the stirring we will always see that the pressure sensor saturates to a constant value.

3 Results and Discussion 3.1 Normal Operation Figure 2 shows three stirrings where the gas flow seems to function adequately. Blue curve describes the vacuum tank pressure (1 Hz sampling). Red and black are the gas flow rates of the two gas flow lines (1 s averages) and they should follow the same set value in normal operation. Gray is the vibration indicator, which is calculated from the vertical accelerometer measurements. The vibration data is first numerically integrated to velocity, then RMS values from 1 s intervals are calculated

34

J. Nissilä et al. 64

204 162

228

203

103

240

134

L7

L9

L17

L2

L3

L6

L5

L18

L10

L4

L11

L8

L14 L16

L12

L1

L20

Ladle ID

Fig. 1 Summary of the LRV by the Ladle ID. Outliers with ID number of stirrings depict an exceptionally high or low LRV compared with the expected LRV range of that specific ladle. This Fig. was originally published in [1] and is licensed under Creative Commons Attribution 4.0 International License (CC BY 4.0)

and finally the time series is further smoothed with a centered median filter using 21 s window length. In Fig. 2, we can see three stirrings where the gas flow rate is kept at a constant value during the vacuum. We also see that the vibration indicator is relatively constant during the vacuum. These cases still represent the fact that the linear correlation between the gas flow rate and the vibration indicator is different between stirrings. Thus, the first case is an example of a typical vibration stirring with a normal LRV value, the second has a lower LRV value and the third has a very large LRV value. The stirring number 64 in Fig. 2 is actually also present in Fig. 1 as the exceptional outlier for ladle ID 17. Figure 3 shows three stirrings where the gas flow rate has been changed in a step like manner during the vacuum phase. The linear correlation between the vibration indicator and gas flow rate is very clearly seen here, a fact that was demonstrated numerically in [1].

3.2 Problem in One of the Gas Lines Figure 4 shows three cases where the gas flow rate of the 2nd gas line (red curve) lags behind the set value for some unknown reason. In stirring no. 165, the gas flow rate of the 2nd gas line remains very low for the entire duration of the vacuum phase. This in turn means that the operator has turned up the set value of the gas flow rate

Detecting Gas Injection Problems in Vacuum Tank Degassing Using … MeltID: 26

0

400

800

Time (s)

MeltID: 64

MeltID: 54

1200

1600

0

400

800

Time (s)

1200

35

1600

0

400

800

Time (s)

1200

Fig. 2 Three stirrings where the gas flow rate is kept at a constant value for the whole duration of the vacuum and the linear correlation between the vibration indicator and gas flow rate is good, though note that the slope for this linear relation changes between stirrings

MeltID: 143

0

400

800

Time (s)

1200

MeltID: 144

0

400

800

Time (s)

1200

MeltID: 255

1600

0

400

800

Time (s)

1200

1600

Fig. 3 Three stirrings where the gas flow rate changes in a step like manner during the vacuum, but the linear correlation between the vibration indicator and gas flow rate is still good in each stirring

to a high value so that the 1st gas line (black curve) can compensate in generating sufficient stirring. In stirring no. 281, the gas flow rate of the 2nd gas line lags behind for approximately 600 s, but slowly increases to the set value after this time period. Since the gas flow rate of the 1st gas line is constant during the vacuum phase of this stirring, it is clear that the vibration indicator mostly follows the shape of the 2nd gas flow rate. Figure 5 shows two cases where the 2nd gas line is also not keeping up with the set value, but the deviations are smaller in value and last for a shorter time period than in Fig. 4. In the third case, the 2nd gas line drops at the start of the vacuum phase and then overshoots the set value near the end before returning to the set value. The vibration indicator reacts to this overshoot.

36

J. Nissilä et al. MeltID: 281

MeltID: 165

0

0

400 800 1200 1600 2000

Time (s)

400

800

MeltID: 282

1200

Time (s)

0

400

800

1200

Time (s)

1600

Fig. 4 Three stirrings where the 2nd gas line’s flow rate lags behind the set value

MeltID: 97

MeltID: 30

0

400

800

1200

Time (s)

1600

0

400

800

Time (s)

1200

MeltID: 91

1600

0

400

800

Time (s)

1200

1600

Fig. 5 Three stirrings where the 2nd gas line is also not following the set value for some period

3.3 Poor Stirring Efficiency and Operator Actions Related to It In Fig. 6 there are examples of three stirrings where exceptionally high gas flow rates are needed to stir the steel. This could indicate that not all of the gas actually reaches the ladle. This could mean for example that the porous plugs at the bottom of the ladle are partly blocked and/or that the automatic stirring gas coupling between the ladle and the support structure is leaking. In the second stirring, the vibration indicator also slowly increases, which indicates that a larger share of the gas reaches the ladle towards the end of the vacuum period. The stirring number 134 is also present in Fig. 1, as the outlier for ladle ID 17 with a very low LRV value. The three stirrings in Fig. 7 are also examples of cases where large amounts of gas are used. The difference is that now the operator also tries to get the gas to flow into the ladle by ramping the gas flow up. This succeeds in the first stirring number 152

Detecting Gas Injection Problems in Vacuum Tank Degassing Using … MeltID: 48

0

400

800

Time (s)

37

MeltID: 134

1200

0

400

800

Time (s)

1200

MeltID: 117

1600

0

400

800

Time (s)

1200

1600

Fig. 6 Three stirrings where an exceptionally high gas flow rate is needed to stir the steel

MeltID: 152

0

400

800

Time (s)

1200

MeltID: 283

MeltID: 279

0

400

800

1200 1600 2000

Time (s)

0

400

800

1200 1600

Time (s)

Fig. 7 Three stirrings where the operator tries to open the possibly blocked gas lines by increasing the gas flow rate rapidly. This is successful in stirring 152 (left)

(a similar success also happens for stirring 282 in Fig. 4) but fails for 279 and 283. The operator then continues to try to open the gas lines by constantly alternating the gas flow between the maximum and minimum values.

3.4 Temporary Increase in Vacuum Tank Pressure Vacuum treatment facilitates degassing, decarburization and deoxidation of steel [9, pp 239–252]. Moreover, the efficiency of desulphurization is related to the stirring power but also to the vacuum pressure [10]. Thus, it is important to maintain a good vacuum for the duration of the stirring. This is not always successful in practice. Figure 8 shows three examples where the vacuum tank pressure suddenly increases during the first half of the stirring. In the first two stirrings this happens twice, and

38

J. Nissilä et al.

some of these events last for several minutes. In the third stirring, the vacuum is also lost for several minutes during the first half of the stirring. The timing of these events suggests that they could be related to material additions. It is clear that this loss of vacuum is not visible in the vibration indicator at all and thus proves the need for measuring several different variables when monitoring vacuum tank degassing efficiently. Figure 9 shows three cases where the vacuum is only partially lost, and these events are also shorter in duration than the ones in Fig. 8. All of these events also happen during the first half of the stirring. Note that there are similar temporary vacuum losses also in stirrings 279 and 283 shown in Fig. 7. Clearly, multiple different fault states sometimes occur simultaneously. MeltID: 276

0

400

800

MeltID: 278

1200 1600

Time (s)

0

400

800

MeltID: 280

1200 1600

Time (s)

0

400

800

1200

Time (s)

1600

Fig. 8 Three stirrings where the vacuum disappears for several minutes

MeltID: 275

0

400

800

Time (s)

1200

MeltID: 274

MeltID: 277

1600

0

400

800

Time (s)

1200

1600

0

400

800

Time (s)

Fig. 9 Three stirrings where the vacuum almost disappears for a short period of time

1200

Detecting Gas Injection Problems in Vacuum Tank Degassing Using …

39

3.5 Features for Automated Fault Classification In previous sections, we have demonstrated that it is possible to identify several fault states in vacuum tank degassing by inspecting the simultaneously recorded measurements of multiple variables. This however requires expert knowledge and careful inspection of the recorded time series. For these reasons, we will now take steps towards automating this process by defining simple features which will separate the different fault states. First, we define the notation for weighted lp norms  x

(α)

p =

N 1   (α)  p x N n=1 n

1/ p ,

(3)

where α is the possible order of derivative, p is the order of norm and N is the length of calculation. Here N = 25600 samples, since we calculated the norms from 1 s intervals. When p = 1, we get absolute average values and when p = 2, we get RMS values. In the case of vibration measurements, x represents displacement, and thus x (1) denotes the velocity signal that has been numerically integrated from the acceleration x (2) measurements. RMS values calculated from some 1 s norms for the whole vacuum period of a stirring are denoted by RMS(·). The four features that we define are single numbers for each stirring calculated over the vacuum period. The first feature is the LRV of Eq. (2). If we denote the median filtered 1 s RMS values of vertical velocity measurements by med21s || x (1) ||2 and the 1 s averages of the two gas lines by ||g1 ||1 and ||g2 ||1 , then the LRV is explicitly

med21s x (1) 2 dt . LRV =  g1 1 + g2 1 dt

(4)

and we approximate the integral over the vacuum period by a sum. The second feature is the RMS of the difference between the two gas flow rate values over the vacuum period.  RMS g1 1 − g2 1 .

(5)

Next, we fit a linear regression model that models the vibration in terms of total gas flow rate, i.e., a least squares line that estimates med21s ||x (1) ||2 in terms of ||g1 ||1 + ||g2 ||1 . Denote the regression prediction by pred||x (1) ||2 and the third feature is then the RMS of the residual of the regression estimate fit over the vacuum period, i.e., RMSE (Root Mean Square Error)  RMS predx (1) 2 − med21s x (1) 2 .

(6)

Finally, the fourth feature is the RMS of pressure measurements from 1 s intervals P over the vacuum period

40

J. Nissilä et al.

RMS (P).

(7)

Figure 10 shows the features (4) and (5) calculated from all of the 24 stirrings which were analyzed in previous sections. The symbols and colors are based on the initial groupings to which they were labeled to. Since several stirrings actually had many faults simultaneously, it is no surprise that the groups that form in Fig. 10 are not only based on the initial labels. For this reason, we have also highlighted some common properties by dashed rectangles. When feature (5) is large (at the top part of the figure), there are problems in one of the gas lines. This group includes all the six stirrings from Sect. 3.2, but this fault is also present in stirrings 279 and 152 of Fig. 7. The stirrings in the left part of Fig. 10 have very low LRV values and thus poor stirring efficiency. This includes all the stirrings from Sect. 3.3, but also stirrings 282 and 91 from Figs. 4 and 5, stirring 277 from Fig. 9 and stirring 54 from Fig. 2. Stirring 64 from Fig. 2 is the most efficient stirring in this dataset and thus occupies a spot at the right part of Fig. 10 by itself. The remaining 4 stirrings from Sect. 3.1 and 5 stirrings from Sect. 3.4 occupy the “normal LRV and gas line functionality” group at the lower center part of Fig. 10. Figure 11 shows the features (6) and (7) calculated from all the 24 stirrings which were analyzed in the previous sections. When feature (7) is large (at the top part of the figure), temporary vacuum pressure increase occurs during the stirring. This group includes all the six stirrings from Sect. 3.4, but this fault is also present in stirrings 279 and 283 of Fig. 7. Stirring 152 from Fig. 7 and stirring 282 from Fig. 4

Fig. 10 Features (4) and (5), i.e., LRV and RMS of the difference of the two gas flow rate values over the vacuum period for all the 24 stirrings analyzed in Sects. 3.1–3.4

Detecting Gas Injection Problems in Vacuum Tank Degassing Using …

41

Fig. 11 Features (6) and (7), i.e., RMSE of a linear model of vibration velocity in terms of total gas flow rate and RMS of vacuum pressure over the vacuum period for all the 24 stirrings analyzed in Sects. 3.1–3.4

are the only ones on the right side of Fig. 11, which means that the linear model between vibration and gas flow rate does not fit them well. Observing their time series, we can see that they both had very small vibration values for the first half of the stirring and hence a probable blockage. This then opened during the second half of the stirring. Thus, there was a big and quite fast change from a blocked state to a functional gas stirring state which explains why the linear model is bad for them. This is a good example of a grouping which was revealed by the feature (7) and was not in the original labeling. Stirrings 283, 134 and 30 from Figs. 7, 6 and 5 have been given the label “bad linear model of vibration”. Observing their time series, we see that they have periods of relatively constant gas flow rate, but large variation in the vibration values. Rest of the stirrings occupy a space at the lower left corner of Fig. 11 and can be labeled collectively as having “normal vacuum and linear model of vibration”. Of these 12 stirrings, we can see that 5 at the farthest left in the figure are from Sect. 3.1, i.e., stirrings that were originally labeled as having a very normal operation.

42

J. Nissilä et al.

4 Conclusions Ladle vibration measurements from a ladle support beam of a vacuum tank degasser at SSAB Europe Raahe were combined with vacuum tank pressure and argon gas flow rate measurements of two gas channels in an effort to recognize different fault states in the vacuum tank degassing process. The stirrings were grouped into the following categories: normal operation, problem in one of the gas lines, poor stirring efficiency and operator actions related to it and also temporary increase in vacuum tank pressure. Four numerical features were also tested on 24 labeled stirrings and found to successfully separate different fault states and reveal their combinations. These features were the ladle responsiveness value (LRV), RMS of the difference between the two gas flow rate values, RMSE of the linear regression model between vertical velocity and total gas flow rate and RMS of pressure measurements, all of these calculated over the vacuum period of the stirring. It is noteworthy, that all these features can also be calculated online during the stirring. In this study, we only plotted the final values that the features had at the end of the vacuum period. The corresponding online calculations can be utilized for monitoring and controlling the vacuum tank degassing process.

References 1. Pylvänäinen M, Visuri V-V, Nissilä J, Laurila J, Karioja K, Ollila S, Fabritius T, Liedes T (2020) Vibration-Based Monitoring of Gas-Stirring Intensity in Vacuum Tank Degassing. Steel Res Int 91(6):1900587. https://doi.org/10.1002/srin.201900587 2. Hwang H, Lee YM, O’loughlin A, Strobel E (2014) Back pressure correlation—a method to determine gas flow status at LMF. In: Proceeding of the Iron & Steel Technology Conference AISTech 2014, Volume II, Indianapolis, Indiana, USA, 5–8 May 3. Burty M, Pussé C, Bertoletti C, Wetta P, Cariola E (2006) Kettlor: efficient stirring in ladle metallurgy. Rev Met Paris 103(11):493–499. https://doi.org/10.1051/metal:2006128 4. Engh TA (1992) Principles of Metal Refining. Oxford University Press, New York, NY, USA 5. Oeters F (1989) Metallurgie der Stahlherstellung. Verlag Stahleisen, Düsseldorf, Germany 6. Yenus J, Brooks G, Dunn M (2016) Multivariate analysis of ladle vibration. Metall Mater Trans B 47:2681–2689. https://doi.org/10.1007/s11663-016-0707-9 7. Alia N, Pylvänäinen M, Visuri V-V, John V, Ollila S (2019) Vibrations of a laboratory-scale gasstirred ladle with two eccentric nozzles and multiple sensors. J Iron Steel Res Int 26:1031–1040. https://doi.org/10.1007/s42243-019-00241-x 8. Yenus J, Brooks G, Dunn M, Li Z, Goodwin T (2018) Study of Low Flow Rate Ladle Bottom Gas Stirring Using Triaxial Vibration Signals. Metall Mater Trans B 49:423–433. https://doi. org/10.1007/s11663-017-1118-2 9. Deo B, Boom R (1993) Fundamentals of steelmaking metallurgy, prentice hall international, Hertfordshire, UK 10. Lachmund H, Xie Y, Harste K (2001) Thermodynamic and kinetic aspects of the desulphurisation reaction in secondary metallurgy. Steel Res 72(11+12):452–459. https://doi.org/10.1002/ srin.200100151

Feature Assessment for a Hybrid Model Antonio Gálvez , Dammika Seneviratne , Diego Galar , and Esko Juuso

Abstract This paper proposes an assessment of features orientated to improve the accuracy of a hybrid model (HyM) used for detecting faults in a heating, ventilation, and air conditioning (HVAC) system. The HyM combines data collected by sensors embedded in the system with data generated by a physics-based model of the HVAC. The physics-based model includes sensors embedded in the real system and virtual sensors to represent the behaviour of the system when a failure mode (FM) is simulated. This fusion leads to improved maintenance actions to reduce the number of failures and predict the behaviour of the system. HyM can lead to improved fault detection and diagnostics (FDD) processes of critical systems, but multiple fault detection models are sometimes inaccurate. The paper assesses features extracted from synthetic signals. The results of the assessment are used to improve the accuracy of a multiple fault detection model developed in previous research. The assessment of features comprises the following: (1) generation of run-to-failure data using the physics-based model of the HVAC system; the FMs simulated in this paper are dust in the air filter, degradation of the CO2 sensor, degradation of the evaporator fan, and variations in the compression rate of the cooling system; (2) identification A. Gálvez (B) · D. Seneviratne · D. Galar TECNALIA, Basque Research and Technology Alliance (BRTA), 48170 Derio-Vizcaya, Spain e-mail: [email protected]; [email protected] D. Seneviratne e-mail: [email protected] D. Galar e-mail: [email protected]; [email protected] A. Gálvez · D. Galar Division of Operation and Maintenance Engineering, Luleå University of Technology, 97187 Luleå, Sweden E. Juuso Control Engineering Group, Faculty of Technology, University of Oulu, PO Box 4300, FI-90014 Oulu, Finland e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_4

43

44

A. Gálvez et al.

of the individual features that strongly distinguish the FM; (3) analysis of how the features selected vary when components degrade. Keywords Feature assessment · Fault detection · HVAC system · Diagnostics

1 Introduction Safety is becoming more important than reliability or efficiency in transportation, oil and gas, and chemical industries. Therefore, restrictive regulations, such as EN 50126–1:2017 [1], define the requirements for railway systems’ reliability, availability, maintainability, and safety. Because of these standards, critical components are often replaced in the early stages of degradation. An early substitution of components implies a loss of useful life and a lack of information about advanced stages of degradation. This lack of information can be overcome using a combination of datadriven and physics-based models, or HyMs. The fusion of models can improve the accuracy of the estimations of the remaining useful life (RUL), clarify the health state of the system, and reduce maintenance actions while keeping the system operating in the required condition. This improvement is based on diagnostics and prognostics processes, the main techniques of prognostics and health management (PHM). Diagnostic detects, isolates, and identifies degraded subsystems or components. The diagnostic process starts when a fault or abnormal behaviour is detected. The prognostics process includes the estimation of the RUL by continuously assessing changes in the behaviour of the system. It evaluates the accumulated degradation to predict the future health state of the system. If there is evidence of a failure, information from data observed in the diagnostics process is used to identify and assess the damage before RUL estimation. There are three main approaches to estimate RUL through prognostics [2]: datadriven approaches, model-based (physics-based) approaches, and hybrid modelbased approaches (HyMAs). These approaches draw on techniques used in the reliability domain, such as engineering experience and expert knowledge.

1.1 Model-Based Approaches Model-based approaches use physics-based models to predict the future state of a system. Physics-based models are built using mathematical models of the physical system to give an understanding of the physics of the monitored system [3]. Material properties, and thermodynamic and mechanical responses are some characteristics incorporated into these models to improve the monitoring of industrial systems when the state variable cannot be directly measured. Physics-based models are used for FDD processes; they also yield a good understanding of a system’s behaviour and permit the precise monitoring of its key variables. These approaches cannot be always applied, however, because characteristics or key parameters are

Feature Assessment for a Hybrid Model

45

sometimes difficult or impossible to obtain. Nevertheless, an accurate model-based approach is more effective than other approaches [2]. A dynamic model of an HVAC system for FDD was proposed by [4], and a model-based approach for FDD was recently presented in [5].

1.2 Data-Driven Approaches Data-driven approaches use mathematical models and weight parameters to predict faults and RUL. Predictions use data collected by sensors embedded in the real system. However, system complexity is not considered, so these approaches miss the relations between the data and the physical world [6]. There are many data-driven techniques, but they can be classified into two main categories [3]. The first category includes approaches based on artificial intelligence (AI). The second includes statistical approaches; these can also be used to reduce the dimensionality of the dataset [7]. Supervised, unsupervised, or semi-supervised learning processes are used to build data-driven models. The supervised learning process uses labelled datasets to predict future events; the unsupervised learning process uses unlabelled datasets; the semisupervised learning process uses both labelled and unlabelled datasets. The results of the models are evaluated in terms of accuracy, or the expected precision. This is a crucial factor of the model [7]. When there are difficulties obtaining run-to-failure data, a generative adversarial network (GAN) approach can be applied to generate synthetic data in faulty asset conditions [8]. Various data-driven approaches to fault diagnostics for HVAC chillers were presented in [9, 10]. More recently, data-driven methods were deployed in air handling units in [11].

1.3 Hybrid Model-Based Approaches Hybrid model-based approaches (HyMAs) combine model-based and data-driven approaches. This combination of models has advantages over using one or another of the root approaches [12]. The main advantage of HyMAs is the reduction of the amount of historical information required to train a data-driven model and the information needed to build a robust physics-based model. This data is combined to complete datasets that otherwise lack information on the system or process. This improves the ability to detect FMs and reduces the appearance of hidden FMs, metaphorically termed “black swan losses” [13], thus improving the FDD and prognostics processes. HyMs can improve the FDD process of critical systems. Nevertheless, the accuracy of multiple fault detection models is sometimes unacceptable. This paper assesses features extracted from synthetic signals generated by a physics-based model

46

A. Gálvez et al.

of an HVAC system installed in a passenger train carriage. The assessment leads to a proposal to improve the accuracy of a multiple fault detection model developed and presented in previous research [14]. The assessment comprises the following: • Generation of run-to-failure data using the physics-based model of the HVAC system. • Identification of the individual features that strongly distinguish the FM. • Analysis of how the features selected vary when the components degrade. The assessment of features leads to a proposed method to enhance the accuracy of the fault detection and identification process. One recent study compared the features used in the fault detection of a ball bearing [15], while another compared the ranking of features to improve the prediction of faults [16]. The paper proceeds as follows. Section 2 describes the methodology used to combine the measured and synthetic data. Section 3 describes the physics-based model developed for data generation, how the faults are modelled, the features extracted from the selected signals, the ranking of features, and the classifiers used to build the data-driven models. Section 4 discusses the results. Section 5 concludes the work and suggests future work to implement the proposed HyMA.

2 Hybrid Modelling Methodology HyM can improve the FDD process for HVAC systems. Figure 1 illustrates a methodology for combining a physics-based model with a data-driven model; this methodology was presented in previous research [14]. A multiple fault detection model was also developed in the work. The model was built using synthetic and measured data to detect faults in different components. The physics-based model was developed and validated in previous research [17]. The accuracy of the multiple fault detection model was unacceptable, so this paper attempts to improve the accuracy. It develops simpler data-driven models trained to detect faults individually, and it assesses the features that strongly identify FMs. The HyMs presented in this paper use a supervised learning method. Once the data are generated by the physics-based model and organized in a table, each simulation is labelled with a fault code indicating the presence of a fault and the type of fault. Then, the features are extracted from every signal loaded in the table. The features are related to a fault code and are used to train, validate, and test the data-driven model.

3 HVAC System Model The physics-based model of the HVAC system installed in the passenger train carriage is separated into heating subsystems, cooling subsystems, ventilation subsystems, and vehicle thermal network systems. The temperature and the concentration of CO2

Feature Assessment for a Hybrid Model

47

Fig. 1 Methodology used in the hybrid model

are managed by two ventilation subsystems, two cooling subsystems, and two heating subsystems. Figure 2 illustrates the principal components. The physics-based model also contains the model of the vehicle’s thermal network which is connected to the HVAC system. Table 1 shows the set of sensors embedded in the real system; these sensors are labelled as real. Soft sensors or virtual sensors are labelled as virtual.

Fig. 2 Principal items considered in the studied HVAC system

48 Table 1 List of parameters for feature extraction

A. Gálvez et al.

Parameter

Class

Temperature after compressor 1

Virtual Signal

Type

Temperature after compressor 2

Virtual Signal

Temperature before compressor 1 Virtual Signal Temperature before compressor 2 Virtual Signal Pressure after compressor 1

Real

Signal

Pressure after compressor 2

Real

Signal

Pressure before compressor 1

Real

Signal

Pressure before compressor 2

Real

Signal

Compression rate 1

Virtual Signal

Compression rate 2

Virtual Signal

Pressure after filter

Virtual Signal

Pressure before filter

Virtual Signal

Real heat transfer

Virtual Signal

CO2 level

Real

Signal

Vehicle temperature

Real

Signal

Impulsion temperature

Real

Signal

Fault code

–

Condition variable

3.1 Fault Modelling The FMs modelled are the (1) accumulation of dust in the air filter, (2) degradation of the CO2 sensor, (3) degradation of the evaporator fan, and (4) variations in the compression rate of the cooling system. The sensor’s drift is modelled by introducing an offset in the sensor model; the offset is controlled by a model parameter indicating the presence of a fault and its degradation state. The rest of the faults are modelled after the failure mode and effects analysis (FMEA) in which an FM is analyzed, and its causes and effects are specified. The causes of the FM are defined in the model as faults. These causes must be linked to the effects which are expected to be represented by the sensors. As indicated in Table 1, there are not enough sensors embedded in the real system to distinguish between FMs. This means the effects of several FMs are reflected in the set of sensors; therefore, a set of soft sensors is defined in the model to improve the detectability of FMs. Thus, the soft sensors are related to an FM.

Feature Assessment for a Hybrid Model

49

3.2 Ranking of Features for the Data-Driven Models The paper presents four data-driven models to detect the FMs individually. The models are trained, validated, and tested using real data and synthetic data generated by introducing noise and the presence of faults at different stages of degradation. The data-driven models presented here compute the following features for each signal listed in Table 1: mean (μ), standard deviation (σ ), skewness (γ ), kurtosis (κ), peak value (x peak ), root mean square (RMS), crest factor (CF) and shape factor (SF). These features are calculated as follows: μ=

t1 t  x(t) t1 − t0 t=t

(1)

0

  t1  t  σ = [x(t) − μ]2 t1 − t0 t=t

(2)

0

γ = κ=

t t1 t=t0 [x(t) t1 −t0 σ3 t t1 t=t0 [x(t) t1 −t0 σ4

− μ]3 − μ]4

xpeak = max|x(t)|   t1  t  RMS =  [x(t)]2 t1 − t0 t=t

(3)

(4) (5)

(6)

0

CF = SF =

xpeak RMS RMS

t t1 t=t0 |x(t)| t1 −t0

(7) (8)

Once the features are extracted from the selected signals, the features are ranked before the data-driven model is trained. The rankings are built using ranking algorithms which compare features and locate those that best distinguish the FM at the top of the ranking. The results of algorithms such as t-test, entropy, Bhattacharyya, receiver operating characteristic (ROC), and Wilcoxon are compared before the final ranking of features.

50

A. Gálvez et al.

The t-test uses the absolute value of two-sample t-tests with pooled variance estimation. The t-score of each feature is generalized as follows [18]: 1. Suppose the feature set if F = { f 1 , . . . , f i , . . . , f g } and feature i has m i different nominal values represented as f i = {xi(1) , xi(2) , . . . , xi(m i ) . 2. Transform each nominal feature value into a vector with the dimension m i : xi(1) ⇒ X i(1) = (0, . . . , 0, 1), xi(2) ⇒ X i(2) = (0, . . . , 1, 0), . . . , xi(m i ) ⇒ X i(m i ) = (1, . . . , 0, 0). 3. Replace all the numerical features in (9) and (10) with those vectors; then, see Eqs. (12) and (13). Equations 9, 10 and 11 evaluate the difference between the mean of one class and the mean of all the classes, where the within-class standard deviation difference is used to standardize the difference. tic = Si2 =

x ic − x i Mc · (Si + S0 )

C    2 1 xi j − x ic N − C c=1 j∈c

Mc =

(10)

1/n c + 1/N

X ic − X i , c = 1, 2, . . . , C ti = max Mc Si Si2 =

(9)

(11) 

C 



T  1 X i j − X ic X i j − X ic N − C c=1 j∈c

(12)

(13)

The entropy algorithm ranks features by relative entropy using an independent evaluation criterion. Further information about relative entropy is presented in [19]. Bhattacharyya ranks features using minimum attainable classification error or Chernoff bound. As it is presented in [20], the Bhattacharyya distance is simpler than Chernoff bound. However, all discussions about Bhattacharyya distance mentioned in the book cited could be extended to the Chernoff. The Bayes error is computed by Eq. 14: εu =

 P1 P2 p1 (X ) p2 (X )d X = P1 P2

(14)

  The term μ 21 defines the Bhattacharyya distance, and is used as the separability of two distributions. In general, and for normal distributions the Bhattacharyya distance is computed by Eq. 15:

Feature Assessment for a Hybrid Model    −1    + | 12 2| 1 1 1 1+ 2 T μ = (M2 − M1 ) (M2 − M1 ) + ln   2 8 2 2 | 1 || 2 |

51

(15)

ROC uses the area between the ROC curve and the random classifier slope for evaluating the performance of classification algorithms. ROC curve provides a graphical representation of a classifier’s performance. The ROC curve is obtained by computing and plotting the true positive rate against the false positive rate for a classifier at a variety of thresholds [21]. Finally, the Wilcoxon technique is a nonparametric test for two-sample problems. This technique uses the absolute value of the standardized U-statistic of a two-sample unpaired Wilcoxon test to rank features. The Wilcoxon rank sum test is a nonparametric test for two populations when samples are independent. This technique is detailed and explained in [22]. The rankings obtained from the aforementioned algorithms are used to train the data-driven models. The best accuracy is obtained using the t-test techniques; therefore, the rankings presented in this paper are developed using these techniques. In the next step, the features located at the top of the rankings are plotted and assessed. The classifiers trained in this paper are decision trees (DTs), discriminant analysis, support vector machines (SVMs), logistic regression, k-nearest neighbors (kNNs), Naïve Bayes (NB), and ensemble classification. These classifiers contain the following characteristics: • DTs is based on the structure of a tree to divide a dataset into leaves and branches while performing the decision tree incrementally. The leaves represent class labels and branches represent conjunctions of features that lead to those labels [23]. • Discriminant analysis models make predictions by finding a combination of features that characterises or separates two or more classes. The resulting combination commonly is used for dimensionality reduction before later classification [24]. • SVM is a classifier that employs linear combinations of different features to make classification decisions. SVM can manage a significant number of features as attempts to reach an optimized solution, thus avoiding overfitting and making feature selection less critical [25]. • LR are statistical models that evaluate the relationship between the mean response and one or more predictor variables which can be qualitative [26]. Otherwise, when the variables are dichotomic variables, the LR model is called binary or binomial logistic regression, and when there are variables with more than two values, they are called multinomial logistic regression. • k-NN is a nonparametric algorithm that generates predictions. The inputs are the k-nearest samples in the feature space and the output can be either the most represented class among them using a distance function in k-NN classification or a value in k-NN regression, a value which is the average of the values of k-NN. • NB classifiers are a family of simple probabilistic learner based on applying the Bayesian rule. NB assumes that the values of features are independent of any other feature, given the class label [27].

52

A. Gálvez et al.

• Ensemble classification integrates multiple models in order to improve the accuracy and reliability of estimations or decisions that can be obtained from using single model alone [28]. These models are configured to avoid overfitting by applying cross-validation with five folds. These specifications are loaded to train, validate, and test the four models using supervised learning.

4 Results and Discussion The paper develops four models; one model detects the FM of dust in the air filter; the second is trained to detect degradation of the CO2 sensor; the third detects degradation of the evaporator fan; and the fourth identifies variations in the compression rate of the cooling system.

4.1 First Case: Dust in Air Filter This model is trained using features containing both healthy data and data when the air filter contains a mass of dust. A few classifiers obtain an accuracy under 95%; DT classifiers reach an accuracy of 100%. Thus, the ranking of features developed by t-test techniques reduces the dimensionality of this model efficiently. The features that best detect an obstruction of the air filter are plotted in Fig. 3. The figure shows the evolution of the features as the mass of dust increases in the air filter. The values of features are normalized between 0 and 1 to ease the comparison. For the CO2 level, the mean is the unique feature plotted in Fig. 3 that is extracted from signals which can be measured in the real system. The rest of the features are extracted from signals from soft sensors. This highlights the importance of defining soft sensing to improve FDD and diagnostics processes. The feature labelled as real heat transfer—peak value best identifies the appearance of a mass of dust in the air filter. As Fig. 3 shows, the value changes dramatically after early stages of degradation. Nevertheless, this feature is not good for the prognostics process; even when the values change after the early stages of degradation, the evolution of this feature is not predictable as the mass of dust increases. The rest of the features plotted are also good at identifying the FM, and they are better than the real heat transfer—peak value for assessing the severity of the FM because the trending is more linear; therefore, it can be predicted.

Feature Assessment for a Hybrid Model

53

Fig. 3 Feature evaluation for muss of dust in the air filter

4.2 Second Case: Degradation of the CO2 Sensor This model is trained and validated using real data and synthetic data to detect faults and deviations in the CO2 sensor. The accuracy obtained is 100% in SVM and k-NN classifiers, and 98% in DT and NB classifiers. Thus, this FM is well detected using an HyM. Figure 4 shows the features located at the top of the ranking and how they vary when the degradation of the sensor changes. The degradation of the sensor shows values from 0.5 to 1.5, with 1 indicating a healthy state. All features plotted in Fig. 4 come from the signals collected by the CO2 sensor. The features extracted from the CO2 are expected to be the most representative for detecting this FM because the CO2 sensor itself is failing. The values of this sensor directly affect the regulation of fresh air inside the cabin. Thus, if a failure is detected or a deviation is prolonged for a long period, the HVAC system will supply more fresh air flow than required, increasing the consumption of the system to keep comfortable conditions in the cabin. Alternatively, the HVAC system may supply less fresh air flow than needed; this will increase the concentration of CO2 in the cabin, a critical parameter for passengers’ safety and comfort. Therefore, features obtained from signals related to those parameters are also useful for detecting that fault, even when the score is not remarkable.

54

A. Gálvez et al.

Fig. 4 Most representative features for detecting the CO2 sensor deterioration

4.3 Third Case: Degradation of Evaporator Fan The evaporator fan is used to move the air flow through the HVAC system. The model is built using synthetic data; the highest accuracy obtained is 90% using DT classifiers. Figure 5 represents the features that strongly identify this FM. The features are related to signals coming from the temperature of the air going from the HVAC system to the cabin, the fresh air flow, and the pressure after the air filter. The healthy state is defined by the number 1 on the X axis of Fig. 5. The features vary as the degradation increases, making it difficult to identify the severity of the failure. These variations may be the main reason why the accuracy obtained is 90%.

4.4 Fourth Case: Variations in the Compression Rate of the Cooling System The compression rate is evaluated with the pressure measured before and after the compressor of the cooling subsystem. The models developed to detect this fault reach an accuracy of 100% using different DT classifiers. The features obtaining the best score in the ranking do not provide a good illustration of the failure because they directly depend on the behaviour of the compressor. The relations of all these features are directly proportional to the degradation level, and they show the same trending, as Fig. 6 indicates.

Feature Assessment for a Hybrid Model

Fig. 5 The most representative features for detecting the evaporator fan degradation

Fig. 6 The most representative features for detecting variations in the compression rate

55

56

A. Gálvez et al.

Summarizing, Fig. 3 illustrates the features labelled pressure drop filter—std and pressure drop filter—clearance factor. Figure 4 illustrates how the features have the same trending for the degradation of CO2 . Moreover, Fig. 6 shows the same trending related to other signals and FM. These features are extracted from the signal defined as pressure drop filter. Thus, as shown in Figs. 3, 4, and 6, some features may show a similar trending and obtain similar scores in the ranking. This must be considered because one of the main purposes of building a ranking of features is the removal of features whose score positions them at the bottom of the ranking, thus reducing the dimensionality of datasets and making the data-driven model more agile. But if many features positioned at the bottom of the ranking are removed from the dataset, the final datasets will contain redundant features and may miss features that could detect a fault in strange conditions, even if they do not get a high score. The definition of a threshold for an unacceptable score is crucial to develop a robust, agile, and accurate data-driven model.

5 Conclusions and Outlook The paper assesses features and HyMs for an HVAC system located in a passenger train carriage. A physics-based model is used to generate data by simulating different levels of degradation in several components of the HVAC. This data is used to build various data-driven methods for FDD. The data-driven methods trained for comparative purposes are DTs, discriminant analysis, SVM, logistic regression, k-NN, NB, and ensemble classification. A ranking of features is developed to select the features that strongly identify the FM. The four models are most accurate when the features come from the ranking developed by t-test techniques; they recognize the FM most accurately. Some features coming from the same signal have the same trending. Moreover, some features that strongly identify an FM are found to match. These problems can be attenuated by defining a threshold to indicate the minimum acceptable value of a feature. The accuracy of the multiple fault detection HyM will be improved by selecting a threshold that considers all the features that obtain a representative score, higher than 1. The consideration of features with a low score can also distinguish the FM, helping to overcome the aforementioned issues. The results suggest the use of larger datasets to develop a multiple fault detection model. Once multiple faults can be accurately detected using a unique model, the prognostics process will be able to estimate the RUL of the system. This will provide useful information and extend the useful life of the HVAC system.

Feature Assessment for a Hybrid Model

57

References 1. EN 50126-1 (2017) Railway applications—the specification and demonstration of reliability, availability, maintainability and safety (RAMS). European Committee for Standardization 2. Liao L, Köttig F (2014) Review of hybrid prognostics approaches for remaining useful life prediction of engineered systems, and an application to battery life prediction. IEEE Trans Reliab 191–207 3. An D, Kim NH, Choi J-H (2015) Practical options for selecting data-driven or physics-based prognostics algorithms with reviews. Reliab Eng Syst Saf 133:223–236 4. Bendapudi S, Braun J, Groll E (2002) A dynamic model of a vapor compression liquid chiller. In: International refrigeration and air conditioning conference, vol 568 5. Yul Chu S, Avestruz A-T (2019) Electromagnetic model-based foreign object detection for wireless power transfer. In: 2019 20th Workshop on control and modeling for power electronics (COMPEL), Toronto, Canada 6. Lyengar S, Lee S, Irwin D, Shenoy P, Weil B (2018) WattHome: a data-driven approach for energy efficiency analytics at city-scale 7. Mirnaghi MS, Haghighat F (2020) Fault detection and diagnosis of large-scale HVAC systems in buildings using data-driven methods: a comprehensive review. Energy Build 229 8. Ortego P, Diez-Olivan A, Del Ser J, Sierra B (2020) Data augmentation for industrial prognosis using generative adversarial networks. Lect Notes Comput Sci 12490:113–122 9. Choi K, Namburu SM, Azam MS, Luo J, Pattipati KR, Patterson-Hine A (2005) Fault diagnosis and HVAC chillers. IEEE Instrum Meas 24–32 10. Madhavi Namburu S, Azam MS, Luo J, Choi K, Pattipati K (2007) Data-driven modeling, fault diagnosis and optimal sensor selection for HVAC chillers. IEEE Trans Autom Sci Eng 4(3):469–473 11. Montazeri A, Mohamad Kargar S (2020) Fault detection and diagnosis in air handling using data-driven methods. J Build Eng 31(101388) 12. Hernández Mejías ÁM, Galar D (2014) Techniques of prognostics for condition-based maintenance in different types of assets, 1 edn. Luleå University of Technology, Graphic Production, Luleå 13. Aven T (2013) On the meaning of a black swan in a risk context. Saf Sci 57:44–51 14. Galvez A, Diez-Olivan A, Seneviratne D, Galar D (2020) Synthetic data generation in hybrid modelling of railway HVAC system. In: Proceedings of the 17th IMEKO TC 10 and EUROLAB virtual conference: “Global trends in testing, diagnostics & inspection for 2030”, pp 79–84 15. Vakharia V, Gupta VK, Kankar PK (2016) A comparison of feature ranking techniques for fault diagnosis of ball bearing. Soft Comput 20:1601–1619 16. Rathore SS, Gupta A (2014) A comparative study of feature-ranking and feature-subset selection techniques for improved fault prediction. In: Proceedings of the 7th India software engineering conference, pp 1–10 17. Gálvez A, Seneviratne D, Galar D. Development and synchronization of a physics-based model for heating, ventilation and air conditioning system integrated into a hybrid model. Int J Hydromechatron (in press). https://doi.org/10.1504/IJHM.2021.10034926 18. Zhou N, Wang L (2007) A modified T-test feature selection method and its application on the HapMap genotype data. Genomics Proteomics Bioinform 5:242–249 19. Li Y, Cai W, Li Y, Du X (2020) Key node ranking in complex networks: a novel entropy and mutual information-based approach. Entropy 1(52) 20. Fukunaga K (1990) Chapter 3—Hypothesis testing. In: Introduction to statistical pattern recognition, 2nd edn. Academic Press, pp 51–123 21. James G, Witten D, Hastie T, Tibshirani R (2017) An introduction to statistical learning with applications in R. Springer, Texas 22. Lovric M (2011) International encyclopedia of statistical science. Springer, Berlin Heidelberg 23. Piryonesi SM, El-Diraby TE (2020) Data analytics in asset management: cost-effective prediction of the pavement condition index. J Ingrastruct Syst 26(1)

58

A. Gálvez et al.

24. Martinez AM, Kak AC (2001) PCA versus LDA. IEEE Trans Pattern Anal Mach Intell 23(2):228–233 25. Moosavian A, Ahmadi H, Sakhaei B, Labbafi R (2014) Support vector machine and K-nearest neighbour for unbalanced fault detection. J Qual Maint Eng 20(1):65–75 26. Wakiru J, Pintelon L, Muchiri P, Chemweno P (2020) A data mining approach for lubricantbased fault diagnosis. J Qual Maint Eng 27. Zhang C, Liu C, Zhang X, Almpanidis G (2017) An up-to-date comparison of state-of-the-art classification algorithms. Expert Syst Appl 82:128–150 28. Santos P, Amado C, Coelho ST, Leitão JP (2017) Stochastic data mining tools for pipe blockage failure prediction. Urban Water J 14(4):343–353

Fatigue and Wear Prediction

Operability Forecasting Combining Neural Network and Survival Analysis with an Application to Hot Corrosion in Turbofan Raphaël Langhendries and Jérôme Lacaille

Abstract In this article, we present a statistical model that aims to predict the wear of turbofan components. The model is composed of a neural network calculating the wear of the targeted component and a parametric model describing the exposure of the material. The two parts are trained together using gradient-based optimization. We tested the model on simulated data. In addition, we have studied the problem of hot corrosion on turbofans and used our model for the corrosion prognosis of high-pressure turbine components. Keywords Artificial neural network (ANN) · Survival analysis · Prognostic health management · Hot corrosion · Turbofan engine

1 Introduction Aircraft engines generate the power necessary to propel the aircraft; modern civilian aircraft engines are turbofans. A simplified diagram of a turbofan is given below (Fig. 1). Maintenance is essential to ensure good operating conditions. Due to the normal aging process, some components may need to be replaced after a certain period of operation. Many maintenance strategies exist depending on the characteristics of the targeted component and the maintenance policy. In any case, every maintenance operation results in costly downtime.

R. Langhendries Université Paris I - SAMM, 90 rue de tolbiac, Paris, France e-mail: [email protected] R. Langhendries · J. Lacaille (B) Safran Aircraft Engines, 77550 Moissy-Cramayel, France e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_5

61

62

R. Langhendries and J. Lacaille

Fig. 1 High bypass ratio turbofan engine. The gas generator is shown in magenta; it consists of a compressor, a combustion chamber and a high-pressure turbine. Most of the airflow comes out of the generator and is propelled by a fan driven by a low-pressure turbine through a second shaft, in green in the figure. Source K. Aainsqatsi, CC BY-SA 3.0, via Wikimedia Commons

In this context, the condition monitoring of turbofans is decisive in properly planning maintenance in order to limit downtime and thus reduce operating costs. In this regard, research has been undertaken to develop physical and statistical methods to model various aspects of turbofan aging [10, 18]. We propose a machine learning model which aims to provide a probability that a component must be maintained conditionally for the use of the engine.

2 Background and Data 2.1 Inspections As a part of the engine’s maintenance policy, regular maintenance operations are planned. If performance degradation is observed, unforeseen operations may also occur. In both cases, the engine is inspected and some of its components are checked (whatever the initial purpose of the maintenance operation). After this control, each component can be classified as operable and no other action is taken or non-operable; in this case, the component is repaired or replaced according to the maintenance policy, in addition, the reason is recorded. We emphasize that in this context, non-operability does not mean that the component is broken; it means that it has reached a certain limit regarding its specification (even if these limits are extremely pessimistic as a precaution). A reason of non-operability is abbreviated by RnO. The RnOs cover a wide range of possible damages due to many phenomena (e.g. wear, fatigue). For example, “fretting”, “corrosion” and “crack” are common RnOs.

Operability Forecasting Combining Neural Network …

63

We emphasize that an RnO can be caused by the joint action of different physical phenomena. For example, the appearance of a crack can result from wear and tear or fatigue and the two can be involved together; see [4]. Throughout the article, we focus on one RnO and a given component of the turbofan. Our goal is to estimate the probability that this component requires maintenance due to this specific cause.

2.2 Engine Utilization We aim to estimate the probability of non-operability conditional on the use of the engine. Each flight performed by an engine is represented by a vector X t (for the flight number t). Then the use of the engine corresponds to the time series (X t )t∈[1,n] where n is the number of flights. One vector X t represents one flight, and we aggregate two types of data: Snapshots. Commercial flights are divided into phases (TakeOff, Climb, Cruise, etc.). For each phase, at a given moment, the engine computer records the measurements of the engine and aircraft sensors. Those records are called Snapshot; we give in Table 1 a non-exhaustive example of data that can be found in a snapshot. Environment Data. Additionally to the snapshot, we also use data describing the environment where the engine is operated (the departure and arrival airport). It includes meteorological data (temperature, pressure, etc.) and concentration in air of several elements (like dust, salt, sulfur dioxide...).

Table 1 A non-exhaustive list of parameters recorded during flight snapshots Parameter Description TAT Altitude Mach number Ground speed N2 T25 EGT N1 PS3 T3 T12 PT2 Fuel flow VBV position Oil quantity/temperature/pressure

External temperature Altitude Instant speed Speed with respect to the ground Core speed Compressor inlet temperature Exhaust gas temperature Fan speed Compressor discharge pressure Compressor discharge temperature Inlet temperature Inlet total pressure Fuel flow Variable bleed valve position Oil quantity/temperature/pressure

64

R. Langhendries and J. Lacaille

2.3 Operability Forecasting Eventually, the problem we address in this paper is to build a statistical model which gives the probability that a component of a turbofan is non-operable because of a specific RnO. Several points are important to note: – An RnO cannot be linked to a single physical phenomenon of aging (such as wear or fatigue) but corresponds to an observation made by an operator during a workshop visit. – Many data describing the use of the engine are available (a time series (X t )t∈[1,n] for each engine). However, there is no prior guarantee that this data contains enough information to effectively predict the RnO. – We only know if the component we are interested in is operational or not when an inspection is carried out during a workshop visit. Therefore, there is a long period during which we have no information about the state of the component. In particular, if the last inspection showed that the component was operational, that does not mean that the component is still operational now. It corresponds to data censored on the right. Likewise, if the last inspection concludes that a component is inoperative, we do not know how long the component has been inoperative. This corresponds to left-censored data.

3 Related Approaches 3.1 Cumulative Damage Model (CDM) Cumulative damage models have been used for a long time to model various types of damages (see [5] for an overview or [6] for a more recent example in the aerospace field). Cumulative models consist in modeling the damage as an increasing function of a damaging parameter. There are many different functions and possible damaging parameters depending on the underlying physical phenomenon (the most common damaging parameters are operating time or the number of cycles). In our situation, the main problem with CDM is that our RnO often cannot be explained by a single physical phenomenon. It is the result of interactions between many of them. Therefore the choice of the damage function becomes crucial.

3.2 Survival Analysis Survival analysis [12] is a statistical area that aims to model a duration until an event occurs. Usually, the event is called death or failure, and the time until failure is called lifetime. The lifetime is described by a survival function S(t) giving the probability to survive (i.e. to not fail) longer than t.

Operability Forecasting Combining Neural Network …

65

The main issue of survival analysis is the estimation of the survival function. For example, the Kaplan-Meier estimator can be used (see [15]). Another possible approach is to approximate the survival function by a parametric model. For this purpose, various parametric models have been proposed. In health sciences, the Cox model is widely used [1]. For material aging, especially in the aerospace industry, Weibull law [19] is a common choice. For example, it has been used in [21]. In our case, to model the lifetime, we propose to consider a large amount of different data (and not only the operating time t). This is why we do not directly use a parametric model to approximate a survival function.

4 Model 4.1 Non-recurrent Model We first introduce the model represented in Fig. 4 in Appendix. In Fig. 4, the time series (X i )i=1..n corresponds to the series of flights achieved by an engine. The model called Exposure model is a neural network that calculates from each flight data X i a vector called Exposure vector Vi which represents the impact on the component resulting from flight X i . Then, we sum all impacts (building a sum vector V = i=1..n Vi in the figure), and we select the most critical feature τ = max(V ) (which we call Exposure count). Finally, the probability of non-operability is computed using Exposure count τ as the input of a Weibull Cumulative Distribution Function (CDF). Parameter learning. In the proposed application, the exposure model is a neural network (in practice, it consists of a few dense layers). Therefore, the parameters of this neural network must be learned. In addition, two parameters appear in the cumulative Weibull distribution function (shape and scale). All these parameters are learned at the same time by the gradient-based descent minimization of an error on a training set. Interpretation. The proposed application has common points with the CDM. Indeed, the exposure model plays the same role as the damage function involved in a CDM. Except that we are using a neural network and not a predefined damage model. The proposed model is also linked to survival analysis in that we use a Weibull cumulative distribution function (CDF) to model the wear process until nonoperability. The Weibull CDF of our model may be seen as 1 − S(t) where S(t) is a survival function and t is the damaging parameter. Usually, in survival analysis t is an operating time. In our proposal, this is a more complex calculation. Implementation. We used PyTorch to implement the neural network, We add the Weibull cumulative distribution function and its partial derivatives (to allow gradient back-propagation). The optimization algorithm was ADAM [11] with a dropout regularization scheme [17]. Another crucial choice is the loss function. In our case, the training set contains engines labeled with 1 if the targeted component has been considered as non-operable

66

R. Langhendries and J. Lacaille

after an inspection (occurring at the end of the time series (X i ), i = 1..n) or with 0 if the component is operable. The result of our model is a probability, so it will be in the interval [0, 1]. As this situation corresponds to binary classification, we select the Binary Cross Entropy Loss provided by PyTorch. The first limit. In our previous proposal, an important implicit assumption is that a flight X i affects the engine component regardless of the current state of the engine. Especially, a flight has the same impact on a recent or an old engine. For many types of wear, this assumption may not be verified. To address this issue, we propose the following extension.

4.2 A Recurrent Model The recurrent model can also be found in Appendix (see Fig. 5); it is identical to that of the previous Fig. 4, except that we add a recurrent component: the GRU block. The GRU block is a Gated Recurrent Unit [2]. It is a kind of recurrent neural network similar to the Long Short-Term Memory (LSTM) with many applications including health monitoring [20]. We add this block to compute a hidden state of the engine called h i representing the state of the engine after each flight X i . Therefore, the wear impact caused by flight X i+1 can be computed taking into account the current engine state h i .

4.3 Multiple Components In both versions (Figs. 4 and 5), the exposure model computes the wear impact and a Weibull cumulative distribution function models the resilience of the components to this wear. Consequently, the two parameters of the Weibull cumulative distribution function only depend on the characteristics of the components (alloy, geometry, etc.). We also emphasize that it is usual to use the same material several times in a turbofan. Therefore, if the same type of component is used multiple times in different places of the turbofan, we can use our survival model to forecast the lifetime of each occurrence of the component. We thus build an exposure model per component occurrence, but we use the same cumulative Weibull distribution for all occurrences. This remark is important. Indeed, in practice, we will often not have many inspected engines at our disposal to train our model. A better advantage of each available engine is allowed by using different occurrences of the same component.

4.4 Weibull Law In our model, we use the Weibull CDF; this choice is motivated by the wide use of Weibull law in the field of reliability engineering. Nevertheless, if a specific RnO is

Operability Forecasting Combining Neural Network …

67

known to be associated with another probability law, the model can easily be adapted. Moreover, to simplify, we chose the Weibull law with two parameters. However, a Weibull law with three parameters could have been used as an extension of the Weibull law.

4.5 Simulation To test the model, we carry out simulations. We have chosen a function f to be a predefined exposure function. We also fix the two parameters for the Weibull cumulative distribution function: the shape k = 6.5 says that the risk of wear increases with time, and the scale λ = 350 gives 63% of observations of non-operability when the exposure reaches 350. We also generate 1000 engines, and select n = 45 parameters per flight and five components per engine. Then we draw randomly numerous flights for each engine and parameters for each flight. Lots of choices for the deterministic exposure function are possible; we build a function that depends on each variable and is not a linear combination of the input (in order not to be easy to learn). f is defined by n f (λ,α) (X t ) =

i=1

αi ln(1 + λi + xti ) n i=1 |αi |

(1)

with X t = (xti )i∈[1,n] the vector representing the flight number t. λ = (λi )i∈[1,n] and α = (αi )i∈[1,n] are parameters independently draw from the same law for each engine. Weibull parameters and probability laws are chosen in order to generate both operable and non-operable engine components. We evaluate our model on a validation set made up of simulated engines with the same survival law and exposure function but not used for the training. The validation cost evolution with the number of learning steps is shown in Fig. 2. Fig. 2 Evolution of the loss function on the validation set during the iterations of the learning process

68

R. Langhendries and J. Lacaille

The validation loss decreases with the number of iterations. Therefore, the model is learning. Nevertheless, we emphasize that the framework of the simulation is clearly favorable to our model (using a deterministic exposure function and Weibull law). This is why we are now going to focus on an application with real data.

5 Application to Hot Corrosion 5.1 Hot Corrosion Issue for Turbofan Hot corrosion is a mechanism of corrosion for metal that occurs at high temperatures [13]. Hot corrosion is a problem for all mechanical systems exposed to high temperatures including gas turbines [8] and turbofans. This is why research has been carried out to forecast different types of hot corrosion damages [3, 16]. Hot corrosion is a challenging issue. Indeed, it has interactions with several other degradation processes (such as fatigue). We also emphasize that in our case, the general word corrosion may refer to three different physical phenomena: hot corrosion type 1, hot corrosion type 2 and oxidation. See Chap. 8 from [7] for further information on hot corrosion and oxidation. Depending on temperature and pressure conditions, corrosion of types 1, 2 or oxidation can be preponderant (see Chap. 8 from [7] and especially Fig. 8.21). When the turbofan is operated, these conditions vary, therefore the two kinds of hot corrosion and oxidation can appear and even coexist. Moreover, when a component is non-operable due to corrosion, we do not know what type of corrosion is involved. Finally, the severity of hot corrosion attacks is known to be influenced by chemicals that can act as a catalytic (for example SO3 [14]). Therefore, air pollution can influence corrosion as well as the combustion reaction in the turbofan. The dataset with which we trained our model consists of 32 engines. Each engine did a different number of flights (in an order of magnitude of 103 ) and we had three occurrences of the same component (see Sect. 4.3). Engines may be controlled more than one time. In practice, engines at our disposal are inspected one or two times. To forecast corrosion non-operability, we used the recurrent model presented in Fig. 5 (the non-recurrent model provides worse results). Our exposure model was a three-dense-layer neural network. We summarize the numeric parameters of our model: 30 neurons per layer in the exposure neural network, 10 neurons for the GRU layer, a dropout rate of 0.4, and a learning rate of 1e−3 .

5.2 k-Fold Cross-Validation To evaluate our model, we use a cross-validation procedure (see Sect. 7.10 from [9]). Cross-validation is a statistical procedure used to estimate how accurately a

Operability Forecasting Combining Neural Network …

69

Fig. 3 For each component, we draw a different box plot. For each box plot, engines are split into two categories: engines classified as non-operable due to corrosion and the other. Then, in the ordinate axis, we display the probability predicted by the model. The first and the third quartiles delimit the box, the median is represented by a horizontal line in the box, “whiskers” are drawn from the 5th percentile to the 95th percentile and outliers are plotted as individual points

predictive model performs in practice. It consists in randomly splitting the dataset into n subsets and successively using one of the subsets as a validation set and training the model with all the others. In that way, each data is used once in the validation set. In practice, we perform a 10-fold cross-validation. Thus, each validation set is composed of 4 or 3 engines. A visualization of the results is provided in the figure below (Fig. 3). We expect that the model associates higher probabilities with corroded engines. We can see in the figure above that for components 1 and 3, the model clearly distinguishes corroded engines from non-corroded. Results for component 2 are less conclusive.

6 Conclusion Our goal was to build an efficient survival model for the wear of turbofan parts. We did not want to use as usual the time or the cycle (number of flights) as a reference but a better indicator corresponding to the real impact of exposure during each flight. Unlike cumulative damage models (CDMs) which are built using expert knowledge, we wanted to deal with complex issues like hot corrosion where a cumulative exposure rate is difficult to design. We wanted to check if artificial intelligence could

70

R. Langhendries and J. Lacaille

automatically solve this problem, even if the number of observations during ground inspections remains low. We have thus designed an application to provide a non-operability criterion for turbofan parts. Two versions of the algorithm are provided: the first one only accumulates data flight after flight, while the second also stores the state of the component to be monitored. Our algorithm can be adapted to predict various types of wear. It integrates a probabilistic survival law to assess a lifetime. We have proposed a Weibull example, but it can be replaced by any relevant probability distribution. Also, note that the model is suitable for handling censored data. This algorithm was first tested on simulated data to test its effectiveness. Then we show that it gives interesting results for predicting the hot corrosion of high-pressure turbine parts.

7 Appendix See Figs. 4 and 5.

Fig. 4 Non-recurrent model: the flight data (xi1 , . . . xim ) at each cycle i comes from the bottom-left and goes into the neural network (called “exposure model”) to calculate an impact vector that accumulates in an overall “exposure vector” flight after flight. Hopefully, this vector describes different types of exposure responsible for wear. We take the worst-case computation τ as input to a Weibull survival model

Nonoperability probability

Weibull CDF

Exposure count

Exposure vector

Exposure vectors

Input

Input

Input Exposure time

Exposure time

Exposure time

Exposure model

Flights

Operability Forecasting Combining Neural Network …

71

Fig. 5 For this recurrent version, the series of flight data still comes from the bottom left, but an additional value is added to the input data of the neural network. This new information h i is a latent variable representing the current hidden state of the engine. h 0 is initiated by zero and evolves through a GRU gate network to keep track of the internal engine state

Nonoperability probability

Weibull CDF

Exposure count

Exposure vector

Exposure vectors

Input

Input

Input Exposure time

Exposure time

Exposure time

GRU

GRU

Exposure model

GRU Hidden engine state

References 1. Cheung YB, Gao F, Khoo KS (2003) Age at diagnosis and the choice of survival analysis methods in cancer epidemiology. J Clin Epidemiol 56(1):38–43. https://doi.org/10.1016/ S0895-4356(02)00536-X 2. Cho K, van Merrienboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, Bengio Y (2014) Learning phrase representations using RNN encoder-decoder for statistical machine translation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp 1724–1734. Association for Computational Linguistics, Doha, Qatar. https://doi.org/10.3115/v1/D14-1179 3. Eliaz N, Shemesh G, Latanision R (2002) Hot corrosion in gas turbine components. Eng Fail Anal 9(1):31–43. https://doi.org/10.1016/S1350-6307(00)00035-2 4. Fan H, Keer LM, Cheng W, Cheng HS (1993) Competition between fatigue crack propagation and wear. J Tribol 115(1):141–147. https://doi.org/10.1115/1.2920967 5. Fatemi A, Yang L (1998) Cumulative fatigue damage and life prediction theories: a survey of the state of the art for homogeneous materials. Int J Fatigue 20(1):9–34. https://doi.org/10. 1016/S0142-1123(97)00081-9 6. Gerhardinger D, Domitrovi´c A, Bazijanac E (2020) Fatigue life prognosis of a light aircraft landing gear leg. In: Annual conference of the PHM Society, p 9. https://doi.org/10.36001/ phmconf.2020.v12i1.1245 7. Gialanella S, Malandruccolo A (2020) Aerospace alloys. oCLC: 1126541492. https://doi.org/ 10.1007/978-3-030-24440-8

72

R. Langhendries and J. Lacaille

8. Gurrappa I, Yashwanth I, Mounika I, Murakami H, Kuroda S (2015) The importance of hot corrosion and its effective prevention for enhanced efficiency of gas turbines. In: Gurrappa I (ed) Gas turbines—materials, modeling and performance. InTech. https://doi.org/10.5772/ 59124 9. Hastie T, Tibshirani R, Friedman J (2009) The elements of statistical learning. Springer series in statistics. Springer New York, New York, NY. https://doi.org/10.1007/978-0-387-84858-7 10. Hwang W, Han K (1986) Cumulative damage models and multi-stress fatigue life prediction. J Compos Mater 20(2):125–153. https://doi.org/10.1177/002199838602000202 11. Kingma DP, Ba J (2015) Adam: a method for stochastic optimization. In: Bengio Y, LeCun Y (eds) 3rd International conference on learning representations, ICLR 2015, San Diego, CA, USA, May 7–9, 2015, conference track proceedings. http://arxiv.org/abs/1412.6980 12. Kleinbaum DG, Klein M (2012) Survival analysis: a self-learning text. Statistics for biology and health. Springer New York, New York, NY. https://doi.org/10.1007/978-1-4419-6646-9 13. Nicholls J, Simms N (2010) Gas turbine oxidation and corrosion. In: Shreir’s corrosion. Elsevier, pp 518–540. https://doi.org/10.1016/B978-044452787-5.00026-3 14. Pettit F (2011) Hot corrosion of metals and alloys. Oxid Metals 76(1–2):1–21. https://doi.org/ 10.1007/s11085-011-9254-6 15. Ragab A (2016) Remaining useful life prediction using prognostic methodology based on logical analysis of data and Kaplan-Meier estimation. J Intell Manuf 16 16. Salehnasab B, Poursaeidi E, Mortazavi S, Farokhian G (2016) Hot corrosion failure in the first stage nozzle of a gas turbine engine. Eng Fail Anal 60:316–325. https://doi.org/10.1016/ j.engfailanal.2015.11.057 17. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R (2014) Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res 15(56):1929–1958. http:// jmlr.org/papers/v15/srivastava14a.html 18. Taha HA, Sakr AH, Yacout S (2019) Aircraft engine remaining useful life prediction framework for industry 4.0. In: 4th North America conference on industrial engineering and operations management, Toronto, Canada, p 12 19. Weibull W (1951) A statistical distribution function of wide applicability. J Appl Mech 18:7. https://doi.org/10.1115/1.4010337 20. Zhao R, Wang D, Yan R, Mao K, Shen F, Wang J (2018) Machine health monitoring using local feature-based gated recurrent unit networks. IEEE Trans Ind Electron 65(2):1539–1548. https://doi.org/10.1109/TIE.2017.2733438 21. Zhong S, Li Z, Lin L (2017) Probability evaluation method of gas turbine work-scope based on survival analysis. In: 2017 Prognostics and system health management conference (PHMHarbin). IEEE, Harbin, China, pp 1–6. https://doi.org/10.1109/PHM.2017.8079225

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission Housam Mohammad , Pavel Mazal , Frantisek Vlasic , and Baraah Maya

Abstract This paper demonstrates the results of testing toroidal bearings using the parameters extracted from Acoustic Emission testing. The aim is to find the characteristics of the Rolling Contact Fatigue (RCF) in these special bearings during testing. Toroidal bearings are characterized by their special design, which combines the self-aligning capability with the axial displacement ability, which contributed to the significant increase in their applications, especially in the wind turbine industry. The main concentration of this paper has been on the period of test before the appearance of pitting. We believe that analyzing and studying the Acoustic Emission (AE) signal in the period before pitting will make it possible to specify the pattern of the signal that can be later used to predict the initiation of pitting. The experiments took place in a specially designed stand for this type of bearing to obtain the best results from AE sensors. This study is an initial step toward other more advanced steps to better understand the crack initiation and pitting mechanisms in the especially important toroidal bearings. Keywords Acoustic emission · Rolling contact fatigue · Toroidal bearings · Pitting

1 Introduction Toroidal roller bearings are single-row bearings with long, crowned rollers. The concave raceways in the inner and outer rings are concentric relative to the center of the bearing. The raceway profiles are matched to each other and ensure optimum distribution of stresses in the bearing as well as low operating friction.

H. Mohammad (B) · P. Mazal · F. Vlasic · B. Maya Faculty of Mechanical Engineering, Brno University of Technology, Technicka 2, Brno 61669, Czech Republic e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_6

73

74

H. Mohammad et al.

The rollers are self-guiding. They will always automatically adopt the position at which the load is distributed over the length of the roller, Fig. 1. Where a shaft is liable to temperature-induced elongation and misalignment defects, the non-locating bearing is a particularly important concept. In this case, toroidal roller bearings have proved ideal as non-locating bearings [10], Fig. 2. In comparison with normal non-locating bearing arrangements, toroidal bearings offer significant advantages: Substantial changes in shaft length are compensated without constraint between the raceways and the rolling elements within the bearing.

Initial position where the rings are aligned

Accommodation of axial misalignment

Fig. 1 Basic characteristics of the toroidal bearing

Fig. 2 The non-locating concept

Accommodation of angular misalignment

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission

75

The design work involved in other solutions is unnecessary. Even more considerable axial displacements have no effect on the locating bearing. There is no axial distortion of the bearing system [10]. The Acoustic Emission method (AE) is a Non-destructive Testing (NDT) method that depends on capturing and processing acoustic waves, which are caused by local changes in the material. Acoustic waves propagate from the place of origin of the defect to the surface of the material, so we can locate the defects and monitor the course from their origin [1, 5, 7]. Contact fatigue is a very unfavorable phenomenon, and it is better to minimize it as much as possible. Rolling Contact Fatigue (RCF) occurs in surface layers by repeated contacts of two non-conformal surfaces [8]. Contact bodies are contacted by a mutual contact at a point, on a straight line or along a curve. Because the surfaces of the bodies meet in a small area, very high local contact pressures occur. In pure rolling, the main stresses are compressive. Shear stresses reach their maximum at a certain depth below the surface. In these places, cracks arise and grow, which penetrate the surface and lead to the peeling or crumbling of small volumes of material. This creates small cavities, which are collectively referred to as pitting. The main parameters influencing the contact fatigue strength include Hertz stress, number of cycles, surface quality and hardness, temperature, and presence of lubricants [2]. The acoustic emission AE method is used to precisely capture the beginning of pitting. It is one of the methods of non-destructive material testing based on the principle of elastic stress wave propagation in the material [3, 4]. It is a non-directional technique and the typical frequency range of AE takes place from 100 kHz to 1 MHz. AE is initiated at the microscopic level; it is much sensitive to detect the loss of any mechanical integrity; thus, it provides the condition of defects at the incipient stage [11]. Combining the AE monitoring and the AE source locator can be used to detect the incipient damage and to forecast the position of the damage in rollers, and this technique allows the monitoring of the deterioration rate of the rolling elements [9]. The basic parameters of the acoustic emission signal evaluation include Root Mean Square RMS (RMS value), counts, and frequency characteristics of the signal [5]. The most commonly used time-domain acoustic emission signal parameter is RMS. A very often used parameter of the time-domain signal is the count parameter, which is defined as the number of overshoots over the set threshold level per unit of time [6]. According to the definition of signal energy, the area enclosed by the squared amplitude of the acoustic emission signal can be used as a measure, and the energy of the acoustic emission signal can be calculated by root mean square voltage (V rms ) or mean square voltage (V ms ):

76

H. Mohammad et al.

T 1 Vms = V 2 (t)dt, T 0  Vr ms = Vms

(1)

where ΔT is average time, and V (t) is time-varying signal voltage [12]. A wide range of studies has been conducted to study the rolling contact fatigue in ball bearings and spherical bearings; in the case of toroidal roller bearings there is almost none. Toroidal bearings share some similarities with spherical roller bearings, except the fact that they can also accommodate axial displacement as well as angular misalignment. In this paper, the AE method has been used to experimentally study the behavior of toroidal bearing and the beginning of pitting under various loading and rotational regimes.

2 Material and Equipment The experiments took place in one of the facilities of the Czech bearing company ZKL, which is the company that produces the toroidal bearing under test, and where the special testing stands were designed and produced as well. Figure 3 shows the experiment setup, where the tested bearing is in the center of the big housing, right below the hydraulic loading cylinder. The tested bearing is a toroidal bearing of the type C4030, the material of all bearing parts is bearing steel 100Cr6 which is hardened and tempered to hardness HRC 58 for the rings and HRC 61 for the rollers. The basic geometrical specifications of the tested bearings are as follows: Bore diameter (d) = 150 mm, Outer diameter (D) = 225 mm, Width (B) = 75 mm, and Weight = 10.50 kg. The radial internal tolerance of the bearing is 150 µm. The tested bearing has been put under variable controllable loads, and it runs according to various testing regimes, as in Table 1. The main part of the stand is the housing that contains the tested bearing with two support bearings (cylindrical bearings), and the hydraulic loading system is mounted on this housing, as shown in Fig. 3. The upper part of the split housing contains two waveguides, which are responsible for transmitting the AE signal from the surface of the outer ring of tested bearing to outside the housing where it can be picked up using AE sensors, as shown in the schematic drawing in Fig. 4. For better accuracy, two AE sensors were attached to the two waveguides, and the results were obtained from both of them. The AE sensors used in this experiment are from the type IDK-14 with a built-in preamplifier 20 dB, frequency range [10–400 kHz], and the AE monitoring system used is the analyzer DAKEL-ZEDO with two channels. The sampling frequency used for frequency spectrum computation is 5 MHz. The frequency bandwidth was limited by digital filters: High-pass filter 50 kHz and lowpass filter 400 kHz (software switchable). The global measurement period for AE parameters was set to 1 s, because it was anticipated that the test will run for many days.

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission

77

Controllable hydraulic loading Temperature and vibration sensors

Support bearing AE sensor connected to the tested toroidal bearing through waveguide.

AE signal analyzer with the analysis software ZDeamon.

Fig. 3 The setup of the toroidal bearing testing station Table 1 The three stages of rotation speeds and loading Stage 1

Stage 2

Stage 3

Speed (rpm)

250

Loading (MPa)

2.4

3

3.6

2.4

500 3

3.6

2.4

750 3

3.6

Duration (hour)

1

1

1

0.5

1

1

0.5

1

1

Fig. 4 Schematic drawing of the main part of the experiment

78

H. Mohammad et al.

3 Results and Discussion The bearing was loaded by different loads during three rotation speeds. Two AE sensors were used to acquire the signals during the operation; they were fixed on the end of the signal waveguides that are in touch with the outer ring of the tested bearing. The received signal was processed and analyzed using a DAKEL Zedo analyzer. Three parameters of AE were used to process and analyze the signals; they are RMS, Energy, and Count of hits. The processed signals were divided into three stages according to the rotation speed of the bearing. From the experiment, as shown in Fig. 5, it was noticed. In Stage 1, where the rotation speed is relatively slow (250 rpm), the effect of different loading was significantly small. The RMS value of the signal was very low, less than 0.1 mV, although there was a small peak immediately after the loading with 3.6 MPa, which is related to the running-in phase of the bearing. The Energy parameter was almost redundant at this stage. The count hit just followed the same pattern. In Stage 2, the rotation speed was raised to 500 rpm which was reflected in a big impact on the values of RMS of the signal that reached the value of 0.3 mV. By implementing the same previous loading regime on the bearing, we notice a very small, almost negligible effect of change when raising the loading during the same rpm. The two other parameters, energy and count of hits, as the RMS, showed a relevant change in values according to the increase of rotation speed, but almost no change for the increase in loading. In Stage 3, the rotation speed reached 750 rpm, and we see a different relationship pattern between the loading and the signal. The increase in loading had a

Fig. 5 The relationship between AE parameters, the rotation speed, and loading of the toroidal bearing

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission

79

proportionate impact on the RMS values and the other two parameters, as shown in Fig. 5. It is noticed from the figure that the increase in the rotation speed of the bearing has a bigger effect on the signal parameters. The increase of the loading of the bearing during the relatively small rotation speeds has a very low effect on the signal parameters, as in Fig. 5, the energy of the signal, and the duration of the hit, while the same change in loading during higher rotation speed has resulted in a more proportionate increase in the values of signal parameters. These notes will be taken into consideration while choosing the conditions for the following series of tests that are planned on this type of bearing. In the next experiment, the rotation speed was fixed at 375 rpm, and the loading was also fixed at 172.5 kN. Two AE sensors were used, the first (A) was attached to the outer ring of the tested bearing, with the help of a waveguide, the second (B) was attached to the housing of the bearing to act as a reference. Then a temperature sensor (thermocouple, range: −40 to 260 °C) and a vibration sensor (frequency response: 2 Hz to 10 kHz ± 5% and sensitivity > 100 mV/g ± 5%) were added to the system, both of them attached to the bearing outer ring. The results of the test, which lasted for about 3.5 days, are presented in Fig. 6. The upper graph (a) in Fig. 6 shows the RMS of the AE signal received from sensor A, plotted with the vibration curve, which represents the RMS of the vibration value, and the lower graph (b) shows the RMS of AE signal received from the sensor B, plotted with the temperature curve. As shown from the upper graph (a) in Fig. 6, at the beginning of the test there were high levels of AE due to the running-in stage of the new tested bearing, then the signal stabilized for most of the time of the test until the third day when the beginning of the assumed effects of RCF started to take place. This assumption comes from the big increase in the energy, and RMSAE of the AE hits as well as the increase in vibration level which also indicated the initiation of the cracks that developed later to form the pits, which increased in number to raise the value of the vibration dramatically and with it the temperature of the bearing, and to trigger the stop of the test. The main observation of this graph is that the temperature of the outer ring, and supposedly of the inner ring as well, started to rise after a long time of the start of an increase in vibration level which indicated the beginning of pitting formation on the tracks in the inner and outer rings. This is important because it means that the temperature of the rings is indicative but less responsive to the changes regarding the effects of RCF in the bearing parts. In other words, the temperatures of the inner and outer rings are not good parameters for a fast and accurate representation of the changes in the bearing parts regarding the RCF effects.

80

H. Mohammad et al.

Fig. 6 a RMS of AE from the sensor mounted on the bearing outer ring and the vibration values from the same location, and b RMS of AE from the sensor mounted on the housing and temperature of the bearing outer ring

Active Monitoring of RCF in Toroidal Bearings Using Acoustic Emission

81

4 Conclusion A toroidal bearing of the type C4030V was tested in a special stand and AE was used as a NDT method. The received signal was analyzed and divided into three stages according to the rotation speed, and the same gradual increase of the loading regime was applied at each stage. The AE signal parameters were more sensitive to the increase in rotation speed than the changes in loading, therefore, the selection of the rpm is more important in any following test. The temperature of the bearing rings is not a reliable parameter for fast and accurate detection of the beginning of the RCF effects on the bearing, but it still can be used as an indicative parameter. This work is the first step of long-term research to study and monitor toroidal bearing behavior with regard to rolling contact fatigue, and it is still in its first stage, where we are collecting, analyzing, and presenting the data. Acknowledgements This work was supported by the project of the Technology Agency of the Czech Republic No. TH 02010306 “Research and development of new generation design and technology of spherical roller thrust bearings”.

References 1. Choudhury A, Tandon N (2000) Application of acoustic emission technique for the detection of defects in rolling element bearings. Tribol Int 33(1):39–45 2. Harris TA, Kotzalas MN (2006) Advanced concepts of bearing technology 5th edn. Taylor & Francis: CRC Press 3. Hase A, Mishina H, Wada M (2012) Correlation between features of acoustic emission signals and mechanical wear mechanisms. Wear 292–293:144–150 4. Kopec B (2008) Nedestruktivní zkoušení materiál˚u a konstrukcí: (nauka o materiálu IV). Akademické nakladatelství CERM, Brno. ISBN 978-80-7204-591-4 [publication in Czech] 5. Lympertos EM, Dermatas ES (2007) Acoustic emission source location in dispersive media. Signal Process 87(12):3218–3225 6. Mahmoud H, Mazal P, Vlasic F (2020) Relationship between acoustic emission signal and loads on pneumatic cylinders. Nondestr Testing Eval 35(2):222–238 7. Mba D (2003) Acoustic emissions and monitoring bearing health. Tribol Trans 46(3):447–451 8. Nohál L et al (2013) An experimental investigation of rolling contact fatigue of steels using acoustic emission method. Insight Non-Destr Testing Condition Monit 55(12):665–669 9. Rahman MZ et al (2008) A study on incipient damage monitoring in rolling contact fatigue process using acoustic emission. Tribol Trans 51(5):543–551 10. Schaeffler M (2015) Toroidal roller bearings. Available at: https://medias.schaeffler.com/med ias/en!hp.info/C..-XL-K#ST4_21019524875 11. Sharma RB, Parey A (2019) Modelling of acoustic emission generated in rolling element bearing. Appl Acoust 144:96–112 12. Zhang J (2018) Investigation of relation between fracture scale and acoustic emission timefrequency parameters in rocks. Shock Vib 2018:1–14

Fatigue Risk Analysis with Intelligent Digital Twins Based on Condition Monitoring Esko K. Juuso

Abstract Fatigue mechanisms proceeding through the formation and growing of cracks are well-known but the progress is not easy to detect with measurements during the operation. The localised structural damage is caused by repeated loading and unloading when the load exceeded certain thresholds. The effect of the loading is highly nonlinear and structures fracture suddenly when a crack reaches a critical size. This research focuses on the advanced data analysis aimed at detecting effective stress impacts by using generalised norms and intelligent stress indices based on nonlinear scaling to provide good severity indicators. Digital twin type solutions can help in detecting changes in fatigue risk analysis. Contributions of the stress are calculated in each sample time, which is taken as a fraction of the cycle time. The Wöhler curve is represented by a linguistic equation (LE) model where the stress part is represented by the intelligent stress indices. The cumulative sum of the contributions indicates the deterioration of the condition, and the simulated sums can be used to predict failure time. Scheduling the maintenance actions can be extended to avoiding risky stress levels. The generalised statistical process control (GSPC) is a feasible solution to demonstrating in real time these risky levels during the operation. The analysis is adapted to changing operating conditions by updating recursively the parameters of the scaling functions. Algorithms of the model and digital twin remain unchanged. In a rolling mill, torque measurements are collected and analysed with a combination of two norms scaled with the nonlinear scaling approach. Keywords Fatigue risk detection · Nonlinear scaling · Intelligent stress indices · Condition monitoring

E. K. Juuso (B) Control Engineering, University of Oulu, PO Box 4300, FI-90014 Oulu, Finland e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_7

83

84

E. K. Juuso

1 Introduction The fatigue analysis has a long history: the importance of cyclic stress and the concept of endurance limits have been known for almost 200 years. In 1837, Wilhelm Albert published the first fatigue test results [1]. Wöhler concluded that cyclic stress range is more important than peak stress and introduces the concept of the endurance limit. Fatigue is progressive, localised structural damage that is caused by repeated loading and unloading. The nominal maximum stress dvalues are less than the ultimate tensile stress limit and may be below the yield stress limit of the material. The fatigue life consists of crack initiation and crack propagation [2]. The mechanism proceeds through cracks formed when the load exceeds certain thresholds. Structures fracture suddenly when a crack reaches a critical size. The shape of the structure will significantly affect fatigue life; square holes or sharp corners will lead to elevated local stresses where fatigue cracks can initiate. Round holes and smooth transitions or fillets are therefore important in order to increase the fatigue strength of the structure. The effects of each stress level are taken into account in the calculations of cumulative damage from individual contributions [3, 4]. Material properties are important and can be tested by applying a regular sinusoidal stress on samples of the material by a testing machine which also counts the number of cycles to failure. ASTM International defines fatigue life, N f , as the number of stress cycles of a specified character that a specimen sustains before the failure occurs [7]. In high-cycle fatigue situations, material performance is commonly characterized by a S-N curve, also known as the Wöhler curve (Fig. 1). The magnitude

Fig. 1 S-N curves in typifying fatigue test results [5, 6]

Fatigue Risk Analysis with Intelligent Digital Twins …

85

of a cyclic stress (S) is represented against the logarithmic scale of cycles to failure (N ). These curves are material-specific [8]. Probability distributions are reasonable in data analysis and design [9]. In practice, the sequences of loads are complex, often random, including large and small loads. The rainflow-counting algorithm and histograms of cyclic stress are used to assess the safe life in these cases. The effects of each stress level are taken into account in the calculations of cumulative damage or in a crack growth equation to calculate crack increments [10]. In the Palmgren-Miner linear damage rule (LDR), individual contributions are combined by means of the algorithm which assumes that there are m different stress magnitudes in a spectrum {Si , i = 1 . . . m}, each contributing n i (Si ) cycles, and Ni (Si ) is the number of cycles to the failure of a constant stress Si . The failure occurs when the cumulative damage D=

m  n i (Si ) Ni (Si ) i=1

(1)

reaches an experimental constant Dmax . Theoretically, Dmax = 1 corresponds to failure but experimentally the limit is between 0.7 and 2.2. The rule (1) does not include the handling of interactions and sequences of stress situations. The effects of dynamic stress changes are not taken into account either [11]. Various nonlinear damage models are discussed in [12]. More than 50 fatigue damage models have been developed. Fatemi and Yang [13] categorized the reviewed theories and models into six major categories: (a) linear damage evolution and linear summation, (b) nonlinear damage curve and two-stage linearisation approaches, (c) life curve modification to account for load interactions, (d) approaches based on crack growth concepts, (e) continuum damage mechanics (CDM) models, and (f) energy-based theories. The applicability of each model varies from case to case. The LDR model dominates in applications. Advanced signal processing methods and intelligent fault diagnosis have been developed to detect different types of machine faults reliably at an early stage [14]. Generalised moments and norms include many well-known statistical features as special cases and provide compact new features capable of detecting faulty situations [15–17]. These methodologies, which have been developed in connection with the vibration signals, can be extended to various measurement signals [16], e.g. the torque sensor technology suits for rough industrial applications [18]. Intelligent methods extend the idea of dimensionless indices to nonlinear systems, also known as linguistic equation (LE) systems: the basic idea is nonlinear scaling, which was developed to extract the meanings of variables from measurement signals [19, 20]. In the present systems, the scaling functions are developed by using generalised moments and norms [21, 22] and tuned with genetic algorithms [23]. The parameters of the scaling functions can be recursively updated with data analysis: the scaling is upgraded gradually and even the initial estimates are not necessary [24].

86

E. K. Juuso

This article addresses the prediction of the fatigue risk by using intelligent stress indices (Sect. 2). Their contributions to the fatigue risk are summed in dynamic models of intelligent digital twins (Sect. 3). The fatigue risk analysis is shortly presented in two applications: a roller mill and an underground loading machine (Sect. 4). The methodology is compared with previous fatigue models in Sect. 5. Conclusions and further research are discussed in Sect. 6.

2 Stress Indices Meanings of features and index levels are essential in stress monitoring since high loads or stresses reduce the time to failure (Fig. 1). Nonlinear effects are taken into account with nonlinear scaling functions obtained from measurements and features. Resulting stress indices can be monitored with statistical process control and used in developing digital twins.

2.1 Feature Extraction The behaviour of the load or stress can be represented with features extracted from measurements. A flexible set can be obtained with generalised norms defined by  ||

τ

p M j || p

=(

τ

p M j )1/ p

=

N 1  p (x j )i N i=1

1/ p ,

(2)

where p = 0 is calculated from N values of a sample, and τ is the sample time. With a real-valued order p ∈ , this norm can be used as a central tendency value p if ||τ M j || p ∈ , i.e. x j > 0 when p < 0, and x j ≥ 0 when p > 0. The norm (2) is calculated about the origin, and it combines two trends: a strong increase caused by the power p and a decrease with the power 1/ p. Therefore, all the norms have the same dimensions as the signal x j . All the features are positive since the generalised norms of absolute values |x j | were introduced for signal analysis in [15]. New feature values are calculated at every sample time by using (2). The new value can also include several sequential samples. Several specialised features {Si , i = 1 . . . M} with different orders p ∈  can be extracted from measurement signals x j . Each feature is defined by the order p and the signal number j since several features can be extracted from a single signal in a chosen frequency range. Real order derivatives provide additional possibilities [16, 17].

Fatigue Risk Analysis with Intelligent Digital Twins …

87

2.2 Nonlinear Scaling Nonlinear mappings are used inside the feasible ranges for representing measurements and features as dimensionless indices defined in a real-valued interval [−2, 2]. The basic scaling approach presented in [20] has been improved later by improved constraint handling [23] and a skewness-based methodology for signal processing [21]. The monotonously increasing functions, which consist of two second-order polynomials, have also monotonously increasing inverse functions (Fig. 2). The scaling functions are defined in the same way for the original signals and any specialised features. Monotonously increasing scaling functions can be constructed by adjusting the centre point ci , the core [(cl )i , (ch )i ] and the support [min (Si ), max (Si )]. In the data-based solution, the value range of Si is divided into two parts by the central tendency value Si and the core area, [(cl )i , (ch )i ], is limited by the central tendency values of the lower and upper parts. The approach is based on the normalised moments generalised by replacing the expectation with the norm (2) as the central value: p

γk =

N 1  p [(Si ) j − ||τ Mi || p ]k N σik j=1

(3)

where σi is calculated about the origin, and k is a positive integer [21]. The monotonous increase is achieved with a sequential approach introduced in [23]: first define the centre point c j , then the core by choosing the ratios αi− = αi+ =

(cl )i −min (Si ) ci −(cl )i max (Si )−(ch )i (ch )i −ci

(4)

from the range [ 13 , 3], and finally calculate the support [min (x j ), max (x j )]. The norms (2) are used together with the generalised skewness, k = 3 in (3), in the datadriven approach to define the centre and corner points. The ratios (4), which are checked in all data-driven cases, are also guiding the manual construction of the

Fig. 2 Examples of feasible shapes of the scaling functions [23]

88

E. K. Juuso

scaling functions. Additional constraints are used, e.g. to introduce local linear parts that can be included if they are feasible. The nonlinear scaling methodology provides good results for the automatic generation of scaling functions. Even small faults and anomalies are detected. The approach has been tested with normal, Poisson and Weibull distributions and used in condition monitoring applications [21]. This approach is suitable for a very large set of statistical distributions [22].

2.3 Stress Indices Stress indices obtained from the scaled feature values [21] are used as indications of the severity of the load. The indices are calculated with problem-specific sample times, and variation with time is handled as uncertainty by presenting the indices as time-varying fuzzy numbers. The classification limits can also be considered fuzzy. Practical long-term tests have been performed, e.g. for diagnosing faults in bearings, in the supporting rolls of lime kilns and for the cavitation of water turbines [21]. The indices obtained at every sample time are aimed for use in the same way as the process measurements in process control. The stress indices {Si , i = 1 . . . M} are consistent with the measurement and health indices developed for condition monitoring [25]. The cavitation index is an example of a stress index: level is −2 when the stress is negligible, and levels −1, 0, 1 are analogous to the lower limits of the vibration severity ranges usable, still acceptable, not acceptable defined in the VDI 2056 [26, 27]. In the fatigue analysis, the levels start from the fatigue limit or endurance limit which is the stress level below which an infinite number of loading cycles can be applied to a material without causing fatigue failure: – – – – – –

(I S )i (I S )i (I S )i (I S )i (I S )i (I S )i

= −2: The fatigue limit or endurance limit. ∈ (−2, −1]: Polishing operation with hardly any increase in the risk level. ∈ (−1, 0]: Careful operation with a very small increase in the risk level. ∈ (0, 1]: Normal operation with small increase in the risk level. ∈ (1, 2): The risk level is increasing strongly—avoid these levels. = 2: A fatigue failure is activated.

Statistical process control (SPC) is a feasible solution for demonstrating in real time the risky levels of stress during the operation. SPC is based on continuously analysing and reducing variation in manufacturing processes [28]. Various control charts have been developed for early detection, but the standard control charts are often based on normal distributions, although non-Gaussian data need to be analysed in many cases. In the fatigue risk analysis, both process measurements and condition monitoring measurements are highly nonlinear. The statistical process control (SPC) can be extended to nonlinear and non-Gaussian data by using the generalised SPC introduced in [29]. The GSPC is suitable for a large set of statistical distributions. The

Fatigue Risk Analysis with Intelligent Digital Twins …

89

Fig. 3 Generalised statistical process control (GSPC) results for a roller mill: stress indices obtained from a difference of two indicators [29]

parameters of the scaling function provide the upper control limit U C L corresponding to (I S )i = 2. A better quality performance can be achieved if the limit is moved to U C L ∗ , i.e. I S = 1. Removing the values exceeding the level corresponding to (I S )i = 2 will then change distribution to the quality control which is the main area of the SPC. Normal control rules can be used in charts shown in Fig. 3.

3 Intelligent Digital Twins Online solutions are challenging since the fatigue risk can increase although the maximum stress values are kept much less than the ultimate tensile stress limits. The structures have significant effects, and cracks can activate in various locations. Therefore, data-based solutions are needed. Digital twins are aimed at calculating the cumulative damage with time. The feature values corresponding to different stress levels may also decrease with increasing cumulative damage or wearing.

3.1 Fatigue Model In the fatigue model, the damage contributions are calculated at every sample time from the stress indices (Sect. 2.3). Wöhler curves were in [9] represented by a linguistic equation (5) I S = log10 (NC ) where the stress index I S can be a scaled value of stress, f −1 (Si ), or a scaled value of a generalised norm obtained from signals: f −1 (||τ Mαp ||). The scaling of the logarithmic

90

E. K. Juuso

values of the number of cycles, NC , is linear. As the LE model is nonlinear, it covers a wide operating range. The system could also contain several specific equations corresponding to different operating points, e.g. low, normal and high stress. The Wöhler curves can be generated from material tests. For the existing curves, the scaling functions of the stress are generated by defining the corner points from the selected points (S, NC ). Then the corner points are modified if the limits of the shape factors α−j and α+j are violated. For process equipment, the S-N curves are gradually refined, as extensive tests cannot be performed in the same way as for materials. The approach is similar to the one used in recursive modelling for prognostics [24].

3.2 Dynamic Model The continuous linguistic equation model (5) extends the principle of the PalmgrenMiner linear damage hypothesis (1). For each sample time, the cycles NC (k) obtained from I S (k) by (5), and the resulting contribution NCτ(k) are summarised to the previous contributions: τ D(k) = D(k − 1) + , (6) NC (k) which can also be used for predictions based on use scenarios. Since the stress is not constant for the whole cycle, the sample time is taken as a fraction of the cycle time. The previous history can be updated whenever the scaling functions are changed [9]. The value range of the sum D is scaled to provide the fatigue risk in percentages. The cumulative sum of the contributions presented by (6) indicates the deterioration of the condition, and the simulated sums can be used to predict the failure time. The high-stress contributions dominate in the summation. Correspondingly, the very low-stress periods have a negligible effect, which is consistent with the idea of an infinite lifetime (Fig. 1). The summation of the contributions also reveals repeated loading and unloading, and the index I S indicates the severity of the effect.

3.3 Recursive Updating The fatigue model (5) uses stress indices defined by five corner points {min (Si ), (cl )i , ci , (ch )i , max (Si )},

(7)

which correspond to specific generalised norms. The recursive updating of these norms facilitates the adaptation of the scaling functions to changing operating conditions. The parameters of the nonlinear scaling functions can be recursively updated by including new equal-sized sub-blocks in calculations. The number of samples can be increased or fixed with some forgetting, and the weighting of the individual sam-

Fatigue Risk Analysis with Intelligent Digital Twins …

91

ples can be used in the analysis. If the definitions should cover all the operating areas, also suspicious values are included as extensions of the support area. In each adaptation step, the acceptable ranges of the shape factor α−j and α+j defined by (4) are checked and corrected if needed. The analysis has two levels: the parameters of the scaling functions and the corresponding orders of the norms [30].

4 Fatigue Risk Analysis The prediction of the fatigue risk has been tested by using torque measurements from a hot rolling mill. Rolling mills are heavily loaded process equipment, where the monitoring of critical components becomes more and more important when loads are increased. For expanding the feasible operating time, the key point is to maintain or upgrade equipment as late as possible, but as early as necessary [18, 31].

4.1 Measurements In rolling mills, the whole power required for material forming is transmitted via drive trains to the rolls which means that the main drives are highly dynamically loaded. Torque measured directly at spindles or motor shafts provides a robust basis for the efficient monitoring in the rough operating conditions. The current signal from the motor offers only limited information, especially in terms of signal dynamics [18]. The huge measurement material requires feature extraction. Already the data preparation was a very time-consuming phase in this project where the torque measurements were analysed from 35,000 passes [31].

4.2 Stress Indices The torque measurements indicate very well the progress of the fatigue risk, e.g. very high torques are detected in immediate failure situations. Several norms and their combinations can be used as alternative features. In [11], the feature selection was based on the information theory: the best similarity between signals is achieved for a feature obtained as a difference between the effective and average values, i.e.  Feature =

N 1  (x j )i2 N i=1

1/2

N 1  − (x j )i N i=1

(8)

92

E. K. Juuso

where x j is the fillet split. The time interval τ can be different for the passes. Since the orders of the norm are here 1 and 2, also negative values of x j can be used. The resulting features are shown in Fig. 5. The scaling functions are highly nonlinear: feature values corresponding to stress levels −2 and −1 are very small and then values rise very fast for levels 1 and 2. The increase is even faster than in Fig. 2. The support area can be changed to ensure that the scaling function is monotonously increasing. All the feature values are positive: the negative value corresponding to the very low level is needed for function definitions. The shape of the scaling function is consistent with S-N curves shown in Fig. 1. The stress monitoring can be done with the generalised statistical process control (GSPC): Fig. 3 represents this analysis for a short period of passes.

4.3 Digital Twin The stress indices I S obtained from the features presented in Fig. 5 fit well in a linear model (Fig. 4) where the stress indices are based on the nonlinear scaling and the life cycles are presented in a logarithmic scale, i.e. the scaling function shown in Fig. 2 extracts successfully the nonlinear effects. A large number of passes have low-stress indices resulting in long periods of operation without failures (Fig. 6). The high-stress cases are seen as a very steep rise in the semilogarithmic curve which supports the previous idea to count only the number of overloads. New failures can also happen quite fast after each other. Counting the high loads is not sufficient since the fatigue risk is increasing also within normal operating conditions. An approach for this was introduced in [9]: the contribution of each pass is obtained by the continuous model (5) and summarised to the previous contributions by (6). The risk is increasing fast when the stress index is high, but the increase can also be very slow for a quite long time (Fig. 6) if the stress is kept at moderate levels. At a risk level higher than 60%, a single high torque level can have a strong effect on the activation of a failure.

Fig. 4 IS-N curve for a hot rolling mill [11]

Fatigue Risk Analysis with Intelligent Digital Twins …

93

0.14

Feature value (kNm)

0.12 0.1 0.08 0.06 0.04 0.02 0 0

0.5

1

1.5

2 Pass (n)

2.5

3

3.5 4

x 10

Fig. 5 Feature = fillet split effective − fillet split average (kNm) [11]

Fig. 6 Calculated fatigue risk (%): O is a failure point and  a pass with high torque, which does not cause a failure [11]

The fatigue risk should be calculated separately for each work roll resulting in a classification plot which refines the idea. The number of overloads is replaced by a quantitative, tunable fatigue risk calculated by (6). Torque measurement technology has been discussed in [18]. Intelligent stress indices based on nonlinear scaling were introduced to fatigue detection in [9]. The Wöhler curve is represented by a linguistic equation (LE) model, where the stress index can be a scaled value of stress or a scaled value of a generalised norm obtained

94

E. K. Juuso

from vibration signals. Torque measurements are informative in fatigue prediction [11]. The monitoring of the rolling mill main drives requires torque to be measured directly at spindles or motor shafts since the main drives are highly dynamically loaded, which affect the product and the residual lifetime of drive components.

5 Discussion The S-N curves relate the different stress magnitudes and the cycles to the fracture. The fatigue risk is accumulating from these contributions during the operation. Material properties can be analysed, but this is not sufficient for complex structures where the stress has strong local variations. Measuring the progress of the fatigue risk is usually impossible and even periodic measurements are expensive and timeconsuming. The crack initiation and crack propagation are caused by complex loading and unloading sequences. In the Palmgren-Miner hypothesis, the cumulative damage (1) is based on discretised calculation which assumes a spectrum of stress magnitudes with specific contributions. Load levels are independent without interactions and their sequence is not taken into account. The fatigue model (5) is based on the nonlinear scaling of the stress values which makes the interaction curve linear (Fig. 4). The spectrum of the stress indices is continuous within [−2, 2]. Interactions and sequences of different damage values D(k) can be taken into account by extending the dynamic model (6). The fatigue models, stress indices, digital twins and the generalised statistical process control (GSPC) operate in the same way for all the scaled values. Modifications of the S-N curve can be done by updating the scaling function of the stress. Two-stage damage theories can be built by using different scaling functions for crack initiation and propagation, respectively. However, dynamic changes can be introduced at any time if the condition of the object is seen to deteriorate. This is done by moving the function in Fig. 2 towards lower feature values. The higher the capital asset of the equipment and the more disastrous the potential consequences of a failure are, the higher the necessity for condition monitoring. Typical areas of application are drive spindles, gear boxes, roll housings and electrical motors. The fatigue risk is naturally based on the stress analysed during all different operating conditions. The condition of the object is monitored within specific operating conditions. Suitable feature extraction solution depends on the application area (Sect. 2.1). In the application case, the torque monitoring enables condition-based maintenance of the main drive spindles to optimise rolling schedules and minimise the risk of overloads. Overload risk rises and the service factor of equipment is reduced [31].

Fatigue Risk Analysis with Intelligent Digital Twins …

95

6 Conclusion The Wöhler curves represented by linguistic equation (LE) models are feasible in calculating the contributions of a complex load that varies with time. In the case study, torque measurements were collected from a rolling mill and analysed with a combination of two norms and scaled with the nonlinear scaling approach. The fatigue analysis changes into a compact parametric solution where the stress index is linked to the fatigue by a linguistic S-N curve, which is linear. The high-stress cases are seen as a very steep rise in the semilogarithmic curve. At a risk level higher than 60%, a single high torque level can have a strong effect on the activation of a failure. Long operating periods can be achieved if the risk levels are low. In the digital twins, the parametric solutions can be recursively updated at any time. They are not just for scheduling the maintenance actions: the generalised SPC approach is a feasible solution to demonstrate the risky stress levels during the operation.

References 1. Schütz W (1996) A history of fatigue. Eng Fract Mech 54(2):263–300 2. Schütz W (1979) The prediction of fatigue life in the crack initiation and propagation stages—a state of the art survey. Eng Fract Mech 11(2):405–421. https://www.sciencedirect.com/science/ article/pii/0013794479900158 3. Palmgren A (1924) Die Lebensdauer von Kugellagern. Verfahrenstechnik 68:339–341 4. Miner MA (1945) Cumulative damage in fatigue. ASME J Appl Mech 67:159–164 5. Bathia C (1999) There is no infinite fatigue life in metallic materials. Fatigue Fract Eng Mater Struct 22(7):559–565. https://doi.org/10.1046/j.1460-2695.1999.00183.x 6. Boyer HE (ed) Atlas of fatigue curves. ASM International, 518 pp 7. Stephens RI, Fuchs HO (2001) Metal fatigue in engineering, 2nd edn. Wiley, New York 8. Marines I, Bin X, Bathia C (2003) An understanding of very high cycle fatigue of metals. Int J Fatigue 25:1101–1107. https://doi.org/10.1016/S0142-1123(03)00147-6 9. Juuso E, Lahdelma S (2012) Intelligent stress indices in fatigue detection. In: 9th International conference on condition monitoring and machinery failure prevention technologies, CM 2012—MFPT 2012, 12–14 June 2012, London, UK, vol 2. BINDT, pp 654–664. https://www. proceedings.com/16283.html 10. Sunder R, Seetharam SA, Bhaskaran TA (1984) Cycle counting for fatigue crack growth analysis. Int J Fatigue 6(3):147–156. https://doi.org/10.1016/0142-1123(84)90032-X 11. Juuso E, Ruusunen M (2013) Fatigue prediction with intelligent stress indices based on torque measurements in a rolling mill. In: 10th International conference on condition monitoring and machinery failure prevention technologies, CM 2013—MFPT 2013, 18–20 June 2013, Krakow, Poland. vol 1, pp 460–471. https://www.proceedings.com/21456.html 12. Hectors K, Waele WD (2021) Cumulative damage and life prediction models for high-cycle fatigue of metals: a review. Metals 11:1–32. https://doi.org/10.3390/met11020204 13. Fatemi A, Yang L (1998) Cumulative fatigue damage and life prediction theories: a survey of the state of the art for homogeneous materials. Int J Fatigue 20(1):9–34. https://doi.org/10. 1016/S0142-1123(97)00081-9 14. Lahdelma S, Juuso E (2007) Advanced signal processing and fault diagnosis in condition monitoring. Insight 49(12):719–725. https://doi.org/10.1784/insi.2007.49.12.719 15. Lahdelma S, Juuso E (2008) Signal processing in vibration analysis. In: 5th International conference on condition monitoring and machinery failure prevention technologies, CM 2008— MFPT 2008, 15–18 July 2008, Edinburgh, UK. BINDT, pp 867–878

96

E. K. Juuso

16. Lahdelma S, Juuso E (2011) Signal processing and feature extraction by using real order derivatives and generalised norms. Part 1: Methodology. Int J Cond Monit 1(2):46–53. https:// doi.org/10.1784/204764211798303805 17. Lahdelma S, Juuso E (2011) Signal processing and feature extraction by using real order derivatives and generalised norms. Part 2: Applications. Int J Cond Monit 1(2):54–66. https:// doi.org/10.1784/204764211798303814 18. Mackel J, Fieweger M (2010) Condition monitoring in steel industry. In: Maintenance, condition monitoring and diagnostics. POHTO Publications, Oulu, pp 26–47. ISBN 978-951-981135-2 19. Juuso EK, Leiviskä K (1992) Adaptive expert systems for metallurgical processes. IFAC Proc Vol 25(17):119–124. https://doi.org/10.1016/B978-0-08-041704-2.50027-3 20. Juuso EK (2004) Integration of intelligent systems in development of smart adaptive systems. Int J Approx Reason 35(3):307–337. https://doi.org/10.1016/j.ijar.2003.08.008 21. Juuso E, Lahdelma S (2010) Intelligent scaling of features in fault diagnosis. In: 7th International conference on condition monitoring and machinery failure prevention technologies, CM 2010—MFPT 2010, 22–24 June 2010, Stratford-upon-Avon, UK, vol 2. BINDT, pp 1358– 1372. https://www.proceedings.com/12147.html 22. Juuso E (2013) Integration of intelligent systems in development of smart adaptive systems: linguistic equation approach. PhD thesis, University of Oulu, 258 pp. http://urn.fi/urn:isbn: 9789526202891 23. Juuso EK (2009) Tuning of large-scale linguistic equation (LE) models with genetic algorithms. In: Kolehmainen M, Toivanen P, Beliczynski B (eds) Adaptive and natural computing algorithms, ICANNGA 2009. Lecture notes in computer science, vol 5495. Springer, Berlin, Heidelberg, pp 161–170. https://doi.org/10.1007/978-3-642-04921-7_17 24. Juuso EK (2015) Recursive data analysis and modelling in prognostics. In: 12th International conference on condition monitoring and machinery failure prevention technologies, CM 2015—MFPT 2015, 9–11 June 2015, Oxford, UK. BINDT, pp 560–567. https://www. proceedings.com/26860.html 25. Juuso E, Lahdelma S (2008) Intelligent condition indices in fault diagnosis. In: 5th International conference on condition monitoring and machinery failure prevention technologies, CM 2008—MFPT 2008, 15–18 July 2008, Edinburgh, UK. BINDT, pp 698–708 26. VDI (1964) VDI 2056 Beurteilungßstäbe für mechanische Schwingungen von Maschinen VDI-Richtlinien 27. Collacott RA (1977) Mechanical fault diagnosis and condition monitoring. Chapman and Hall, London 28. Oakland JS (2008) Statistical process control, 6th edn. Routledge, New York 29. Juuso EK (2015) Generalised statistical process control GSPC in stress monitoring. IFACPapersOnline 48(17):207–212. https://doi.org/10.1016/j.ifacol.2015.10.104 30. Juuso EK (2015) Recursive dynamic modelling in changing operating conditions. In: Proceedings of the 56th conference on simulation and modelling (SIMS 56), October, 7–9, 2015, Linköping University, Sweden. Linköping University Electronic Press, Linköping, pp 169–174. https://doi.org/10.3384/ecp15119169 31. Juuso EK, Ruusunen M (2015) Stress indices in fatigue prediction. In: International conference on maintenance, condition monitoring and diagnostics, and maintenance performance measurement and management, MCMD 2015 & MPMM 2015, Oulu, Finland, 30 September–1 October 2015. POHTO, Oulu, pp 89–96. ISBN 978-951-98113-7-6

Low-End Hardware in Stress Monitoring of CNC Machines Konsta Karioja , Kristóf Lajber , and Esko Juuso

Abstract Stress monitoring systems are not always included in machinery, and this might create a need for a retrofitted system. Programmable Logic Controller (PLC) units, mini PCs and other similar hardware are nowadays widely available at a reasonable price, which makes this type of hardware an interesting choice for retrofitted monitoring systems. However, the cheap price often comes with some drawbacks, e.g. relatively low computing power and low memory capacity. In this paper, we study how low-priced computers perform on tasks which might be required for vibration-based stress monitoring. In addition, we discuss a technique of vibration analysis, which could contribute to stress monitoring of Computer Numeric Control (CNC) machines in general. The target of the case study was a CNC machine for milling wood. Keywords Low-price computers · Stress monitoring · Vibration

1 Introduction The need for stress monitoring of CNC machines can emerge due to several reasons. The goal can be to predict tool wear, to detect tool defect or to prevent the misuse of

K. Karioja (B) Intelligent Machines and Systems, University of Oulu, PO Box 4200, FI-90014 Oulu, Finland e-mail: [email protected] K. Lajber Faculty of Informatics, Savaria Institute of Technology, ELTE, Eötvös Loránd University, Budapest, Hungary E. Juuso Control Engineering, University of Oulu, PO Box 4300, FI-90014 Oulu, Finland e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 E. Juuso and D. Galar (eds.), Proceedings of the 5th International Conference on Maintenance, Condition Monitoring and Diagnostics 2021, Lecture Notes in Mechanical Engineering, https://doi.org/10.1007/978-981-99-1988-8_8

97

98

K. Karioja et al.

the machine. Especially when highly automated machining is desired, monitoring of machinery is often important. There are numerous research publications on stress monitoring of machine tools. In 2014, Lauro et al. published a review about monitoring machining processes in several ways [1], which partly overlap with the topic of this article. Ambhore et al. have published a review [2] with a rather extensive look at different methods of tool condition monitoring, which can be considered a pretty similar work to stress monitoring. Some online capable tool monitoring experiments have been conducted already in the 1990s [3–6], so it seems pretty clear that low-end devices today could perform several calculations which can be utilised in online monitoring, if suitable interfaces to required sensors are available. There are some studies about different ways to carry out work similar to ours. For instance, Madahusdana et al. have shown in [7] that artificial neural networks might be a feasible solution for determining milling tool condition based on vibration measurements. Artificial neural networks might require rather high computational power, but, e.g. Tarng et al. [8] and Hong et al. [9] have published studies on applying neural networks for similar applications as in this study already in 1996. Though in these studies accelerometers were not utilised, it is notable that even with computers that were in every respect inferior to the computers nowadays, the results of the studies were rather promising. This paper studies the possibilities of using low-end hardware in stress monitoring. Hardware-oriented solutions are discussed in Sect. 2. Monitoring methodologies are based on spectral norms (Sect. 3). Analyses of results in CNC machines are presented in Sect. 4 and extended to computational tests in Sect. 5. Conclusions are drawn in Sect. 6.

2 Utilising Low-End Hardware on Stress Monitoring Low-end computers can perform most calculations needed for vibration analysis. At least, if the time domain signal does not have to be very long, even a very cheap computer is suitable, though the calculation process might be rather slow. In condition monitoring, this is rarely a problem, because a delay of minutes or even hours can often be allowed, and continuous measurements are rarely obligatory. In stress monitoring the issue may be different. Of course, if one is looking for information, which helps to determine cumulative stress over a longer period of time, stress monitoring could be conducted despite the extensive delay. However, our goal here was somewhat different, because the CNC machine was located in a fabrication laboratory (Fab Lab), where the machine operators are not always very skilled. To prevent the misuse of the milling machine, the online monitoring system should warn the user rapidly when stress limits are exceeded, and thus the allowed delay is only a few seconds. In the case of severe overload, the monitoring system should even be able to send an emergency shutdown signal to the control system, and naturally in this case the delay should be minimal.

Low-End Hardware in Stress Monitoring of CNC Machines

99

In this study, we aimed to use low-end devices to determine whether or not their computational power is adequate for online stress monitoring. We consider at least three of the tested devices potentially applicable, and the tests are discussed in the following.

2.1 Single Board Computers Due to enormous recent development, a so-called single board computer (SBC) can be considered an interesting choice here. A very well-known device in this branch is the Raspberry Pi, which we have utilised in our study as well. These devices run some kind of operating system on them, which makes use of some peripherals easier. In addition, this hardware usually has a bit more computing power, adequate for rather complicated signal processing tasks. Nonetheless, the architecture of SBCs is not the most suitable for real-time or timing-critical applications. For this type of use, an external data acquisition (DAQ) device is required to access the signal from a sensor. In some cases, this might even be a show-stopper, because DAQ suppliers do not always provide driver software compatible with SBC type computers. The benefit of SBCs is that they are a quite inexpensive solution while their computing power is often relatively high. Nevertheless, the total price may not be that cheap because of the required external DAQ modules. The software development for a SBC is still a lot less laborious than for some kind of hardware-oriented device, if one can find a way to perform bus communication to the DAQ device.

2.2 Industrial Computers Another concept, which is possibly the easiest one to apply, are the high-end industrial solutions, which in our study was represented by industrial PC (IPC) from Beckhoff. These types of DAQ and automation systems were often designed for measurement purposes, but not necessarily for vibration measurements. Many of these systems have vast support for different sensor types. The benefit of these is that they are an “off the shelf” solution. On the other hand, this often means a relatively high price. Moreover, some trouble can be caused by the fact that these are often closed systems and do not support the implementation of experimental or specially designed signal processing solutions. To deploy such functionality, external computers should be used. These can be anything from a single board computer to a huge cloud computing service. However, the Beckhoff IPC tested here is to our knowledge rather freely programmable, which is one of the reasons this device was acquired for this purpose.

100

K. Karioja et al.

2.3 Hardware-Oriented Solutions The third concept to be considered is hardware-oriented solutions, like microcontrollers or other devices which generally do not require an operating system. However, this is not always the case, because some microcontrollers run a real-time operating system. In this case, the solution is hardly ready to be implemented and special hardware functions like Direct Memory Access (DMA), and special digital signal processing modules could be necessary. The benefit of this concept is that really cheap hardware can be used, and the solution can still be robust. The drawback of options of this kind is that they require a lot more development work not only for software but for hardware as well. In addition, the computing power of the microcontroller type devices is often very limited, and their built-in Analog to Digital Converters (ADCs) usually have rather low resolution. Of course, an external DAQ module could be used, but in this type of setup the computational load due to bus communication might cause problems. At this point, special digital signal processors and Field-Programmable Gate Array (FPGA) solutions should be mentioned as interesting options for these kinds of tasks. This type of hardware along with accelerometers based on MicroElectroMechanical Systems (MEMS) could give a solution with cheap hardware, but deploying the devices can be very laborious.

3 Stress Monitoring and Vibration Measurements Different ways of stress monitoring via vibration are discussed, e.g. in [10–12]. Vibration is often fairly straightforward to measure, which can make it worth considering when aiming to stress monitoring. When measuring mechanical stress, other measuring strain, cutting force or torque might often be interesting [1], but retrofitting these measurements to CNC machines can be difficult. Especially if vibration monitoring is done with relatively cheap IoT-capable hardware, the low price and the less laborious data acquisition become notable advantages.

3.1 Computational Spectral Analysis In vibration monitoring, spectral analysis is a very common method. Online monitoring is often based on calculating some features, but spectral monitoring using calculated features is not always considered. There are several studies which have shown that spectral analysis is possible this way [13–16], but it does seem to be common to use this type of methodology for stress monitoring, or at least we are not aware of this type of applications.

Low-End Hardware in Stress Monitoring of CNC Machines

101

In the case presented here, we have applied the concept of spectral norms, which can be written as follows: X (α)  p,w = (

N −1 

1

wk |X k (α) | p ) p ,

(1)

k=0

where X k(α) is a complex number resulting from the Discrete Fourier Transform (DFT) of the vibration signal and α stands for the order of derivative, p is the order of norm and wk is a weighting factor. The spectral norms can be used to calculate Spectral Norm Index (SNI), which is defined by X (α)  pm 1 bαm . (α) M m=1 X ref  pm M

,...,α M SNIαp11,,αp22 ,..., pM =

(2)

(α) In this case, bαm is a weighting factor and X ref is a norm obtained from the reference case. This is to say that the SNI is a ratio of the generalised spectral norms, where a value of a norm is compared with a reference case. More precise discussion on this concept can be found in [16, 17].

4 Test Case In our tests, the target was a milling machine for wood working located at the University of Oulu. The need for monitoring was considered to be relatively high because the machine is used in Fab Lab, where operators vary and even a short break in supervision could potentially risk misuse and damaging the machine. In this case, we used signals measured using the Beckhoff ELM3604-0002 measurement device and KS81B accelerometer manufactured by MMF. The sensitivity of the sensor was determined using a handheld shaker from PCB. In this study, we present results on four different features. The feature calculated was always root-mean-squared (RMS), but different orders of derivatives were utilised in different cases, and in one case the RMS was calculated to the envelope spectrum, in other words, it was a spectral norm [16]. The filtering was performed using an ideal filter in the frequency domain, and the derivation or integration was performed in the frequency domain as well. In the test run analysed here, a 6 mm end milling tool was used. The testing sequence consisted of a short CNC code, where a similar machining track was followed four times in a row in the same workpiece, but the depth of the cut was one millimetre deeper each time, and the depth exceeded the safety limit set by fab lab personnel on the third time. This is to say, we have one cut which should exceed the action limit, one cut which is on the top limit and two which should trigger a warning, or perhaps, an alert.

102

K. Karioja et al.

4.1 Analysis Relative differences of a set of features were calculated for three signals: velocity, acceleration and jerk, which is the derivative of acceleration with respect to time. The corresponding indices, SNI12 , SNI22 and SNI32 , are obtained for the amplitude spectrum. A general form of the spectral norm indices is given in (2). For the envelope spectrum of acceleration, the index was calculated to frequency ranges around the rotational frequency f and its harmonics up to 5 f . For this feature, we have used notation SNI22ENV . The relative change index of the root-mean-square of velocity, denoted by SNI12 , is shown in Fig. 1. The RMS of velocity is a quite common feature used in vibration monitoring, when the severity of the overall vibration is evaluated. It can be seen that in this case the feature is obviously not sensitive enough for the change in signal. It even seems that this feature cannot detect the time periods when the machine is actually cutting wood. The relative change indices of the root-mean-square of the acceleration and jerk signals are shown in Figs. 2 and 3, respectively. As can be seen, the higher order derivative produces better results in this frequency range. In the indices calculated from the acceleration signal, the difference between different cases can be seen, but the difference is slightly more distinguishable in the indices of the jerk signal. In addition, the difference between time periods when the actual milling takes place and when the machine runs idle is a lot more clear in the case of the jerk signal. (3) is summed to an index of envelope spectrum In Fig. 4, the relative change of xrms of acceleration. The feature calculated to envelope spectrum is rms as well, and the band pass used to create the envelope is from 2300 to 4000 Hz. Differences between indices are quite clear. The RMS of velocity seems almost useless in this case, but RMS might be useful when the order of the derivative is

Fig. 1 Cut getting deeper, SNI12 from frequency range 10 to 1000 Hz

1.4 1.2

SNI 12

1 0.8 0.6 0.4 0.2 0

0

50

100

Time [s]

150

200

250

Low-End Hardware in Stress Monitoring of CNC Machines Fig. 2 Cut getting deeper, SNI22 10 to 1000 Hz

103

1.4 1.2

SNI 22

1 0.8 0.6 0.4 0.2 0

0

50

100

150

200

250

200

250

200

250

Time [s]

Fig. 3 Cut getting deeper, 6 mm tool, SNI32 10 to 1000 Hz

1.4 1.2

SNI 32

1 0.8 0.6 0.4 0.2 0

0

50

100

150

Time [s]

Fig. 4 Cut getting deeper, 1 1 3 2 2 SNI2 + 2 SNI2ENV , frequency range 10 to 1000 Hz SNI32 and 2300 to 4000 Hz for enveloping

1.4 1.2

SNI 23 +SNI 22ENV

1 0.8 0.6 0.4 0.2 0

0

50

100

Time [s]

150

104

K. Karioja et al.

(3) higher, and the sum of xrms and the envelope index shows perhaps the most promising results. The envelope spectrum carries information about cyclostationary vibration with certain cyclic frequencies, thus the goal of using an envelope index here was to determine the change in the stress of the machine imposed by the milling process. This is to say the index is meant to be sensitive to vibration imposed by the spindle rotation and the cutting process, which is expected to cause vibration at a frequency which is the spindle rotational frequency, and this frequency is multiplied by the number of teeth in the milling tool. Producing the best results shown here does not seem to require an unacceptable amount of computing power, as it is discussed in Sect. 5.

5 Computation Time Tests For testing the algorithm, the code tested in Matlab was exported to C++ language to be adapted for smaller devices. Our choice was C++ because it is still one of the most commonly used programming languages, and is widely used in hardware like microcontrollers which have high potential in this kind of applications. The code originally written in Matlab was converted to C++ code using the Matlab Coder. To be able to compile and port it to different devices, the source code export was selected instead of a library. This solution might sound quite complicated, but actually rather straightforward. Some additional work was required, e.g. to make Matlab Coder to detect the input variable types correctly and to enable the resulting code to operate with different signal lengths. We had to use application programming interface (API) functions to wrap external data in a proper format, but this function was provided by the Matlab Coder. We noted that a lack of error handling, and feedback on the exported code can make the debugging process quite time-consuming. To measure the time required for the calculation, some functions had to be added to the program. A console application was written for reading the data from a file. After loading the data into the memory of the device, it is wrapped in the required structure using the API functions. Then the calculations were performed 10 times because the average calculation time was calculated. This was considered necessary because some inconsistencies were experienced in the results during the testing period. Each time before the calculation loop was started, and after finishing it, the current time was recorded into an array. The difference between these times indicated the time required for the calculation. After all of the runs completed, the results were written to a file. This method aimed to eliminate the effect from other hardware parameters such as the read-write speed of the used storage medium (Fig. 5). These tests were conducted on different devices from three different categories. The laptops were used as a reference. An industrial PC from Beckhoff and three single board computers were the main subject of these tests. The devices are listed along with their Central Processing Units (CPUs) and basic specifications in Table 1.

Low-End Hardware in Stress Monitoring of CNC Machines

Read data from file

Stop time measurement save result to array

105

Wrap to matlab specific format

Feature calculation

Start time measurement

i = 1; i++; i == 10

Write results to file

Fig. 5 Flowchart of the test program Table 1 Tested devices Device type Device Laptop Laptop IPC SBC SBC SBC

HP EliteBook Lenovo Thinkpad T580 Beckhoff C6015-0010 Raspberry Pi 3 Model B Raspberry Pi 1 Model B BeagleBoard Black

CPU (No. of cores) RAM capacity

Clock frequency

i7-7600U (4) i5-8350U (4)

16 GiB 16 GiB

max. 3.9 GHz max. 3.6 GHz

Intel Atom E3845 (4) Broadcom BCM2837 (4) ARM11 76JZF-S (1) AM335x ARM Cortex-A8 (1)

4 GiB

1.91 GHz

1 GiB

1.2 GHz

512 MiB

700 MHz

512 MiB

1GHz

The test programs are executed on all of the mentioned devices. The calculation times are represented in Table 2. These calculation times lead to conclusion that everyday laptops are vastly overpowered for the task, which was an expected outcome. The signal processing on these always took less than 0.2 s. This means they are capable of handling the data amount. These tests were conducted to obtain some kind of a reference. The Beckhoff industrial PC performed the best of the following candidates. Originally, it was selected for special experimental applications, so it was expected to perform well in this test. The results from this device are good; even if the computation time was clearly longer than with the tested laptops, the computation time of 1.2 s for 48 s test signals can be considered very good results. In the single board computer category, the results were promising as well, even if the performance of low-end SBC devices was considerably weaker in most of the cases than the IPC. We tested two versions of a well-known SBC, the Raspberry Pi,

106 Table 2 Running times Device HP EliteBook Lenovo Thinkpad T580 Beckhoff C6015-0010 Raspberry Pi 3 Model B BeagleBoard Black Raspberry Pi 1 Model B

K. Karioja et al.

Running time [s]