Smart Sensors Measurement and Instrumentation: Select Proceedings of CISCON 2021 9811969124, 9789811969126

Citation preview

Lecture Notes in Electrical Engineering 957

Shreesha Chokkadi Rajib Bandyopadhyay   Editors

Smart Sensors Measurement and Instrumentation Select Proceedings of CISCON 2021

Lecture Notes in Electrical Engineering Volume 957

Series Editors Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli Federico II, Naples, Italy Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán, Mexico Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology, Karlsruhe, Germany Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China Gianluigi Ferrari, Università di Parma, Parma, Italy Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid, Madrid, Spain Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität München, Munich, Germany Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA, USA Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt Torsten Kroeger, Stanford University, Stanford, CA, USA Yong Li, Hunan University, Changsha, Hunan, China Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University, Palmerston North, Manawatu-Wanganui, New Zealand Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan Luca Oneto, Department of Informatics, BioEngineering, Robotics and Systems Engineering, University of Genova, Genova, Genova, Italy Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University, Singapore, Singapore Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China Walter Zamboni, DIEM—Università degli studi di Salerno, Fisciano, Salerno, Italy Junjie James Zhang, Charlotte, NC, USA

The book series Lecture Notes in Electrical Engineering (LNEE) publishes the latest developments in Electrical Engineering—quickly, informally and in high quality. While original research reported in proceedings and monographs has traditionally formed the core of LNEE, we also encourage authors to submit books devoted to supporting student education and professional training in the various fields and applications areas of electrical engineering. The series cover classical and emerging topics concerning: • • • • • • • • • • • •

Communication Engineering, Information Theory and Networks Electronics Engineering and Microelectronics Signal, Image and Speech Processing Wireless and Mobile Communication Circuits and Systems Energy Systems, Power Electronics and Electrical Machines Electro-optical Engineering Instrumentation Engineering Avionics Engineering Control Systems Internet-of-Things and Cybersecurity Biomedical Devices, MEMS and NEMS

For general information about this book series, comments or suggestions, please contact [email protected]. To submit a proposal or request further information, please contact the Publishing Editor in your country: China Jasmine Dou, Editor ([email protected]) India, Japan, Rest of Asia Swati Meherishi, Editorial Director ([email protected]) Southeast Asia, Australia, New Zealand Ramesh Nath Premnath, Editor ([email protected]) USA, Canada Michael Luby, Senior Editor ([email protected]) All other Countries Leontina Di Cecco, Senior Editor ([email protected]) ** This series is indexed by EI Compendex and Scopus databases. **

Shreesha Chokkadi · Rajib Bandyopadhyay Editors

Smart Sensors Measurement and Instrumentation Select Proceedings of CISCON 2021

Editors Shreesha Chokkadi Department of Instrumentation and Control Engineering Manipal Institute of Technology Manipal Academy of Higher Education Manipal, India

Rajib Bandyopadhyay Department of Instrumentation and Electronics Engineering Jadavpur University Kolkata, India

ISSN 1876-1100 ISSN 1876-1119 (electronic) Lecture Notes in Electrical Engineering ISBN 978-981-19-6912-6 ISBN 978-981-19-6913-3 (eBook) https://doi.org/10.1007/978-981-19-6913-3 © 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 by Prof. Rajib Bandyopadhyay

The Control Instrumentation System Conference (CISCON) 2021 was held during 26–27 November 2021. This conference was for the eighteenth time organized by the Department of Instrumentation and Control Engineering Department, Manipal Institute of Technology. Due to the pandemic, it was organized in virtual mode. This year it was arranged in partnership with the International Society of Automation (ISA), Education and Research Division and Bangalore Section and with financial support from the Board of Nuclear studies and the Institution of Engineers, Mangalore. Around 120 manuscripts were received, and the acceptance rate was 25%. Four plenary lectures were delivered. While on the first day, Prof. Soumyo Mukherji of IIT Bombay and Prof. Rajib Bandyopadhyay the undersigned delivered the keynote addresses, Prof. Russel Rhinehart of the Okalama State University, USA, Dr. Jayesh Barve, Principal, General Electric, Bengaluru, and Prof. Karabi Biswas of IIT, Kharagpur, were invited to present the keynote addresses on the second day. There were four panel discussions on contemporary topics participated by eminent industry experts, academicians, and doctors. There were thirty-four (34) oral presentations. The papers were categorized into four groups—(a) Signals and Computational Techniques, (b) Automation, Robotics and Machines, (c) Electronics and Instrumentation and (d) Control Systems and were presented in three sessions. The proceedings of CISCON 2021 is an electronic publication for the accepted and presented papers at the conference. All the papers submitted went through a double-blind review process by two or three reviewers prior to being accepted to the conference. On the whole, the eProceedings presents a comprehensive overview of ongoing research in the field of instrumentation, control and automation in India and abroad and the current interests and research directions of the instrumentation and control community. I hope that we will meet at the next CISCON conference in 2022 in physical mode. Kolkata, India

Prof. Rajib Bandyopadhyay

v

Preface by Prof. Shreesha Chokkadi

I am very glad that CISCON 2021 held during 26–27 November 2021, XVIII in the series, is a grand success. Started in 2004 when the first batch of Instrumentation and Control Engineering undergraduates of MIT Manipal are in the verge of entering final year of engineering, with an objective to give them an opportunity to explore about trends in instrumentation, control and systems engineering, this annual event has grown year after year with support from MAHE management, industry and research organizations of the country like ISRO, DRDO, VSSC, BRNS, DST, CSIR, etc. Since 2020 due to the pandemic, we conducted this event as a virtual conference. In spite of this, the response from researchers across the country is tremendous, and this year, we also had an international contribution as well as invited talk. With the proceedings of the presentation being published as Scopus-indexed Lecture Notes in Electrical Engineering by Springer Nature, the value of the presentations has been enhanced. The quality of the conference can be measured with the fact that only 34 of the submitted 128 articles have been accepted for presentation. This years’ conference had representation from both academia with professors from reputed IITB, IITKGP, NIT Jadavpur, and experts from industry like GE presenting invited talks. A feather on cap of this version of CISCON is two debates where the panellists were industry experts from Honeywell, GE and Mylabs Discovery Solutions along with experts from academia like IITB, NITJ and MIT Manipal, wherein deliberations on contemporary issues like sensor design for biomedical applications and relevance of AIML in Electronics and Instrumentation curriculum were very critically introspected. Overall, all deliberations and presentations are very informative and provided very good input to all participants especially to our faculty and students. I congratulate and acknowledge all participants, resource persons, conveners for taking out their

vii

viii

Preface by Prof. Shreesha Chokkadi

time and expertize to organize and make the event a very successful one. I wish this annual event of Department of ICE, MIT Manipal, would leave up to the expectations of all key stakeholders and grow as a sought-after multidisciplinary conference in contemporary domains. Manipal, India

Prof. Shreesha Chokkadi

About the Conference

Control Instrumentation System (CISCON) is the annual conference event organized by the Department of Instrumentation and Control Engineering, Manipal Institute of Technology. The department initiated CISCON in the year 2004 to provide a platform for its first batch of B.E. in Instrumentation and Control Engineering students to have interaction and exchange of ideas with their counterparts in and outside the institution. This is the first of its kind in the institute and under the able leadership of Dr. V. I. George. With very few institutes in the country offering this specialized interdisciplinary course, people working in both Instrumentation and Control Engineering sought after for this conference every year and have gained lots of recognition. The conference has been sponsored by national research organizations like Defence Research and Development Organization (DRDO), Board of Research in Nuclear Sciences (BRNS), Indian Space Research Organization (ISRO), and Council of Scientific and Industrial Research (CSIR) to name a few. The proceedings of CISCON has been brought out regularly since its inception. In 2015, it was decided to bring out the published papers in Scopus-indexed journals to give additional incentive to authors who put forward their research articles to CISCON, and the same trend has continued till 2017 with the rapid increase in submission. Later, presented papers were published in Lecture Notes in Electrical Engineering published by Springer Nature. The conference has attracted a large number of papers in varied disciplines like process control, automation, renewable energy, robotics, image processing, sensor and instrumentation, etc. Out of the total 120 papers submitted, 85 papers were sent for double-blind review after preliminary inspection and plagiarism check. Out of these, 36 papers have been accepted and presented in the conference and would be considered for publication in this book as chapters. We believe that the proceedings of the conference will be well received by researchers working in the domain and get inspiration for budding researchers to explore more into the varied domains in which the papers are presented. The papers presented in this proceedings are mainly in the domain of process control automation, instrumentation, robotics, image processing and many more. The readers of this proceedings will get an insight into the varied areas in which contemporary research is being carried forward in this domain and get started to go ahead. These papers will ix

x

About the Conference

give openings for beginners and also the direction for those who are working in these specific domains already. We are confident that the proceedings will be accepted by prospective researchers very well and give encouragement for us to go ahead with organizing CISCON every year with many new ideas and scope. This event was made possible by the utmost support from Chancellor of MAHE Padmashree Awardee Dr. Ramadas M. Pai, Pro-Chancellor Dr. H. S. Ballal, Vice-Chancellor Lt. Gen. (Dr.) M. D. Venkatesh, Registrar Dr. Narayana Sabhahit, Section Heads of finance and other logistic services, and they deserve our heartfelt gratitude. Director of Manipal Institute of Technology Cdr. (Dr.) Anil Rana, Joint Director Dr. Somashekara Bhat and Dr. Shreesha C. and Head of the Department, Instrumentation and Control Engineering deserve lots of appreciation for their constant guidance and motivation. Our sincere gratitude to Prof. Soumyo Mukherji, Prof. Rajib Bandyopadhyay, Prof. R. Russel Rhinehart, Prof. Karabi Biswas and Dr. Jayesh Barve for sharing their knowledge and views at the conference gathering. The convener of the conference, Dr. Kapil Sadani, deserves special recognition for his several months of untiring work towards this conference and a special mention to the administrative staff of Manipal Academy of Higher Education (MAHE), Manipal Institute of Technology and also the Department of Instrumentation and Control Engineering for their wholehearted support in making the conference event. Our sincere acknowledgement to the unanimous technical reviewers, to all contributing authors for taking time and effort to send their research work and adhering to all review comments and formatting requirements. We also wish to place our gratitude to Springer Nature for accepting our request to publish the accepted/presented papers in CISCON 2021 and finally our acknowledgement for all who have directly or indirectly helped us in organizing this event successfully and bringing out these proceedings.

Contents

Muffler Transmission Loss Optimization for a Vehicle Using Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Riziyamaalisa Gavit and Kiran Wani

1

Design and Simulation of a Wireless Charging System for Electric Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nikhil Kadam and Archana Thosar

19

Recent Advances in Sensor Technology for Biomedical Applications: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Niharika Karnik, Karan Bhadri, and Pankaj Dhatrak

37

Performance Analysis of Diode Clamped and Flying Capacitor Multilevel Matrix Converter Used for DFIG-Based Wind System . . . . . . G. Pandu Ranga Reddy, D. Mahesh Kumar, K. Rajesh, Y. Chintu Sagar, and J. Nageswara Rao

59

Real Time Feedback System for Speech Dysfluency in Children . . . . . . . . Jennifer C. Saldanha and Rohan Pinto

75

Nonlinear Model-Predictive Control Using First-Principles Models . . . . R. Russell Rhinehart

93

DC Motor System Identification and Speed Control Using dSPACE Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 S. Menaka and S. Patilkulkarni Oil Quality Analysis Using Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . 129 Nivedita Daimiwal, Revati Shriram, Harish Shinde, Radhika Kulkarni, and Apeksha Galewad Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive Moment for Washing Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 S. Elavaar Kuzhali, Kotha Manvitha, Anisha Singh, Lakshmi Pranathi, Shreya Dhavule, and M. Poorvita xi

xii

Contents

System Identification, Stability Analysis and PID Controller Design for PEM Electrolyzer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Aruna Rajaiah and Jaya Christa Sargunar Thomas Sliding Mode Hybrid Control of PMSM for Electric Vehicle . . . . . . . . . . . 165 Ajay Pawar and S. V. Jadhav Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 M. Bala Abhirami and I. Thirunavukkarasu Performance Prediction of Solar Cell Using Virtual Production Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 B. Ashok Kumar, T. S. Bagavat Perumaal, S. Senthilrani, and Parthasarathy Seshadri Optimisation of FPGA-Based Designs for Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 P. L. Bonifus, Ann Mary Thomas, and Jobin K. Antony Design and Implementation of an Automated Fuel Station . . . . . . . . . . . . . 223 M. Jyothirmayi, Vibha B. Raj, V. Lekhana, and P. Manjunath Heart Disease Prediction Using Machine Learning Algorithms . . . . . . . . . 239 Rea Mammen and Arti Pawar Design and Performance Evaluation of a Simple Resistance-to-Digital Converter for Tunneling Magneto-Resistance-Based Angular Position Sensor with 180° Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 Kishor Bhaskarrao Nandapurkar The Use of LBP Features in Transform Domain for Object Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 R. Ahila Priyadharshini and S. Arivazhagan Design and Simulation of Capacitive Pressure Sensor for Monitoring Lead-Acid Battery Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 Yashwant Adhav, Dayaram Sonawane, and Chetankumar Patil Development of Screw Press-Dewatering Unit for Biogas Slurry . . . . . . . 303 Madhuri More, Chitranjan Agrawal, and Deepak Sharma The Use of Photoplethysmography for Blood Glucose Estimation by Noninvasive Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 Vandana C. Bavkar and Arundhati Shinde Single-Stage Stand-Alone Induction Motor Driven Solar Water Pumping System with Minimal Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 Anup Shetty, K. Suryanarayana, and L. V. Prabhu

Contents

xiii

Automation of Weight-Based Sorting System Using Programmable Logic Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353 P. Chenchu Saibabu, R. Anjana, Manisha Kumari, and C. R. Srinivasan Design, Development and Verification of a Fault Injection Capable Synchronous Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365 Sona Meiyappan, P. Chaithanyasai, S. Swetha, M. Vishnu Deepika, and P. V. Sunil Nag UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 K. Bipin and P. V. Sunil Nag Design and Implementation of Efficient IoT-Based Smart Oil Skimmer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 S. Rajesh Kannan, V. G. Rajagopalan, H. Ramakrishnan, S. Sibi Selvan, and Sushanth Krishnamithran Comparison of Discrete Time Sliding Manifold and Its Impact on System Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Shaktikumar R. Shiledar and Gajanan M. Malwatkar Volkswagen Emission: An Analysis on the VW Vento Using Automotive Network Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423 Suprava Sarkar and Nithin Mohan Hand Gesture-Controlled Wheeled Mobile Robot for Prospective Application as Smart Wheelchairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437 Leon Muli Suryavanshi, Ananth Jnana Chandraraj, Kshetrimayum Lochan, and Pooja Nag Manual Dexterity Assessment Using a Nine-Hole Pegboard Test . . . . . . . 449 K. Aneesha Acharya and Amartya Choudhary Implementation of Indoor Navigation Control for Two-Wheeled Self-balancing Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461 B. Vignesh, Deepa Jose, and P. Nirmal Kumar Application of NIR Spectroscopy with Chemometrics for Discrimination of Indian Black Pepper Berries . . . . . . . . . . . . . . . . . . . . 475 Arnab Giri, Dilip Sing, Sudarshana Ghosh Dastidar, Pallab Kanti Halder, Nanaocha Sharma, Pulok K. Mukherjee, and Rajib Bandyopadhyay

xiv

Contents

A Comparative Study Between Partial Least Squares and Principal Component Regression for Nondestructive Quantification of Piperine Contents in Black Pepper by Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483 Dilip Sing, Sudarshana Ghosh Dastidar, Wasim Akram, Sourav Guchhait, Shibu Narayan Jana, Subhadip Banerjee, Pulok Kumar Mukherjee, and Rajib Bandyopadhyay Power Quality Data Mining Using Hybrid Feature Extraction Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 Vidhya Sivaramakrishnan, Balaji Mahadevan, and Kamaraj Vijayarajan Low-Cost, IOT-Based Child Safety Monitoring Robot with User-Friendly Mobile App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503 Kalyan Kasturi, Rajani Dharanikota, Khaleelu Rehman, Senthilkumar Meyyappan, and Akhil Kommineni Secure Image Classification Using Deep Learning . . . . . . . . . . . . . . . . . . . . 513 K. Gururaj, Alaka Ananth, and Sachin S. Bhat

Editors and Contributors

About the Editors Prof. Shreesha Chokkadi completed his graduation with B.E. in Electrical and Electronics Engineering from Government BDT College of Engineering Davanagere in 1988. He obtained M.E. from Walchand College of Engineering Sangli in Control Systems (Electrical) during 1992. In 2002 he was awarded with Ph.D. from IIT Bombay for his thesis entitled New Approaches to Control Relevant Identification. He has published about 30 research articles in indexed International journals and more than 40 technical papers in peer reviewed conferences. He has guided 8 doctoral students with three on rolls. Professor Shreesha is member of multiple scientific professional bodies such as IEEE, ISTE, ISLE and IEI. Prof. Rajib Bandyopadhyay completed bachelors, masters and Ph.D. in Electronics from Jadavpur University in 1984, 1987 and 2001 respectively. He has been serving the department of Instrumentation and Electronics as a professor since 2001. Prof. Bandopadhyay works on electronic olfaction, gas sensing and design and testing of analytical instruments. He has successfully delivered over 13 externally funded projects, guided over 20 PhD students and published over 92 research articles in the domain of Instrumentation. Professor Bandyopadhyay is the president of ISA Kolkata chapter.

Contributors Yashwant Adhav Cummins College of Engineering for Women, Pune, India Chitranjan Agrawal Department of Mechanical Engineering, CTAE, MPUAT, Udaipur, Rajasthan, India R. Ahila Priyadharshini Mepco Schlenk Engineering College, Sivakasi, India

xv

xvi

Editors and Contributors

Wasim Akram Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India Alaka Ananth NMAM Institute of Technology, Nitte, India K. Aneesha Acharya Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India R. Anjana Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India Jobin K. Antony Department of Electronics & Communication Engineering, Rajagiri School of Engineering & Technology, Kochi, Kerala, India; APJ Abdul Kalam Technological University, Thiruvananthapuram, Kerala, India S. Arivazhagan Mepco Schlenk Engineering College, Sivakasi, India B. Ashok Kumar Department of Electrical and Electronics Engineering, Thiagarajar College of Engineering, Madurai, India T. S. Bagavat Perumaal Department of Electrical and Electronics Engineering, Thiagarajar College of Engineering, Madurai, India M. Bala Abhirami Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India Rajib Bandyopadhyay Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India Subhadip Banerjee School of Natural Product Studies, Jadavpur University, Kolkata, India Vandana C. Bavkar Department of Electronics, Bharati Vidyapeeth (Deemed to be University), College of Engineering, Pune, India Karan Bhadri Dr. Vishwanath Karad, MIT-World Peace University, Pune, India Sachin S. Bhat Shri Madhwa Vadiraja Institute of Technology and Management, Bantakal, Udupi, India K. Bipin Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India P. L. Bonifus Department of Electronics & Communication Engineering, Rajagiri School of Engineering & Technology, Kochi, Kerala, India; APJ Abdul Kalam Technological University, Thiruvananthapuram, Kerala, India P. Chaithanyasai Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India

Editors and Contributors

xvii

Ananth Jnana Chandraraj Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal, Karnataka, India P. Chenchu Saibabu Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India Amartya Choudhary Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India Nivedita Daimiwal Cummins College of Engineering for Women, Pune, India Sudarshana Ghosh Dastidar Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India Rajani Dharanikota ECE Department, Nalla Malla Reddy Engineering College, Hyderabad, India Pankaj Dhatrak Dr. Vishwanath Karad, MIT-World Peace University, Pune, India Shreya Dhavule Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India S. Elavaar Kuzhali Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India Apeksha Galewad Cummins College of Engineering for Women, Pune, India Riziyamaalisa Gavit MTech Automotive Technology, College of Engineering, Pune, India Arnab Giri Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India Sourav Guchhait Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India K. Gururaj NMAM Institute of Technology, Nitte, India Pallab Kanti Halder School of Natural Product Studies, Jadavpur University, Kolkata, India S. V. Jadhav Department of Electrical Engineering, College of Engineering, Pune, India Shibu Narayan Jana School of Natural Product Studies, Jadavpur University, Kolkata, India Deepa Jose Department of Electronics and Communication, KCG College of Technology, Chennai, Tamil Nadu, India M. Jyothirmayi Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bangalore, India

xviii

Editors and Contributors

Nikhil Kadam Department of Electrical Engineering, College of Engineering Pune, Pune, India Niharika Karnik Dr. Vishwanath Karad, MIT-World Peace University, Pune, India Kalyan Kasturi ECE Department, Nalla Malla Reddy Engineering College, Hyderabad, India Akhil Kommineni ECE Department, Nalla Malla Reddy Engineering College, Hyderabad, India Sushanth Krishnamithran Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai, India Radhika Kulkarni Cummins College of Engineering for Women, Pune, India D. Mahesh Kumar Deparment of EEE, PVKK Institute of Technology, Anantapur, India Manisha Kumari Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India V. Lekhana Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bangalore, India Kshetrimayum Lochan Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal, Karnataka, India Balaji Mahadevan Sri Sivasubramaniya Nadar College of Engineering, Chennai, India Gajanan M. Malwatkar Department of Instrumentation Engineering, Government College of Engineering, Jalgaon, India Rea Mammen ICAS, Manipal Academy of Higher Education, Manipal, India P. Manjunath Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bangalore, India Kotha Manvitha Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India Sona Meiyappan Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India S. Menaka JSS Science and Technology University, Mysuru, India Senthilkumar Meyyappan ECE Department, Nalla Malla Reddy Engineering College, Hyderabad, India Nithin Mohan Influx Big Data Solutions, Influx Technology India, Domlur, Bengaluru, Karnataka, India

Editors and Contributors

xix

Madhuri More Department of Renewable Energy Engineering, CTAE, MPUAT, Udaipur, Rajasthan, India Pulok K. Mukherjee School of Natural Product Studies, Jadavpur University, Kolkata, India; Department of Biotechnology, Institute of Bioresources and Sustainable Development, Government of India, Imphal, India Pulok Kumar Mukherjee School of Natural Product Studies, Jadavpur University, Kolkata, India; Institute of Bioresources and Sustainable Development, An Autonomous Institute under Department of Biotechnology, Government of India, Imphal, India Pooja Nag Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal, Karnataka, India Kishor Bhaskarrao Nandapurkar Indian Institute of Technology (Indian School of Mines) Dhanbad, Dhanbad, Jharkhand, India P. Nirmal Kumar Electronics and Communication Engineering, College of Engineering, Anna University, Chennai, India Chetankumar Patil Department of Instrumentation and Control Engineering, College of Engineering, Pune, India S. Patilkulkarni JSS Science and Technology University, Mysuru, India Ajay Pawar Department of Electrical Engineering, College of Engineering, Pune, India Arti Pawar ICAS, Manipal Academy of Higher Education, Manipal, India Rohan Pinto Department of Electronics and Communication, St Joseph Engineering College, Mangaluru, India; Affiliated to Visvesvaraya Technological University, Belagavi, India M. Poorvita Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India L. V. Prabhu HEXMOTO Controls Pvt. Ltd, Mysuru, India Lakshmi Pranathi Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India Vibha B. Raj Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bangalore, India V. G. Rajagopalan Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai, India Aruna Rajaiah P.S.R. Engineering College, Sivakasi, Tamil Nadu, India

xx

Editors and Contributors

K. Rajesh Deparment of EEE, RGM College of Engineering and Technology, Nandyal, India S. Rajesh Kannan Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai, India H. Ramakrishnan Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai, India J. Nageswara Rao Deparment of Electrical and Computer Engineering, MizanTepi University, Teppi, Ethiopia G. Pandu Ranga Reddy Deparment of EEE, G. Pullaiah College of Engineering and Technology (Autonomous), Kurnool, India Khaleelu Rehman ECE Department, Nalla Malla Reddy Engineering College, Hyderabad, India R. Russell Rhinehart Oklahoma State University, Stillwater, OK, USA Y. Chintu Sagar Deparment of EEE, Ashoka Women’s Engineering College, Kurnool, India Jennifer C. Saldanha Department of Electronics and Communication, St Joseph Engineering College, Mangaluru, India; Affiliated to Visvesvaraya Technological University, Belagavi, India Suprava Sarkar Department of Instrumentation and Control, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India S. Senthilrani Department of Electrical and Electronics Engineering, Velammal College of Engineering and Technology, Madurai, India Parthasarathy Seshadri Department of Electrical and Electronics Engineering, Thiagarajar College of Engineering, Madurai, India Deepak Sharma Department of Renewable Energy Engineering, CTAE, MPUAT, Udaipur, Rajasthan, India Nanaocha Sharma Department of Biotechnology, Institute of Bioresources and Sustainable Development, Government of India, Imphal, India Anup Shetty Department of Electrical and Electronics Engineering, NMAM Institute of Technology, Nitte, Karkala, Karnataka, India Shaktikumar R. Shiledar Department of Instrumentation Engineering, Government College of Engineering, Jalgaon, India Arundhati Shinde Department of Electronics, Bharati Vidyapeeth (Deemed to be University), College of Engineering, Pune, India Harish Shinde Cummins College of Engineering for Women, Pune, India Revati Shriram Cummins College of Engineering for Women, Pune, India

Editors and Contributors

xxi

S. Sibi Selvan Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai, India Dilip Sing Department of Instrumentation and Electronics Engineering, Jadavpur University, Kolkata, India Anisha Singh Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India Vidhya Sivaramakrishnan New Prince Bhavani College of Engineering and Technology, Chennai, India Dayaram Sonawane Department of Instrumentation and Control Engineering, College of Engineering, Pune, India C. R. Srinivasan Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India P. V. Sunil Nag Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India K. Suryanarayana Department of Electrical and Electronics Engineering, NMAM Institute of Technology, Nitte, Karkala, Karnataka, India Leon Muli Suryavanshi Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal, Karnataka, India S. Swetha Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India I. Thirunavukkarasu Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India Ann Mary Thomas Department of Electronics & Communication Engineering, Rajagiri School of Engineering & Technology, Kochi, Kerala, India; APJ Abdul Kalam Technological University, Thiruvananthapuram, Kerala, India Jaya Christa Sargunar Thomas Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India Archana Thosar Department of Electrical Engineering, College of Engineering Pune, Pune, India B. Vignesh Electronics and Communication Engineering, College of Engineering, Anna University, Chennai, India Kamaraj Vijayarajan Sri Sivasubramaniya Nadar College of Engineering, Chennai, India

xxii

Editors and Contributors

M. Vishnu Deepika Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India Kiran Wani ARAI ACADEMY, Chakan, Pune, India

Muffler Transmission Loss Optimization for a Vehicle Using Genetic Algorithm Riziyamaalisa Gavit

and Kiran Wani

Abstract Automotive reactive mufflers attenuate sound using interference which is produced by engine exhaust sound waves that are partially or fully canceled. Its performance depends on back pressure which depends on the structure. The indicator of the performance of the muffler is a measurement of transmission loss (TL) which depends on the geometry of the muffler. In this work, experimental as well as CAE analysis is used to find TL; for experimental, validation test is carried on impedance tube, and conical adapters are designed to accommodate different diameter openings of impedance tube and muffler. Inlet, outlet, and muffler diameter and length are optimized to maximize TL. Genetic algorithm is used for maximization of TL given by CAE analysis of variable models with change in range of − 15% to + 15% of original dimensions. With the help of this work, contribution of each part can be studied. We achieve 10.24% maximization in TL suggested by this optimized solution. The purpose of this paper is to reduce the complexity of the design. Keywords Transmission loss · Muffler · Genetic algorithm

1 Introduction Sound is a normal part of daily life, and there are lots of sounds pleasant and unpleasant, whereas unpleasant and damaging sound is called noise. It can be generated out by humans or machines. Noise and vibration are both variations in air and media pressure that affect the human body. Noise exceeding a certain value can create a lot of damage to human hearing ability [1]. Noise causes discomfort in physical or mental performance. Humans who are in direct exposure to a R. Gavit (B) MTech Automotive Technology, College of Engineering, Pune 41100, India e-mail: [email protected] K. Wani ARAI ACADEMY, Chakan, Pune 410501, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_1

1

2

R. Gavit and K. Wani

high level of noise coming out from engine working of an automobile or farm tractors are under mild but repetitive acoustic danger. Therefore, there is necessary to reduce this unwanted emitted noise using some devices structures [2]. Mufflers are sound-attenuating devices that reduce radiated noise from equipment such as engines. Sound can be affected by some external influences like pressure, temperature, and the density of the medium. There are four types of noise sources in vehicles. They are engine noise, wind noise, road noise, and exhaust noise. To keep up with government legislation [3]. The muffler needs to satisfy the compatibility of appropriate dB level means accepted minimal transmission loss, also its robustness and structural integrity. From the above, we say that searching for the best design within the restricted time for the vibroacoustic optimization applications is always an issue, so to address this issue, geometry optimization of automotive mechanical structure is used to reduce radiated sound power level by different types of optimization methods [4]. Barbieri’s work shows us combining finite element analysis and optimization methods for mufflers’ geometry design [5]. The intention is to obtain the dimensions for the acoustic muffler with the transmission loss (TL), being maximized in the desired frequency range. L.J Yeh’s paper has shown us that a genetic algorithm (GA) can be utilized in the optimization of uproar noise reduction in duct systems by adjusting the muffler geometry under area or space constraints [6]. Also, its extended work showed that parameters like crossover, mutation, and elitism techniques play pertinent roles in GA optimization. TL optimization in a double-chamber attenuator was applicable, and it is validated with calculations used in conventional gradient approaches or performing experiments in the laboratory [7]. Simulated annealing (SA) is also an alternate approach for optimization of sound transmission loss (STL) [8], but research shows that on comparing both simulated results, GA shows better acoustic performance than SA. The methodology of this paper is followed as per the work of Ranjbar in which they have investigated the effect of variation of the dimension of a component individual on overall performance; in this project, they are working on the single-chamber or regular muffler proper. The experimentation is done for changes in between 5 and 30% which is carried out by the method of three-point method which is carried out in MAP software. Later, all the results are compared, analyzed, and optimized using a genetic algorithm. From this work, we can get that inlet diameter is a critical parameter than other variables; the results show a rise in TL by 9 dB [9, 10]. TL consists of audio spectrum which is a function of frequency, so to obtain a unified global measure of noise, TL behavior is provided in the desired frequency range. The root mean square of TL over that frequency waveband is taken for optimization-independent variable [11]. In their extended work, shape optimization of multichamber muffler which carried also the effect of geometrical [12] properties of muffler components on TL optimization is studied. M. Vora checked the outcome result of muffler design on amplitude and tractor noise sound frequency. Quantitative analysis of apex amplitude-frequency observation various muffler installation at distinct engine speed (RPM), the result shows that present existence of higher frequencies in a surplus amount at engine speeds of

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

3

1000, 1500, 2000 RPM. At operator ear level, SPL observations on agricultural tractor showed lowest 930.6 Hz and highest 1502.6 Hz peak frequencies for no and standard muffler, but normal muffler peak frequency is at 1134.9 Hz. For no muffler condition at 1000 and 1500 RPM, peak frequencies generated are (1525.8 Hz, 103.7 Hz) and (665.4 Hz, 2016.5 Hz), respectively. This is taken at OEL and 10 m distance from farm tractor condition, respectively. Now, noise generated from this is (20–55 dB, 50–80 dB) and (20–70 dB, 40–90 dB), respectively [13]. So, studying the acoustic audio spectrum for the values between 930.6 and 1502.6 Hz produces more noise than other frequencies. The laboratory test shows that experimental validation for finding out TL [14] of the multichamber muffler is done using modified four-microphone impedance tube setup which is equipped with power analyzer and DAQ system, in which modified cones do not affect the final TL result.

1.1 The Targets in the Project to Be Achieved Are as Follows • Optimize the multichamber muffler lengths and diameters. • Obtain above 5% maximization of transmission loss compared to base muffler design. • Suggest design modification for the muffler. • Check the effectiveness of the use of impedance tube for experimental calculation of transmission loss for a low cost.

2 Theory 2.1 Acoustic Filter Design Parameters 1. Adequate insertion loss (attenuation) 2. Backpressure 3. Structural.

2.2 Acoustic Filter Performance Parameters 1 Transmission loss 2 Insertion loss 3 Noise reduction.

4

R. Gavit and K. Wani

Transmission loss (TL) It is defined as a power incident on a muffler and that is transmitted downstream into an exhaust outlet. TL = Lw1 − Lw2 (dB) | | | | S2( A2)2 2 | (dB) | TL = 10 ∗ log| 2 2 S1( A1) | ( / ) TL = 20 ∗ log A2 A1 (dB)

(1)

(2) (3)

where S1, S2 are areas of the exhaust pipe and A1, A2 are associated with the incident and transmitted acoustic power [2, 4].

2.3 Calculation of Noise Transmission Loss by Impedance Tube Another way to find out transmission loss at a low cost as well without full laboratory setup at a high cost is that we can use the impedance tube method to calculate TL. According to standards ASTM E2611, transmission loss of a muffler can be calculated by using an impedance tube setup. To measure the transmission loss, they use the four-microphone transfer function method [12] (Fig. 1). In the impedance tube, we can find out sound absorption coefficient and transmission loss of any material; frequency response at four microphones is calculated and imported into the transfer matrix method. Using the impedance tube, we can find TL at lower as well as higher frequencies.

Fig. 1 Diagram of transmission loss measurement setup using impedance tube

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

5

2.4 Genetic Algorithm (GA) Its ideas are like the idea of Darwin’s theory of survival of the fittest. It is normally able to find the global minimum and can neglect local minima, so it can operate at all types of discontinuities. This is an optimization method, and this algorithm shows the process of natural selection where the fittest individuals are selected for procreation to generate offspring of the latest generation, simple law of survival of the fittest. TL comes in the audio spectrum which is a function of frequency, so to obtain a unified global measure of noise, TL gives noise behavior in the desired frequency range, root mean square of TL is done, and the calculation formula is given below for RMSTL. / RMSTL =

2 ∫ ff max min TL ( f )d f

f max − f min

(4)

This method is considered to find the best optimum value of a simple muffler which generates a maximum root mean square of transmission loss (RMSTL). Crossover is a genetic guide used to mix the hereditary information of two parental chromosomes to generate the latest offspring with new properties. The mutation is manipulation as per other data. Discrete variable optimizations are used for engineering problems where the design variable is fixed from a set of discreet values. Solving discreet problems have prone and cone to as we know we have limited range of value, but the combination explosion can occur lead to take much longer time to solve the equation is simple to type of optimization where any function can be used as fitness function and specified as per objective maximization or minimization. Both methods have its advantages and disadvantages, so the combination of function-based and discreet variable optimization can give us a fast and less complex solution. Curve fitting is a method used to find possible best-fitted curves on basis of discreet data. You can use different sorts of postprocessing methods for plotting, interpolation, and extrapolation methods to estimate confidence in intervals and calculate integrals and derivatives once you have created a fit for the curve. TL comprises a spectrum, i.e., it is a function of (f ) frequency. So, to obtain some unified global measure of the noise TL behavior of a muffler in each desired frequency range, the root means the square level of noise transmission loss over that frequency waveband spectrum, known subsequently RMSTL is measured Based on software simulation is used in this project. By computing data of TL at each frequency obtained in numerical software MATLAB [13, 15].

3 Methodology See Fig. 2.

6

R. Gavit and K. Wani

Fig. 2 Workflow of research thesis

4 Simulation and Experimentation The muffler model taken for the study is Mahindra 575 Diesel Direct Injection engine. DI: 4 strokes, 4-cylinder; diesel engine: 2730 cc, 45 Hp; engine rated RPM: 1900 (Table 1). Table 1 Original dimensions of muffler

Component

Dimensions (mm)

Inlet diameter

48

Outlet diameter

42

Muffler diameter

93

Inlet pipe length

400

Outlet pipe length

272

Muffler length

503

Distance from first point to the upstream baffle plate

25

Distance between two baffle plates

75

First baffle plate hole number

6

Second baffle plate hole number

3

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

7

5 CAE Simulation Computer-aided engineering simulation is done in COMSOL Multiphysics software in pressure acoustics. Frequency-domain interface is used for computation of the pressure fluctuation for the propagation of acoustic waves in fluids in inert background conditions. It is suitable for all frequency-domain simulations with harmonic variations of the pressure field (Fig. 3). It is used in acoustics to calculate transmission loss versus frequency. The impacts of all eigenmodes that are adequately resolved by the mesh, as well as how they couple with the applied loads or excitations, are considered in a frequency-domain research. A transfer function, such as magnitude or phase of deformation, sound pressure, impedance, or scattering parameters versus frequency, is often displayed as the result of a frequency-domain analysis (Fig. 4). COMSOL recommends using a minimum of five elements per wavelength for a 3D problem, and according to the acoustic meshing criteria, six elements per wavelength are required to capture the acoustic sound wave, but there is a chance that one may still need to confirm the best-suited mesh size for your model using a mesh refinement study. • • • • •

Maximum frequency of wavelength = 1800 Hz Sound velocity = wavelength * frequency 343 = wavelength * 1800 Wavelength = 0.1905 m Element size = wavelength/6 = 0.03175 m = 31.4 mm.

From the above, better accuracy will be found nearby 31.4 mm element size, but for every model, we see that it does not apply, so we take in a range of 65–28.4 mm (Fig. 5).

Fig. 3 Three-dimensional model of muffler

8

R. Gavit and K. Wani

Fig. 4 Meshed model in COMSOL

Fig. 5 Variable used for calculation

5.1 Experimental Testing In impedance tube, we can find out sound absorption coefficient and transmission loss of any material; for practical validation, experimental results are necessary. Experimental setup To calculate transmission loss of muffler, frequency response at four microphones is calculated and imported into transfer matrix method which will be conducted automatically by LabVIEWsoftware; the DAQ system gives us the plot of transmission loss. The experiment is conducted using white noise and with the lid open and closed alternatively. A random white noise signal is provided from the speaker at the inlet end of the muffler because generally, white noise predicts a good result to evaluate transmission loss performance. At outlet, sound pressure level is measured with the use of microphones. The first three readings are taken keeping the end of the

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

9

Fig. 6 Actual experimental setup

impedance tube open, and the next three readings are taken with the end of the impedance tube closed. Further, muffler transmission loss is computed with the use of the transfer function method. Figure 6 shows actual experimental setup on which experiments are conducted. From Fig. 6, we can see that the changes in the actual setup are made. The tube used to place the samples in the impedance tube is replaced with a muffler tube with modified conical ends, so it can be easily held into the impedance tube, and the remaining setup remains the same.

6 Result and Discussion 6.1 Transmission Loss Experimental Results for Base Muffler From Table 2, we can see that the dominant frequencies lie in the range of 800– 1600 Hz; in this range, we obtain the highest transmission loss (Fig. 7).

6.2 Transmission Loss CAE Simulation Results for Base Muffler Using the COMSOL tool, the transmission loss at each point with the interval of 5 Hz is within the range of 60–2000 Hz, and all 389 points’ data are imported in MATLAB where root-mean-square transmission loss (RMSTL) is calculated which is about 20.90 dB (Fig. 8).

10 Table 2 Transmission loss from experimental results for base muffler

R. Gavit and K. Wani Frequency (Hz)

TL (dB)

63

7.8

80

6.9

100

7.3

125

7.5

160

11

200

9.6

250

13.3

315

0.5

400

11.5

500

9.7

630

0.3

800

15.1

1000

17.1

1250

18.3

1600

22.2

Fig. 7 Transmission loss from experimental result

6.3 Modified Geometry For input variables, we are taking the length and diameter of each component length, muffler, and outlet. Effect of varying lengths of components See Table 3.

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

11

Fig. 8 Transmission loss from the simulation result

Table 3 Effect of length variation of components on TL Inlet pipe length

Outlet pipe length

Muffler length

RMSTL

00

400

272

503

20.90

+ 5%

420

Percentage of change

19.43 285.6

20.68 528.15

+ 10%

440

18.71 299.2

20.42 553.3

+ 15%

460

20.12 578.45

380

19.36 477.85

360

18.97 19.38

272.8

19.81 452.7

− 15%

19.93 20.06

258.4 − 10%

20.02 18.62

312.8 − 5%

19.84

340

18.92 16.74

231.2

21.04 427.55

18.89

12

R. Gavit and K. Wani

Table 4 Effect of diameter variation of components on TL Percentage of change Inlet pipe diameter Outlet pipe diameter Muffler diameter RMSTL 0

48

+ 5%

50.4

42

93

19.56 44.1

19.46 97.65

+ 10%

52.8

19.52 102.3

55.2

19.77 106.95

45.6

19.49 88.35

43.5

19.50 19.47

37.5

18.99 83.7

− 15%

19.13 18.98

39.9 − 10%

19.19 19.99

48.3 − 5%

19.39 19.68

46.2 + 15%

20.90

40.8

19.55 18.77

35.7

18.99 79.05

18.89

Effect of varying diameters of components See Table 4. From the graph, we can see that the base muffler has optimized values than others, but when we reduce output length till the 5% reduction, TL is decreasing, but after that, it is rising; at 15%, we get maximum RMSTL of 21 dB. The change in diameter of the components does not show any significant effect on TL (Fig. 9).

6.4 Curve Fitting to Obtain Polynomial Equation from Varied RMSTL The effective criteria for model selection are comparing all criteria in which the coefficient of determinations, R2, has the highest value among others. It measures the value for variation considered for the best-fitted model. It is often used to compare models and analyze which model gives the best-suited fit to the data. R always escalates with the model size. For example,

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

13

Fig. 9 Combined graph of the effect of variation of length and diameter on RMSTL

%% Muffler diameter % Goodness of fit: % SSE: 0.03593 % R-square: 0.9861 % Adjusted R-square: 0.9164 % RMSE: 0.1895 All data obtained till now are in discreet form; by importing these data in MATLAB curve fitting tool, we can get polynomial function. For example, A(x) = a1 ∗ x 6 + a2 ∗ x 5 + a3 ∗ x 4 + a4 ∗ x 3 + a5 ∗ x 2 + a6 ∗ x + a7

(5)

Coefficient values are different for all six conditions. Figure 10 displays curve fitting tool window of polynomial curve for input diameter.

Fig. 10 Polynomial curve for input diameter

14

R. Gavit and K. Wani

Fig. 11 Optimized inlet length

The MATLAB Optim tool is used to optimize the data, for which genetic algorithm solver is used with the population of 100 and other parameters are adjusted as per best for each equation separately (Figs. 11 and 12). The above plot shows the optimized best fit for inlet length. It is like the way the other six optimizations are done. Now, the result obtained is given in Table 5. From the above data, we can see that the most significant parameter is length of transmission loss. After getting all optimized values for each parameter using a genetic algorithm, we reconstructed the 3D model to test it for transmission loss. Now, the value of obtained RMSTL is 23.04 dB (Fig. 13). If we observe the plot by comparing it with the plot of TL of the base muffler, we can see that its peak muffler TL is higher than the base muffler which is 85 dB, and for the optimized muffler, it is 90.05 dB. If we compare the peak frequency in the base muffler, it is observed at 1000 Hz, but in the optimized muffler, it is at 800 Hz.

7 Conclusion Obtained RMSTL for the optimized muffler is 23.04 dB, and for the base muffler, it is 20.9 dB. So, there is a 10.24% increase in overall RMSTL. Variation of diameter shows the least effect on the maximization of transmission loss, but length reduction of inlet and outlet pipe by − 2.02% and − 2.09% shows significant effect on

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

15

Fig. 12 Optimized inlet diameter

Table 5 Optimization result table Part

Original dimension (cm)

Optimization percentage

Optimized dimension

Inlet diameter

48

− 2.0260

46.992

Inlet length

400

− 15

340

Muffler diameter

93

7.3963

99.882

Muffler length

503

− 0.1055

502.497

Outlet diameter

42

− 2.0938

41.118

Outlet length

272

15

312.8

RMSTL. Increasing the length of the expansion chamber will help to increase the T, and the increase of muffler pipe length area for perforation will also increase flow enhancement. Peak transmission loss in the base muffler is 85 dB, and in the optimized muffler, it is 90.05 dB.

8 Future Work Optimization of the muffler with the approach of the stress analysis using Modal analysis of muffler to check for dominant frequency. As a result, frequency at which the peak occurred is 1000 Hz in the base muffler and 800 Hz in the optimized muffler, which is shifted to a lower frequency, so it is needed to check its overall effect and study the effect of changing of material on the following geometry.

16

R. Gavit and K. Wani

Fig. 13 Transmission loss of optimized muffler in COMSOL

Acknowledgements This work is carried in the NVH laboratory under the support of ARAI, Chakan, India.

References 1. Herrin DW Design of muffler and silencer. NVH short course, lecture notes 11 2. Munjal M (1987) Acoustic of ducts and mufflers, 2nd edn, vol 21. Wiley, New York (NY) 3. Motor vehicle act regulations last amended September 20, 2020, by B.C. Reg. 240/2020, Section 7.03 4. Ranjbar M, Marburg S, Hardtke H-J (2013) Vibroacoustic optimization of mechanical structures: a controlled random search approach. Adv Mater Res 622:158 5. Barbieri R, Barbieri N (2006) Finite element acoustic simulation-based shape optimization of a muffler. Appl Acoust 67:346 6. Yeh LJ, Chang YC, Chiu MC, Lai GJ (2002) Computer-aided design on single expansion muffler under space constraints. In: Proceedings of the nineteenth C.S.M.E. national conference (Taiwan, 2002), pp 625–633 7. Yeh LJ, Chang YC, Chiu MC, Lai GJ, Her MG (2003) Shape optimization of constrained double-chamber muffler with extended tubes by mathematical gradient methods. In: Proceedings of the 3rd conference on S.M.E. (Taiwan, 2003) 8. Yeh LJ, Chang YC, Chiu MC (2005) Shape optimal design on double-chamber mufflers using simulated annealing and a genetic algorithm. Turk J Eng Environ Sci 29:207 9. Ranjbar M, Kermani M (2014) Muffler design by noise transmission loss maximization on narrow band frequency range. In: The 7th automotive technologies congress (OTEKON 2014), 26–27 May 2014, Bursa, Turkey

Muffler Transmission Loss Optimization for a Vehicle Using Genetic …

17

10. Kermani M, Ranjbar M (2015) Muffler design by noise transmission loss maximization. Doctoral thesis, Easter Merriten University 11. Kermani M, Ranjbar M (2016) Comparative study of design optimization of mufflers by genetic algorithm and random search method. J Robot Mechatron Syst 1(2). ISSN: 2399-1550 12. Kermani M, Orak (2018) On sound transmission loss maximation of a multi-chamber exhaust system. In: OTECON 2018 13. Vora M, Swarnkar R (2016) Effect of muffler design on amplitude and sound frequency of tractor noise. IJERT 5(09):227–235. ISSN: 2278-0181 14. Barnard AR, Rao MD (2004) Measurement of sound transmission loss using a modified four microphone impedance tube. In: Noise-Con’04. The 2004 national conference on noise control engineering, Baltimore Maryland, United States 19 15. MathWorks academy documentation on “Genetic Algorithm”, “Curve Fitting”, “RMS” (2021)

Design and Simulation of a Wireless Charging System for Electric Vehicle Nikhil Kadam and Archana Thosar

Abstract This paper presents design process and working of wireless power transfer system. WPT system is designed to add convenience and sense of reliability to end consumer’s life. WPT system has been developed for small electric city car which has power requirement of 560 W. Lead acid battery having maximum terminal voltage of 56 V and maximum charging current is 10 A. Depending on components used to transfer power wirelessly, wireless power transfer (WPT) technology has been divided into three parts, i.e., capacitive WPT, inductive WPT and magnetic gear WPT. Among other methods, resonant inductive coupling method is advantageous. Resonant inductive WPT is used to charge the electric vehicle (EV). Inductive power transfer (IPT) mainly consists of the two coils separated by the air. One coil is placed on ground and other on base plate of electric vehicle. Along with the coil, power converters are also present in the circuitry. On the transmitting side, power supply circuitry is connected to transmitting coil placed on the ground, and on receiving side, power conditioning circuitry is placed on vehicle side. Keywords Wireless power transfer · Plug-in electric vehicles · IPT · Spiral coil

1 Introduction Wired charging has more power transfer efficiency still it has some disadvantages. Wired charging system has some problems such are exposure of wire, problem of electric shock and vulnerable to adverse weather conditions. Wireless charging technology provides advantages over wired charging system. In wireless charging technology, as there is no presence of physical connection risk of electric shock is reduced. Evolution of wireless charging technique can make life of electric vehicle consumer N. Kadam (B) · A. Thosar Department of Electrical Engineering, College of Engineering Pune, Pune, India e-mail: [email protected] A. Thosar e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_2

19

20

N. Kadam and A. Thosar

more convenient. Wireless charging can also introduce possibility of charging while moving concept, which can eventually introduce concept of “infinite driving range” for electric vehicle. Wireless power transfer (WPT) technology is an emerging technology. It is finding its applications in various fields such are wireless information transfer systems and wireless chargers. Wireless charger is being promising feature of WPT, and it is been developed across the world and spreading its area of implementation in E-vehicle charging also. Wireless power transfer technology is a promising technology because it offers lot more advantages over traditional wired power transfer technology. Wireless power transfer systems (WPTS) are formed by two sections, transmitting section, i.e., primary section and receiving section, i.e., secondary section. In WPTS, coils are used to transmit power from primary to secondary, i.e., from transmitter to receiver. Primary section transmits power to secondary section through resonant coupling field technology. Three technologies can be used in the WPTS which are inductive coupling WPT, capacitive WPT and magnetic gear WPT. From these three inductive coupling, technology is most convenient because of its higher power transfer efficiency [1]. To increase power transfer efficiency, further resonance introduced to the system. Resonant inductive coupling technology provides best performance for wireless charging of electric vehicle [2]. Adding resonance to the system makes WPT systems more efficient when they are operated at higher power levels. Capacitors are used on transmitting side as well as receiving side of the coupling coil to introduce resonance in the circuit. Position of the capacitor is used to define the compensation topology. Series-series compensation topology yields maximum efficiency when it comes to inductive coupled wireless power transfer [3]. Efficiency of wireless power transfer in resonant WPT system does not depend on self-inductance of the coil, but it depends on the mutual inductance between transmitting coil and receiving coil. Transferring higher power levels requires higher frequencies to be used. High frequency has impact on efficiency as well [4]. Power transfer efficiency is affected by vertical distance between the coils as well as horizontal misalignment between the coils. More the misalignment higher are the losses [5]. Power electronics is used to support WPT system. Power electronics that is used to develop wireless power transfer system consists of switching devices. Wide band gap devices provide high capacity, high operating frequency range, wide operating range and good thermal stability. Wide band gap switching device such as SiCMOSFET is intended to use in high frequency applications [6]. Bidirectional inductive power transfer in WPT is possible between the two coils, separated by distance, magnetic coupling [7]. A model has been developed along with controller design. Khutwad [8] has implemented wireless IPT system with efficiency level 67% and prototype was made. This paper develops in two stages: at first, it gives wireless power transfer system specifications for which WPT system is intended. Brief description about the wireless power transfer system is included in this stage. In second stage, design procedure has been covered. More in detail, Sect. 2 gives overview of general scheme of wireless power transfer system. Section 3 describes wireless power system for electric

Design and Simulation of a Wireless Charging System for Electric Vehicle

21

vehicle. General block diagram of WPT for electric vehicle is included in it. Section 4 illustrates the battery charging profile which has been taken as a base to design the WPTs. Specifications of the system have been derived from this charging profile itself. Section 5 is devoted for study of coupling section and coils used in WPT systems. Coil simulation and results got in ANSYS are included here. Section 6 deals with the designing of the power electronics part. Simulations and results has been included in the same section. Section 7 concludes the paper.

2 Overview of WPT System WPT can be supported by various techniques. Based on operating principles, these techniques are identified. WPT techniques are capacitive, inductive and magnetic gear. Classification is given in Fig. 1.

2.1 Magnetic Gear WPT Magnetic gear technology consists of two magnets. Interaction between two synchronized permanent magnets is used for coupling. Due to these moving parts, wear and

Fig. 1 Classification wireless power transfer (WPT) techniques

22

N. Kadam and A. Thosar

tear of the system can happen. In MG, current produced at the output is directly proportional to speed of rotation of secondary magnet, which is very difficult to control. This makes it not suitable for WPT of EV.

2.2 Capacitive WPT In capacitive power transfer, two metal plates are used as medium to transfer power wirelessly. The surfaces facing each other of these plates can act as conducting plates can be used to transmit field. As electric field is confined between these two metal plates only there will be less amount of loss of field. Though capacitive power transfer has this advantage of very small loss of field, the capacitive wireless power transfer technology poses a practical disadvantage of small coupling capacitance. Capacitance depends on area between the plates. As distance between the plates in wireless charging system for electric vehicle would be around 15–20 cm, capacitance value will be small. This is the reason capacitive wireless power transfer technique is useful for application with small air gap and not useful for large air gap application such as EV wireless charging.

2.3 Inductive WPT In inductive WPT to transfer power, wirelessly two magnetic coils are set to resonate at the natural frequency. Two coils are placed with air gap in between. An electric current is passed through the transmitter coil that will generate a magnetic field, which causes receiver coil to induce voltage. The magnetic resonance results in the transfer of electric energy through the air from transmitter coil to receiver coil. Capacitive and magnetic gear WPT techniques are not suitable for the wireless charging for the electric vehicle. Inductive WPT being the suitable for transferring power over a distance of 10–20 cm it is useful for wireless charging of EV. Because of which we have chosen inductive WPT to build wireless charging system. WPT system consists of a coupling section. Coupling sections acts as core of the WPT system. Coupling section consists of coil or capacitive plates based on the type of the WPT technique used. In case of the inductive WPT along with inductive coils, capacitor is used to add resonance to the circuit.

3 WPT System for Electric Vehicle Wireless power transfer systems (WPTS) are formed by three sections, transmitting section, receiving section and coupling section. Figure 2 shows all the components of WPTS. Transmitting section aims to produce in phase AC voltage at the input

Design and Simulation of a Wireless Charging System for Electric Vehicle

23

Fig. 2 General scheme of WPT system

of the coupling section. Current from grid supply has distortion in it which results in less power factor. Loss of power factor can be overcome by using power factor circuitry. In transmitter section, power conversion takes place as AC–DC–AC. This double conversion stages are used to increase power frequency from 50 Hz to tens of kHz. Receiving section consists of AC–DC rectifier and DC–DC chopper circuit. Receiving section uses DC–DC chopper circuit to maintain required voltage and regulate charging current at the input of the battery. Coupling section is used as connecting link between transmitting section and receiving section. WPTS of electric vehicle uses inductive coils as coupling device. When current passes through transmitting coil it produces magnetic field around it. Produced magnetic field when interacts with receiver coil, induces current in the receiving coil. This phenomenon is known as coupling. Amount of coupling is defined by using coupling coefficient, more the coupling coefficient more is the power transfer efficiency. To increase the efficiency and to transfer power of higher rating, resonance is to be used. Compensation network used in coupling sections is classified into four types based on position of the capacitor, which are series-series, parallel-parallel, series-parallel and parallel-series. Series-series compensation network yields maximum efficiency. Power conversion cycles used in WPT system requires fast switching switches. Modern wide band gap switches such as SiCMOSFET and SiCIGBT are used to get high switching action. There are two main types of the wide band gap switches, first is SiC, i.e., silicon carbide and second one is GaN, i.e., gallium nitride. Over conventional Si, i.e., silicon material used in switches, SiC has multiple advantages. SiC material is used to develop high capacity and low power loss transistors. In the power conversion stages, power converters build with SiC material transistor helps to reduce power losses. SiC material transistor provides high capacity, high operating frequency range, wide operating range and good thermal stability over conventional Si devices.

24

N. Kadam and A. Thosar

Fig. 3 a Green line—charging current, b blue line—terminal voltage, c violet line—resistive load

4 WPT System Specification Wireless power transfer system is being developed for small electricity car. Before designing this circuits, one should now charging profile of the battery for which system is being designed. From the charging profile, one can get to know about nominal values that can be used to design the system. In this charging profile in first stage, charging current is kept constant and terminal voltage VB is increasing from 36 to 56 V. In second stage, VB is regulated at VB max , by decreasing charging current. The constant current stage ends when charging current falls below threshold value. During the constant current stage power also increases from 360 to 560 W and then reduces to 0. Battery pack appears as a variable resistive load. Which can be defined by VB and IB . Resistance of the load varies from 3.6 to 5.6 Ω during constant current stage. In constant voltage stage when current decreases load resistance increases till 560 Ω (Fig. 3). As this charging cycle is very slow and parameters, i.e., current, voltage and power operates at nominal value for long time these values should be considered for while designing. Specification of the supply voltage and supply frequency comes from the single phase domestic socket connected to power grid. The dimension of the city-car chassis dictates the receiving coil diameter D standard. SAE J294 standard gives standard useable frequency for operation of wireless charging of small light weight vehicles. That frequency is 85 kHz, this is the operating frequency, and we have used to design this paper. The distance between the coils is kept at 15 cm assuming Indian car ground clearance based on the deployment of transmitting coil.

5 Coil Design for Electric Vehicle Coupling section consists of transmitter coil with transmitter compensation network and receiver coil with receiver compensation network. Compensation is used to introduce resonance in the system. Compensation basically is process of adding capacitor along with coil inductance on both sides, transmitter as well as receiver

Design and Simulation of a Wireless Charging System for Electric Vehicle

25

Fig. 4 Circuital scheme of inductive WPTSs (with resonance)

side. The primary and secondary coils of a wireless power transmission system have a high leakage inductance. The main coils require a lot of current to transfer electricity properly. The WPTs system suffers from a large deal of deprivation as a result of the high primary supply current, which leads to decreased effectiveness. The reactive components should be compensated by using a capacitor to solve the problem. As a result, the resistive element persists in the network, lowering losses while increasing effectiveness. To add resonance to the system, capacitors are used in compensation techniques. Placement of capacitor indicates the name of compensation techniques, such as S-S, S-P, P-S and P-P. From all of this S-S compensation, technique is easy to model and efficiency analysis results show S-S helps to compute maximum efficiency (Fig. 4).

η=

ω 2 M 2 RL (RR + RL ) RS + ω2 M 2 (RR + RL ) 2

(1)

The efficiency is increased by using high mutual inductance and operating frequency values. Along with high M and ω values, reducing load resistance value also helps to improve the efficiency.

5.1 Coil Types 1. Helix coil In helical coil structure turns are packed in a rectangular section which has base b and height h and their mean distance is Rm. The self-inductance of helix coil can be calculated by using (2) L Helix =

0.31(Rm N )2 6Rm + 9h + 10b

(2)

2. Spiral coil Spiral coil structure is a planer structure. In spiral coil structure, turns are saturated at the outer side of the circle. The self-inductance of spiral coil can be calculated by

26

N. Kadam and A. Thosar

using (3)  c   2 + c3 ∅ + c4 ∅2 L spi = C1 μo N 2 Ravg ln ∅

(3)

where Ravg and ∅ can be expressed as (4) Ravg =

Ro − Ri Ro + Ri ,∅ = 2 Ro + Ri

(4)

Between helix coil and spiral coil, spiral coil has more uniform field density as compared helix coil which yields 18% more coupling coefficient. Because of which in IPT spiral coil is used (Fig. 5). In inductive WPT, inductive coils made with Litz wire are being used here in our system. Physical standards of coil are given in Table 1. ANSYS is used to perform the FEM analysis of the coil. Spiral shaped coil shown in Fig. 6 has been developed in ANSYS, and performance of the coil is analyzed.

Fig. 5 Coils a helical type b spiral type

Table 1 WPT specifications

Parameter

Value

Nominal supply voltage

230 V

Nominal output power

560 W

WPT operating frequency

85 kHz

Coil radius

0.19 m

Distance between coil

0.15 m

Number of turns (N)

18

Design and Simulation of a Wireless Charging System for Electric Vehicle

27

Fig. 6 ANSYS model for FEM analysis of spiral coil model

5.2 Coupling Section Design Transmitter and receiver coil are assumed to be of similar dimensions. Both are of spiral type and has inner diameter of 0.16 m and outer diameter of 0.38 m (Table 2). R0 = 0.19 m Ri = 0.08 m 1 = 0.3563 k= 2  23  2 D 1 + 2 3 √R R o

Ravg =

i

0.19 + 0.08 Ro + Ri = = 0.135 m 2 2

0.11 0.19 − 0.08 Ro − Ri = = 0.4074 m = Ro + Ri 0.19 + 0.08 0.27 ⎧ ⎪ C1 = 1.00 ⎨  c  ⎪ C2 = 2.46 2 + c3 ∅ + c4 ∅2 = C1 μo N 2 Ravg ln ⎪ C = 0.00 ∅ ⎪ ⎩ 3 C4 = 0.20

∅=

L spi

L spi = 108.655 µH

28

N. Kadam and A. Thosar

Table 2 Coil design parameter

Number of turns (N1, N2)

18

Coil radius

19 cm

Cross section of wire

2.6 cm

Fs

85 kHz

Self-inductance (L T , L R )

118.655 µH

Capacitance (CT , CR )

36 µF

k≙√

M ; M = 35.86 µH LT LR

CT = CR =

1 = 36 µF ω2 L

6 Power Converter Design Figure 7 shows power flow that will be used in this wireless power transfer system. For power conversion in primary, in stage 1, we have used full wave rectifier, followed by in stage 2, we have used high frequency inverter to attain frequency requirement of the system. In secondary stage, we have receiver rectifier circuit along with DC–DC chopper circuit to get stable power output at the input of the battery of electric vehicle. These power conversion stages require fast switching switches. Modern wide band gap switches such as SiCMOSFET and SiCIGBT are used to get high switching action.

6.1 Rectifier with PFC-Boost Circuit See Fig. 8. The design parameters of PFC boost chopper are as follows. The input design data is taken from Table 3. Preliminary values of currents and voltages are required to determine the inductor and capacitor values of the boost chopper, and they may be determined as follows.

Fig. 7 Wireless charging system

Design and Simulation of a Wireless Charging System for Electric Vehicle

29

Fig. 8 Rectifier with PFC-boost circuit

Table 3 Design parameters of PFC boost chopper

VG,min

207 Vrms

VG,max

253 Vrms

Output voltage Vout

370 V

Vout,min

350 V

Grid frequency f G

50 Hz

Switching frequency f sw

85 kHz

Iripple

10%

Vripple

6%

Maximum output power Pout_ max

700 W

The following formula can be used to compute the maximum value of the output current Iout,max =

Pout,max Vout

and value comes out to be 1.89 Arms that can be seen in Fig. 9. The grid current can be calculated as Pout,max = 3.71 Arms ηVin P F √ IG,pk = 2IG = 5.25 A

IG =

IG,avg =

2IG,pk = 3.34 A π

30

N. Kadam and A. Thosar

Fig. 9 Output of AC-DC rectifier with PFC

By considering 10% current ripple of peak grid current, the ripple current will 0.525 VDC,R,min =

√ 2VG,min = 294.7

By considering 6% voltage ripple of peak grid current, the ripple voltage will 17.56 V. Input filter capacitance can be obtained as CR =

IRipple = 44 mF 8 f sw VG,ripple,max

D = (Vout − VDC,R,min )/V _out = 0.209 L> C>

Vout .D(1 − D) = 1.37 mH Iripple . f sw

2ton Pout = 0.597 mF = 0.6 mF 2 − Vout,on

2 Vout

Capacitor at the out of the PFC is used to get smooth output at the output. To get smooth output from the rectifier, capacitor is placed in parallel to the load.

Design and Simulation of a Wireless Charging System for Electric Vehicle

31

6.2 High Frequency Inverter Inverter used in this paper is full bridge inverter. Structure of full bridge inverter has been showed in Fig. 10. Based on a phase shift full bridge, inverter with a phase shift of as near to 100% is feasible. This inverter is designed to provide higher voltage and power levels while maintaining a same load current level. When compared to traditional pulse width modulation (PWM) waveforms, the major advantage of phase shifted control is the ability to regulate not only the duty cycle but also the switching losses to some extent. It is not possible to achieve a 100% phase shift in the planned complete bridge, but it is conceivable to get extremely near, about 99%. The voltages of points A and B of Fig. 10 have the waveforms showed in the first two plots of Fig. 11, while the voltage VAB has the alternate waveform reported in the third plot. Because the supplied circuit resonates at the operating frequency, only the first harmonic V S of VAB affects the coil current I S. Its peak amplitude VS,pk is varied by adjusting φ according to VS,pk = VDC

φ 4 sin π 2

(5)

The maximum amplitude is VDC π4 , reached when φ is set at π and the output voltage VAB of the H-bridge is a square wave. 4 π VCH ∼ 9 A this can be visualized in the graphs (Fig. 12). IS,pk = ηωM min = Using HFI inverter DC power from rectifier circuit has been converted to AC power. Square AC voltage and current has given to coupling section.

Fig. 10 Full bridge high frequency inverter

32

N. Kadam and A. Thosar

Fig. 11 HFI output voltage

Fig. 12 Stage 2 output

6.3 Chopper Circuit According to Fig. 8 which is the functional block diagram of the system at the end of power conversion stage and just before the car battery chopper circuit is used. A chopper is a static power electronic device that is used to convert fixed dc input voltage to a variable dc output voltage (Fig. 13).

Design and Simulation of a Wireless Charging System for Electric Vehicle

33

Fig. 13 Circuit diagram of chopper

The chopper circuit’s job is to regulate the current I B injected into the battery and the voltage VB across its terminal according to the charging profile of the battery. Peak current in the power switch during on-time is larger than ICH,N and equal to ICH,N if current ripple is IB,N . Current ripple is maximum when the chopper duty cycle is 50%. This condition is never going to occur in this WPT system, as minimum duty = 0.55, i.e., duty cycle never going to go down cycle that can occur here is VVB,min CH below 55%. For chopper circuit to operate in the continuous mode during entire process, ripple current should not go below IB,min value. Data 1. 2. 3. 4. 5.

Input nominal voltage, Vi: 65 V. Output required voltage (charging voltage), Vo: 56 V. Current ripple = 20%. Voltage ripple = 2% Switching frequency (Fs ) = 85 kHz Calculations: ΔIL = 20% of load current = 20% of 10 A = 2 A ΔVc = 2% of 56 V = 1.12 V Duty ratio =

Vo VS

=

56 64

= 0.875

L CH =

D(1 − D)Vs ∗ ΔIL

COUT =



1 Fs

= 41.176 µH

D(1 − D)Vs  2  ∗ T = 2.616 µF 8LΔVc

34

N. Kadam and A. Thosar

Fig. 14 Output after final stage

While converting DC–DC power chopper circuit has been developed such as maximum charging voltage and current will not exceed value VB = 56 V and I B = 10 A. Graph shown in Fig. 14 shows that charging current is less than 10 A, charging voltage is nearly equal to 56 V. Power transferred to the battery comes out to be 560 W.

7 Conclusion The paper has given detailed survey which concludes that inductive WPT is best suitable for electric vehicle application. And difference between spiral and helix coil structure given with pictorial view. Spiral coil structure has advantage over helix structure. This paper states design of the wireless power transfer system for electric golf vehicle. While designing the inverter, rectifier and converters they are assumed to have efficiency of 95%, which in turn produce overall 77% efficiency.

Design and Simulation of a Wireless Charging System for Electric Vehicle

35

References 1. Ching TW, Wong YS (2013) Review of wireless charging technologies for electric vehicles. In: 2013 5th International conference on power electronics systems and applications (PESA), Hong Kong, pp 1–4. https://doi.org/10.1109/PESA.2013.6828235 2. Jegadeesan R, Guo YX (2012) Topology selection and efficiency improvement of inductive power links. IEEE Trans Antennas Propag 60(10):4846–4854 3. Feng H, Cai T, Duan S, Zhang X, Hu H, Niu J (2018) A dual-sided tuned series–series compensated resonant converter for wide charging region in a wireless power transfer system. IEEE Trans Ind Electron 65(3):2177–2188 4. Buja G, Bertoluzzo M, Mude KN (2015) Design and experimentation of WPT charger for electric city car. IEEE Trans Ind Electron 62(12):7436–7447. https://doi.org/10.1109/TIE.2015. 2455524 5. Mude KN, Outeiro MT (2017) Coil misalignment analysis under different radius of coil and wire for wireless power transfer system. In: IECON 2017—43rd annual conference of the IEEE industrial electronics society, Beijing, pp 5319–5323. https://doi.org/10.1109/IECON.2017.821 6921 6. Liu J, Zhang Y, Wang Z, Cheng M (2018) Design of a high-efficiency wireless charging system for electric vehicle. In: 2018 1st workshop on wide bandgap power devices and applications in Asia (WiPDA Asia), Xi’an, China, pp 40–44. https://doi.org/10.1109/WiPDAAsia.2018.873 4657 7. Swain K, Neath MJ, Madawala UK, Thrimawithana DJ (2012) A dynamic multivariable statespace model for bidirectional inductive power transfer systems. IEEE Trans Power Electron 27(11):4772–4780 8. Khutwad SR, Gaur S (2016) Wireless charging system for electric vehicle. In: 2016 International conference on signal processing, communication, power and embedded system (SCOPES), Paralakhemundi, pp 441–445. https://doi.org/10.1109/SCOPES.2016.7955869

Recent Advances in Sensor Technology for Biomedical Applications: A Review Niharika Karnik, Karan Bhadri, and Pankaj Dhatrak

Abstract Sensor technology has become an integral part of the diagnosis, monitoring, therapeutic and surgical areas of medical science. Various sensors like glucose biosensors for diagnosis of diabetes mellitus or fluorescent sensors for gene expression and protein localization have become a common part of the biomedical field. Due to their widespread applications, various advances and improvements have taken place in medical sensor technology which has led to an increase in the ease and accuracy of diagnosis as well as treatment of diseases. This review article aims at studying various novel and innovative developments in biosensors, fibre optic sensors, sensors used for microelectromechanical systems, flexible sensors and wearable sensors. This article also explores new sensing methodologies and techniques in different medical domains like dentistry, robotic surgery and diagnosis of severe life-threatening diseases like cancer and diabetes. Various sensors and systems used for rapid detection of the SARS-CoV-2 virus which is responsible for the COVID-19 pandemic have also been discussed in this article. Comparison of novel sensor-based systems for detection of various medical parameters with traditional techniques is included. Further research is necessary to develop low cost, highly accurate and easy-to-use medical devices with the help of these innovative sensor technologies. Keywords Biosensors · Fibre optic sensors · Flexible sensors · MEMS · Robotic surgery · Sensors in dentistry

1 Introduction Sensors are an integral part of the technical and technological advances that have taken place in fields like Robotics, Industrial Automation, Internet of Things and Mechatronics. The Biomedical world is another domain that has seen an exponential rise in the use of sensor technology not just for diagnosis of diseases or monitoring of patients but also for surgical procedures [1, 2]. One category of N. Karnik (B) · K. Bhadri · P. Dhatrak Dr. Vishwanath Karad, MIT-World Peace University, Pune, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_3

37

38

N. Karnik et al.

sensors is used in the medical field measure and monitor physical aspects like temperature of the body, pulse rate, heart sound, blood viscosity, respiratory rate and flow [3]. Other sensors measure chemical and biological aspects like pH value, oxygen/carbon dioxide concentration, enzymes, antigens, antibodies, receptors and hormones. Due to this widespread use of sensors, various advances have taken place in their technology. Due to the advances in biomedical microsystems, a large number of ground-breaking microelectromechanical system (MEMS) devices which make use of specialized sensors have been introduced in the fields of diagnosis, precision surgery and therapeutic systems [4]. Fibre optic sensors have both sensing and communication functions and are hence used in biosensing applications such as robotic surgery, endoscopy and thermal ablation [5]. A certain study investigated the design and performance of an extrinsic fibre optic sensor for measuring the frequency and amplitude of heart rate signal [6]. To overcome the drawbacks of rigid materials, flexible sensors that have improved mechanical, electrical and thermal properties have been presented for strain sensing in various body parts such as elbow, finger, knee and neck [7]. Tissue-based, enzyme-based, DNA biosensors, immunosensors, piezoelectric and thermal biosensors have applications in both the medical field and metabolic engineering [8]. In the cardiac field, optical voltage sensors are applied for cultured human stem cell-obtained cardiomyocytes as well as in whole hearts in vivo [9]. In the urology domain, researchers have developed an innovative Zn2 + sensor that can be applied to the clinical diagnosis of early prostate cancer [10]. These sensor technologies have contributed towards inexpensive, quick and at times noninvasive diagnosis of diseases. This review introduces the progress made by various novel sensors and affiliated systems to improve the healthcare systems through fast detection and treatment of diseases. The further sections in this paper will discuss the various advances and improvements made by sensors in the medical domain.

2 Materials and Methods Various sensors are used for diagnosing, treating and monitoring diseases. They are generally used for sensing blood pressure, temperature, pH value, respiratory rate, oxygen and carbon dioxide concentration and many such applications [2, 3]. Advances in sensor technology have made diagnosis of diseases easier and more accurate [8]. The study of these advances is important to apply them to our daily healthcare systems and to overcome any distribution and technological problems that might arise while commercializing these products. Remote monitoring of patient health conditions is also possible, and hence, the further sections discuss the introduction of novel and innovative sensors in the biomedical field.

Recent Advances in Sensor Technology for Biomedical Applications: …

39

2.1 Biosensors They are investigative devices that translate a biological response to an electrical one, and they should be very specific, not dependent on physical parameters like temperature and pH and must be reusable [8]. They have major benefits like quick speed of determination, high sensitivity and decent selectivity and have applications in safety of food, biomedical research and environmental detection [3]. A biological factor which acts as a molecular recognizing agent and a physicochemical detection transducer makes a biosensor [11]. Figure 1 shows the various transducers used in biosensors and their principle of operation. Recognizing agents are immobilized on the transducer surface, and these interact with target molecules while preventing addition of reagents in the sample solution. For operative incorporation of recognizing agent in the sensor, there are four major coupling mechanisms that are used for obtaining the essential purpose: immobilization of membrane, entrapment of matrix, covalent fabrication and encapsulation by physical adsorption as shown in Fig. 2 [12]. Due to the incessant advances in micro-electronics, micro-machining, optoelectronics and biomedicine, biosensor technology has evolved to a great extent. Due to innovative and useful nanomaterials and investigative technologies, there is a potential for advancement of biosensor devices for various applications such as biomedical, clinical, bio-technological, biological, food industries, health and environmental monitoring [13]. Voltametric, Potentiometric, Conductometric, Optical, Calorimetric, Enzymatic, Impedimetric, Piezoelectric, Immunosensors and DNA sensors are the different types of biosensors in use today [14]. Further research based on biosensors can prove useful in not just detecting cancer in patients but also prevent in-stent restenosis and for study of genomes through detecting biomarkers. Some of the recent advances in biosensors along with their principle and use and

Fig. 1 Various transducers used in biosensors and their operation principle

40

N. Karnik et al.

Fig. 2 Different coupling mechanisms for operative incorporation of recognizing agent in the sensor: a covalent fabrication, b entrapment of matrix, c immobilization of membrane, d encapsulation by physical adsorption [12]

their application are shown in Table 1. The various biosensors discussed in the table will help future researchers to understand the principle and application of recently developed sensors such as the exosome detection biosensor or the hydrogel-based immunosensors.

2.2 Wearable Sensors A wearable device used for biomedical applications consists of sensors, drives, displays and computer elements that through Wireless Network connections gives us a digital world and makes our lives convenient and comfortable [3]. Watches, rings, wristbands, armbands, belts, running shoes, buttons and helmets make up a majority in the market for medical devices that can be worn [19]. The quality of life of healthy individuals and patients has improved drastically cause of implantable sensors and bio-integrated electronics that can be worn [20]. Improvement in surgeries and their outcomes has been achieved due to continuous and real-time monitoring of pathological and physiological aspects for disease detection. Wearable sensors have the potential to revolutionize the modern healthcare systems through their ability to process data in real time. It thus becomes very important to understand the various aspects of these sensors and carry out future research on integrating concepts like neural networks into their processing systems. Uses of wearable sensors have two major categories called biophysical tracking and biochemical monitoring which are explained below in further detail [21].

Biosensor

Exosome detection biosensor

Resonator sensors based on capacitance in bioresorbable stents

Magnoelastic resonant sensors in stents

Hydrogels based immunosensors

Graphene-based biosensors

Name of author

Cheng et al. [15]

Hoare et al. [16]

Hoare et al. [16]

Balahura et al. [17]

Bai et al. [18]

Table 1 Recent advances in biosensors technology (2019–2020)

Various approaches for outputting the signal such as electrochemistry, fluorescence, plasmon surface, Raman scattering which enhances the resonance of graphene and surface are used

Interactions between an antigen and an antibody on transducer surface produce a change in the hydrogel constitution which can be detected by a transducer

Resonance of magnoelastic materials who change their magnetic susceptibility due to mechanical stress application

Sensors made use of passive capacitance and inductance resonators which are biodegradable

Recognition is done with “lock and key” aspect of antigen and antibody interaction. Electrical transduction is either optical or electrochemical

Principle

Biomarkers related to cancer like miRNA, DNA, proteins and small molecules are quantitatively detected

Immunosensor applications consist of microbiological analyses, environmental analyses, genomics and proteomics, microarrays, nanotechnology, agriculture, quality of food and military applications

A term called ISR (in-stent restenosis) causes narrowing or blocking of vessel where the stent is placed, and these sensors are used for detection of stent occlusion

A term called ISR (in-stent restenosis) causes narrowing or blocking of vessel where the stent is placed, and these sensors are used for detection of stent occlusion

These biosensors can be used for quick and easy detection of cancer in regions that have limited medical resources and where diagnostic facilities are not abundant

Application

Recent Advances in Sensor Technology for Biomedical Applications: … 41

42

2.2.1

N. Karnik et al.

Biophysical Tracking

Human motion analysis, temperature and heart rate monitoring play a vital part in the assessment of health, and thus, these are the major types of wearable sensors for biophysical tracking [22]. Cardiovascular diseases are responsible for deaths and can generally be identified by irregularities in the heart rate. A study made a long-term wearable CNT (carbon nanotube) and PDMS (polydimethylsiloxane) and ECG (Electrocardiogram) electrode that can be easily connected to the normal ECG devices and is sweat and motion resistant [23]. Another study developed a biodegradable, single-use pressure sensor useful in cardiovascular monitoring, and it displayed a high sensitivity of 0.78 kPa−1 [24]. Many diseases are defined by change in body temperature, and thus, measuring body temperature is a very important aspect that can help doctors diagnose diseases and also analyse the efficiency of treatments [21]. Research has been done on temperature sensors that are both stretchable and wearable, and analysis of their capacitive response to temperature ranges of 25–290 °C is shown [25]. Another study defined the fabrication of an extremely deformable and wearable temperature sensor made up of an element of sensing, interconnections and the pads which can be used for post-surgery monitoring. [26]. Motion sensing is another important factor of biophysical tracking as it can gauge the strain in various joints and muscles when our body is in motion. Various piezoelectric wearable sensors for tracking human motion have been defined with some of them having energy harvesting capacities too [20].

2.2.2

Biochemical Tracking

Wearable sensors can be used for blood, pH and biomolecule monitoring and tracking which can help in detection of cancer and diabetes and help in monitoring blood pressure. Blood test comprises sample collection which is invasive and is followed with centrifugation for separation of plasma. Analysis of chemicals is the next step, and all this may avert people who need urgent medical help from getting the quick and accurate medical attention needed [27]. Wearable biosensor offers health status tracking in a real-time manner by monitoring non-invasively applicable biomarkers from biofluids. A study presents a skin-worn microneedle sensing device which makes use of a microneedle-based enzyme electrode for nominally invasive monitoring of subcutaneous alcohol [28]. Without any compromise in stability, response time or sensitivity this sensor provides free of interference ethanol detection in artificial interstitial fluid. Another study presents the formation of a wearable and stretchable array of microsensors that simultaneously monitors heavy metals like Cd, Pb, Zn, Cu and ions of Hg via electrochemical SWASV (Square Wave Anodic Stripping Voltammetry) on Au and Bi microelectrodes in human body fluids [29]. Heavy metals like Zn and Cu are important trace elements which may have extremely harmful effects on the body if found in excess as high copper concentration can cause kidney and heart failure or Wilson’s disease and low copper levels may cause anaemia and osteoporosis. CuO-based electrochemical pH sensors that have flexible substrates

Recent Advances in Sensor Technology for Biomedical Applications: …

43

are wearable with interdigitated electrodes, and CuO nanostructures are presented in a paper [30]. The pH value shows properties of various biological, physiological and medical conditions and hence helps in the diagnosis and monitoring of diseases.

2.3 Microelectromechanical Sensors (MEMS) These sensors have advantages like their ability to perform in harsh environments, capability of multiplexing, tiny size and immunity to electromagnetic interference and are applied in microsystems, environmental and biomedical domains [31]. Generally, MEMS devices need less energy to function, are less vulnerable to surrounding vibration, have faster reactions and involve very less volume of fluids [32]. In treatment of sleep apnoea, flow generators are required as patients need air flow into their lungs in continuous cycles. Flow sensors are required in this disease to regulate flow of air to avoid a feeling of breathlessness due to the upper airway being obstructed. A study presented a MEMS LCP (Liquid Crystal Polymer) pressure sensor based on membranes and performed experiments to examine the scope of this flow sensor for sleep apnoea therapy [33]. Another study presented a MEMS pressure sensor based on FPI (Fabry–Perot interferometer) with anodic bonding whose major principles of design and rudimentary mechanical model were introduced [31]. Neural sensing is the pillar of machine-brain interface, research in neuroscience and clinical neuromodulation. In a study, a flexible array of microelectrodes is proposed, which is used for neural stimulation sensing on surface of the neuron, without tissue penetration [34]. Triaxial MEMS tactile sensors placed on the tip of forceps for grasping are used in laparoscopic surgery to evaluate stress. [35]. Despite their extensive applications, various limitations related to them restrained the use of MEMS sensors for a while. These disadvantages included high fabrication cost related to the prerequisite of clean room facilities and brittleness that reduced their application spectrum and saturation of their response over an extended time period [36]. Further for their large-scale production, problems may arise due to their small size and the lack of manufacturing facilities. To overcome these disadvantages and to increase the use of MEMS sensors in various medical systems, it is important for researchers to understand the fundamental principles that govern these sensors and increase their subsequent applications. Table 2 shows some of the MEMS sensors, their sensing methods and the parameter they detect.

2.4 Fibre Optic Sensors (FOS) A growing need for non-stop and remotely monitoring medical parameters has created an ever-increasing demand for fibre optic sensors because of their reduced form and the advantage of spatial monitoring with a single or multiple sensors [2]. Optical sensing mechanisms comprise Raman spectroscopy, fluorescence,

44

N. Karnik et al.

Table 2 MEMS sensors used in diagnosis [4] Sensor

Sensing method

Parameter detection

Micro-cantilevers

Mechanical sensing

Tuberculosis, human papilloma virus (HPV)

Amperometric

Electrochemical sensing

Glucose

Potentiometric

Electrochemical sensing

Urea, penicillin, amino acids

Conductometric

Electrochemical sensing

Gases, non-conducting mediums

Impedimetric

Electrochemical sensing

Human immunodeficiency virus (HIV)

Surface plasmon resonance

Optical sensing

Haemoglobin, pathogens

Interferometric sensors

Optical sensing

Herpes simplex virus (PHV)

colorimetry, absorption, luminescence, RI (refractive index) and SPR (surface plasmon resonance) [5]. Light propagates through the core by total internal reflection as the optical fibre is made of a core which is germanium-doped silica and is bound by a cladding made of pure silica which has a somewhat lower RI. Light spreading through fibres is divided into two parts: the “exponentially” decaying field in cladding which is temporary and the directed field of the core. The field which is temporary in cladding goes down to zero before it reaches the external medium. These sensors are currently applied in the vascular, urologic, neural, and gastrointestinal diagnostic fields. These sensors show promising potential for diagnosis of medical conditions of the heart, kidneys and stomach. Due to their flexible properties and small size, they show large number of applications as wearable sensors. A study presents design, operation principle, performance and experiments done on an extrinsic FOS developed for measuring frequency and amplitude of heart rate [6]. The main application areas of the fibre optic sensor are shown in Fig. 3. The fibre optic sensors are mainly of two types: the Fabry–Perot interferometry (FPI) and Fibre Bragg grating (FBG) sensors [5].

2.4.1

Fabry–Perot Interferometry (FPI)

A small cavity inside the optical fibre that produces large spectral variation having a periodic pattern forms the basis of the FPI. They are one of the major aspects for sensing of pressure in medical devices [37]. The cavity is enclosed by a diaphragm that bends the cavity and controls the length when there is application of pressure as shown in Fig. 4. Extrinsic FPI Pressure Sensors are advantageous as they have a comparatively small size, less expensive and simple fabrication process, fairly high sensitivity to pressure, are suited for operations in high temperature areas due to their high mechanical strength [38]. A certain study proposed and demonstrated a microcavity extrinsic FPI curvature sensor with enhanced sensitivity having a shift sensitivity of −5.9607 nm/m−1 along with cross-sensitivity for lower temperature of 9.9696 × 10−3 m−1 /°C in a range of curvature of 3–15 m−1 [39]. Another study

Recent Advances in Sensor Technology for Biomedical Applications: …

45

Fig. 3 Major application areas of fibre optic sensors in the biomedical field [2]

Fig. 4 General layout of the Fabry–Perot interferometer [2]

presents an extrinsic FPI pressure sensor based on a nano-thick silver diaphragm [40]. The sensor thus designed exhibits a linear response within the range of pressure variation up to 50 kPa. Pressure variations can help measure blood pressure, detect variations in heartbeat and monitor the condition of stents. Further research in the use of FPI sensors may prove to be beneficial in reducing risk of heart attacks and low blood pressure and offer treatment for various cardiovascular diseases.

2.4.2

Fibre Bragg Grating Sensors (FBG)

These are the most widespread optical sensors and they are filters for reflection, having a very fine bandwidth and a rapid response, while also being sensitive to temperature and strain [37]. They have applications in oil and gas, structural health monitoring and medical devices and are compact, less expensive devices. A study presents manometry catheters made using FBG which gives a close to perfect method

46

N. Karnik et al.

Fig. 5 Schematic of sensor design used in the fibre optic manometry catheters [41]

for physiological monitoring of the gut in a nominally invasive way as shown in Fig. 5 [41]. Another paper describes a new FBG pressure sensor comprising two FBG sensors and a stretchable silicone cylinder made for monitoring of patients with ventricular assist devices [42]. In another study, fabrication of a small acoustic and flexible FBG sensor having improved sensitivity to pressure was shown which used a side-polished arrangement of ∼15.5 mm length and ∼53 µm depth. Some studies also use the Fibre Bragg grating sensors and the Fabry–Perot interferometer in conjunction [2, 34].

2.5 Flexible Sensors These sensors are mostly non-invasive, and detection of diseases can be done by checking the composition, lack and abundance of constituent material in saliva, tear, neural activity, blood and its pressure, skin, urine and numerous other parameters [32]. Sensors that change mechanical forces like tension, pressure, stress and strain into electrical signals are vital amongst various flexible sensors [7]. Generally, flexible physical sensors are made up of two separate building blocks, the template material which is flexible and the active agents used for sensing, which have liquid or solid form and can have a principle of mechanical deformation like twisting, pressing, bending and stretching of the sensing device [32]. A study presented the fabrication, design and application of an innovative graphite or PDMS (Polydimethylsiloxane) flexible sensor for biomedical applications for detecting strain [1]. Another paper presented a stretchable strain sensor built on ionic liquids and manufactured with the help of Polydimethylsiloxane mould and the method of simple sealing [22]. Another study fabricated a flexible pressure sensor useful for applications of smart skin using graphene oxide and its response to the various pressures assessed by gauging the electric current [43]. Electronic skin is an artificial device that copies the behaviour of the human skin with respect to carrying mechanical stimulus to the inner layers and assembling and processing the received signals. Like human skin,

Recent Advances in Sensor Technology for Biomedical Applications: …

47

Table 3 Various flexible sensors with their biomedical applications [7] Vitro/Vivo

Body part

Parameter evaluated

Application

In vitro

Hand

Motion

Recovery of hand monitoring

Wrist

Pulse

Cardiovascular diseases diagnosis

Upper arm

Blood pressure

Hypertensive treatment

Knee and elbow

Strain distribution

Posture and movement detection; postoperative rehabilitation

Foot

Pressure distribution

Sport training and gait detection

Throat

Sound vibration

Diagnosis of damaged vocal cords and voice recognition

Face

Muscle movement

Face paralysis diagnosis; recognition of facial expressions

Cardiovascular

ECG

Treatment and prevention of cardiovascular diseases

Heart

Heart rhythm

Sinus arrhythmia correction and cardiac pacing

Bone

Pressure

Osteoporosis and fracture treatments and nerve tissue repair

In vivo

it needs to be flexible and hence latest advances in flexible sensors are encouraging the developments of these electronic skins. Research was also carried out on the fabrication of a 3D stretchable ECL (Electro-chemiluminescence) sensor built on a GPE (graphene paper electrode) with large surface area, exceptional conductivity, high porosity and an extensive electrochemical potential [44]. ECL is a procedure where an electrochemical voltage triggers luminescent signals release and is used as an operative investigative method in drug analysis and immune assay. Table 3 gives a brief overview of the various flexible sensors, their placement in the body and their applications.

2.6 Application of Sensors in Biomedical Fields Generally, sensors are used for diagnosis and treatment of various cardiovascular, urological and neurological diseases [2, 32]. However, recently sensors have also been used in diagnosis of various kinds of cancers and diabetes mellitus [12, 45]. They have revolutionized the concept of robotic surgery with easy force feedback for robotic forceps used in surgeries and are used extensively in the field of dental implants for measuring the stress on the implant as well as the primary stability of the implant [46, 47]. This paper highlights the use of novel sensors for developing monitoring and detection systems in the fields of dentistry and robotic surgery while highlighting methods like biomarker detection and real-time monitoring of

48

N. Karnik et al.

cancer and diabetes. The following section highlights some of the advances in sensor technology and their implementation in the above-mentioned fields.

2.6.1

Dentistry

The field of dentistry involves analysis of forces, stresses on the teeth as well as bacteria concentration that can lead to cavities or plaque formation. Hence, various sensors are used extensively in this field for quick and accurate diagnosis and analysis. Maximum bite force present cause of joint action of mastication muscles, teeth and joints is the main pointer for practical masticatory system condition. A study sheds light on a procedure that uses a FBG sensor for dynamic bite force measurement which is called a Bite Force Measurement Device (BFMD) [48]. Bacteria existing on the tooth enamel dissolve the hard tissue of the teeth and can cause cavities. A study designed a graphene-based wireless sensor for bacteria recognition on the tooth enamel [49]. Dental implants find application majorly for replacing missing teeth and mechanical factors play a key role in their failure. Therefore, it is very important to evaluate key parameters like frequency of resonance, allowable stress and strain (maximum), and fracture resistance for determination of implant durability [50]. A study uses a FBG sensor to evaluate strains at the surface of mandible produced due to impact or static loads and the sensor is verified by loading and recording uniaxial strain dynamically and statically [51]. Osseointegration assessment is another vital factor to prevent premature failure of the dental implant. A study performed the RFA (Resonance Frequency Analysis) for gauging the primary stability of the implant using artificial bone blocks composed of polyurethane foam and its FRF (Frequency Response function) characterized its structural resonance [52]. Two Hall effect sensors were used to gauge the response and excitation signals wherein a dual inductance device provided an excitement source to the bone-implant assembly. Another study highlighted a Non-contacted RF (Radio Frequency) detection device made with a loudspeaker which generates excitation in the acoustic form and for structure’s vibration analysis they used a capacitance sensor [46]. The setup of this device is shown in Fig. 6. Accuracy check of each model was analysed with “Osstell” Mentor device. Another study on RFA for dental implants talks about a “L-shaped” transducer which is compact and is fixed to abutment or the implant by a fastener [53]. Hence, use of sensors for measuring various mechanical parameters of dental implants has caused advances in their technology. Dentists and implantologists benefit greatly due to accurate and non-invasive techniques for monitoring dental health that has developed due to integration of these sensors in biomedical devices.

2.6.2

Robotic Surgery

Robotic surgery has replaced conventional surgical techniques, and thus, many technological advances have taken place in the sensors that assist in these surgeries. Force

Recent Advances in Sensor Technology for Biomedical Applications: …

49

Fig. 6 Non-contacted RF (Radio Frequency) detection device [46]

sensing is vital in increasing the competence of robots in multifaceted environments like the medical field [54]. In micro-surgery of the retina, assistance from robots can solve everyday problems encountered during the routine of micro-manipulations. These problems include tool access, shaking of hands, limits to the resolution for manipulation of tools, target motion, visualization of the target, resolution and use of non-injurious but healing forces [2]. Several sensors for detecting force are used in micro-manipulation, micro-surgery and nominally invasive surgery. In Vitreoretinal surgery (VRS) which is robot-assisted, force feedback from tool-tip can deliver the essential information to help a surgeon perform all manoeuvres, efficiently decrease forces with enhanced precision and possibly increase the efficiency and safety of the procedure as shown in Fig. 7 [47]. In a study, a novel cohesive system consisting of a tremor-cancelling, active and hand-held micro forceps with motorized sensing of force known as Micron was developed [55]. Another paper shows an enhanced bimorph, piezoelectric tactile sensor which can distinguish between soft materials having mechanical characteristics that are similar. The aim for this sensor is to develop surgical robotic tools which are intelligent for resection of brain tumour by means of integrated sensors [56]. Its function is to support neurosurgeons by creating tactile tools that are intra-operative to advance tissue differentiation and delineation. Another study presents the implementation of a tactile or force sensor called HexaTactile which has a 3-DOF (Degree of Freedom) parallel mechanism which is dissociated for Robot-Human interaction based on 6 MEMS barometers array of soft tactile sensors [54]. Another research fabricated a novel MRE (Magnetorheological) elastomer, which produces small movements beneath magnetic and electric fields, and it is used to sense the angles of bending and movements for flexible robots in biomedical applications [57].

2.6.3

Cancer Diagnosis

Recognition of cancer at an initial stage is crucial for cancer treatment and for reducing mortality rate. Hence, new, quick and easy methods for detection of cancer cells continue to be challenging [58]. Previously, the common approaches for analysis of cancer contain biopsy examination, tumour imaging and recognizing biomarkers of cancer. Fluorescent sensors based on Aptamers (single-stranded and short DNA or RNA molecules capable of identifying many targets with specificity) have been developed for detecting cancer biomarkers [59]. Another article developed a (PDA

50

N. Karnik et al.

Fig. 7 The sensor must be near the tip of the tool so that it can remain in the eye during vitreoretinal surgery [55]

NPs) Polydopamine nanoparticles sensor for free of label electrochemical (A-5490) lung cancer cell detection, and its capability for detection was also investigated [58]. Antigens associated with cancer have (CDE) Cancer-Derived Exosomes on the surface, and exosome detection biosensors facilitate early detection in individuals from resource-limited areas who formerly did not have access to innovative tools for diagnosis [15]. Another paper presented an innovative microcantilever sensor for detecting cancer prematurely with the help of a combined ring resonator [60]. In this study, the reaction of the antibody and antigen will cause the mass of tumour biomarkers on the microcantilever to increase. The sensing action will be performed due to this increase. Thus, a lot of progress has been made in sensor technologies for detecting cancer using cancer biomarkers.

2.6.4

Diabetes Mellitus Diagnosis

The growing number of people affected by diabetes mellitus each year has caused an increase in the use of biosensors for glucose-sensing technologies [12]. This section discusses the novel and innovative sensors that have been developed for non-invasive and accurate glucose sensing. A certain study developed a nonenzymatic, disposable, simple, fast and inexpensive sensor using a disposable PGE (pencil graphite electrode) improved by [Cu(NP)] nanoparticles of copper which showed a peak of absorption at 560 nm [45]. There is a variation in the composition of tears due to metabolic and ocular factors and tears are thus used for the evaluation of health. In another study, contact lenses made of microfluid were made for analysing in situ tear glucose, protein, pH, and sensing of nitrite ions as shown in Fig. 8 [61]. Another paper presents a monitoring system for glucose which is not just real time but can also reference its electrochemical potential with an electrode made of wired glucose dehydrogenase [62]. It evaluates changes in the glucose concentration between 3.5

Recent Advances in Sensor Technology for Biomedical Applications: …

51

Fig. 8 Microfluidic lens developed for in situ tear analysis [61]

and 8.5 mM. These recent advances have proved extremely useful in continuous and real-time monitoring of glucose which has helped patients with diabetes.

2.6.5

Sensors for Battle Against COVID-19

The coronavirus disease presents many problems for its fast clinical diagnosis due to its varied symptoms including loss of sense of smell and taste, headache, breathlessness, high fever, throat ache and diarrhoea and in some cases lack of any symptoms [63]. As of now, the sole established gold standard is the Reverse transcriptase (RT-PCR) test that uses the mechanism of nucleic acid sequencing for definitive detection of COVID-19 [64]. However, for rapid detection of the virus, various biomedical sensors and systems have been developed in the last two years. An article proposes the development of an optical fibre sensor for the detection of the virus using the concept of Evanescent wave absorbance (EWA) technology [65]. This sensor detects the elevated levels of immunoglobin M (IgM) and immunoglobin G (IgG) biomarkers caused by an exposure to the SARS-CoV-2 virus. The infection can also be detected by affinity-based biosensors such as the DPV-Based sensor which uses calixarene functionalized graphene oxide for targeting RNA of SARS-CoV-2 and Plasmonic-based sensor which is built on the basics of surface plasmon resonance (SPR) phenomenon [66]. Another study proposed the use of built-in smartphone sensors such as colour, humidity, cameras, temperature, proximity, inertial and wireless sensors to develop an AI-based framework for rapid detection of the virus [67]. Research was also conducted on detection of the virus in both symptomatic and asymptomatic cases using available wearable sensors and integrating them with Deep Neural Networks (DNN) [68]. The DNNs were trained from data obtained from 87 subjects who were either healthy, COVID positive (symptomatic) or COVID positive (asymptomatic) and were deployed on devices like smartphones and smartwatches. Thus, various biosensors, wearable sensors and optical fibre sensors have been developed for faster diagnosis of the coronavirus. However, these tools

52

N. Karnik et al.

are still not commercially available in the market and hence more research needs to be carried out to overcome various technological challenges to make these sensors and systems viable for mass production [69].

2.7 Advances in Sensor Technology The use of sensors has helped in providing faster, non-contact and non-invasive methods of diagnosis, monitoring and treatment of diseases and medical conditions as compared to traditional methods. A comparative study between modern systems comprising novel sensors and traditional methods has been made as shown in Table 4. Thus, in terms of technical parameters, the modern sensors that are developed have a smaller size, better quality, are thermally and chemically stable and support wireless communication and high rate of data transfer. They can also be implantable and have improved noise cancellation techniques. Their sensitivity is higher, and they have optimized timing which makes them perfect for real-time monitoring of biomedical parameters. Table 4 Comparison of traditional techniques and novel sensors developed for detection of various medical parameters Sr. No.

Medical parameters

Traditional technique

Novel sensors developed

1

Heart rate

Heart rate monitor, electrocardiogram (ECG)

Wearable sensors [70], Fibre optic sensors [71], Biosensors [72], Flexible sensors [73]

2

Blood pressure

Sphygmomanometer

Wearable sensors [74], Fibre optic sensors [75], Biosensors [76]

3

Glucose levels

Traditional glucometer (Finger prick test)

MEMS sensors [4], Biosensors [77]

4

Blood urea

Blood urea nitrogen (BUN) test

MEMS sensors [4]

5

Temperature

Thermometer

Wearable sensors [22], Fibre optic sensors [75]

6

Human motion analysis

Gait analysis

Wearable sensors [78], Fibre optic sensors [79], Flexible sensors [80]

Recent Advances in Sensor Technology for Biomedical Applications: …

53

References 1. Nag A, Afasrimanesh N, Feng S, Mukhopadhyay SC (2018) Strain induced graphite/PDMS sensors for biomedical applications. Sens Actuators A Phys 271:257–269. https://doi.org/10. 1016/j.sna.2018.01.044 2. Tosi D, Poeggel S, Iordachita I, Schena E (2018) Fiber optic sensors for biomedical applications. Elsevier Inc., 3. Wan H et al (2020) Biomedical sensors. Elsevier Inc., https://doi.org/10.1016/B978-0-12-816 034-3.00002-X 4. Sonetha V, Agarwal P, Doshi S, Kumar R, Mehta B (2017) Microelectromechanical systems in medicine. J Med Biol Eng 37(4):580–601. https://doi.org/10.1007/s40846-017-0265-x 5. Yin MJ, Gu B, An QF, Yang C, Guan YL, Yong KT (2018) Recent development of fiberoptic chemical sensors and biosensors: mechanisms, materials, micro/nano-fabrications and applications. Coord Chem Rev 376:348–392. https://doi.org/10.1016/j.ccr.2018.08.001 6. Yhuwana YGY, Apsari R, Yasin M (2017) Fiber optic sensor for heart rate detection. Optik 134:28–32. https://doi.org/10.1016/j.ijleo.2017.01.035 7. Cheng M et al (2020) A review of flexible force sensors for human health monitoring. J Adv Res. https://doi.org/10.1016/j.jare.2020.07.001 8. Mehrotra P (2016) Biosensors and their applications—a review. J Oral Biol Craniofac Res 6(2):153–159. https://doi.org/10.1016/j.jobcr.2015.12.002 9. Acker CD, Yan P, Loew LM (2020) Recent progress in optical voltage-sensor technology and applications to cardiac research: from single cells to whole hearts. Prog Biophys Mol Biol 154:3–10. https://doi.org/10.1016/j.pbiomolbio.2019.07.004 10. An Y et al (2020) A novel tetrapeptide fluorescence sensor for early diagnosis of prostate cancer based on imaging Zn2+ in healthy versus cancerous cells. J Adv Res 24:363–370. https://doi. org/10.1016/j.jare.2020.04.008 11. Nguyen HH, Lee SH, Lee UJ, Fermin CD, Kim M (2019) Immobilized enzymes in biosensor applications. Materials (Basel) 12(1):1–34. https://doi.org/10.3390/ma12010121 12. Ali J, Najeeb J, Asim Ali M, Farhan Aslam M, Raza A (2017) Biosensors: their fundamentals, designs, types and most recent impactful applications: a review. J Biosens Bioelectron 08(01):1– 9. https://doi.org/10.4172/2155-6210.1000235 13. Maduraiveeran G, Sasidharan M, Ganesan V (2018) Electrochemical sensor and biosensor platforms based on advanced nanomaterials for biological and biomedical applications. Biosens Bioelectron 103(Decem 2017):113–129. https://doi.org/10.1016/j.bios.2017.12.031 14. Singh S, Kumar V, Dhanjal DS, Datta S (2020) Biological biosensors for monitoring and diagnosis. Microb Biotechnol 317–335. https://doi.org/10.1007/978-981-15-2817-0 15. Cheng N et al (2019) Recent advances in biosensors for detecting cancer-derived exosomes. Trends Biotechnol 37(11):1236–1254. https://doi.org/10.1016/j.tibtech.2019.04.008 16. Hoare D, Bussooa A, Neale S, Mirzai N, Mercer J (2019) The future of cardiovascular stents: bioresorbable and integrated biosensor technology. Adv Sci 6(20). https://doi.org/10.1002/ advs.201900856 17. Balahura LR, Stefan-Van Staden RI, Van Staden JF, Aboul-Enein HY (2019) Advances in immunosensors for clinical applications. J Immunoassay Immunochem 40(1):40–51. https:// doi.org/10.1080/15321819.2018.1543704 18. Bai Y, Xu T, Zhang X (2020) Graphene-based biosensors for detection of biomarkers. Micromachines 11(1). https://doi.org/10.3390/mi11010060 19. Kaewkannate K, Kim S (2016) A comparison of wearable fitness devices. BMC Public Health 16(1). https://doi.org/10.1186/s12889-016-3059-0 20. Zhou H et al (2020) Stretchable piezoelectric energy harvesters and self-powered sensors for wearable and implantable devices. Biosens Bioelectr 168(May):112569. https://doi.org/10. 1016/j.bios.2020.112569 21. Nasiri S, Khosravani MR (2020) Progress and challenges in fabrication of wearable sensors for health monitoring. Sens Actuators A Phys 312:112105. https://doi.org/10.1016/j.sna.2020. 112105

54

N. Karnik et al.

22. Zhang SH, Wang FX, Li JJ, Peng HD, Yan JH, Pan GB (2017) Wearable wide-range strain sensors based on ionic liquids and monitoring of human activities. Sensors (Switzerland) 17(11):1–10. https://doi.org/10.3390/s17112621 23. Jung HC et al (2012) CNT/PDMS composite flexible dry electrodes for long-term ECG monitoring. IEEE Trans Biomed Eng 59(5):1472–1479. https://doi.org/10.1109/TBME.2012.219 0288 24. Boutry CM, Nguyen A, Lawal QO, Chortos A, Rondeau-Gagné S, Bao Z (2015) A sensitive and biodegradable pressure sensor array for cardiovascular monitoring. Adv Mater 27(43):6954– 6961. https://doi.org/10.1002/adma.201502535 25. Park SJ, Jeon JY, Kang BC, Ha TJ (2020) Wearable temperature sensors based on lanthanumdoped aluminum-oxide dielectrics operating at low-voltage and high-frequency for healthcare monitoring systems. Ceram Int (Sept 2020). https://doi.org/10.1016/j.ceramint.2020.10.023 26. Wu X et al (2015) Thermally stable, biocompatible, and flexible organic field effect transistors and their application in temperature sensing arrays for artificial skin. Adv Func Mater 25(14):2138–2146. https://doi.org/10.1002/adfm.201404535 27. Mohan AMV, Rajendran V, Mishra RK, Jayaraman M (2020) Recent advances and perspectives in sweat based wearable electrochemical sensors. TrAC Trends Anal Chem 131:116024. https:// doi.org/10.1016/j.trac.2020.116024 28. Mohan AMV, Windmiller JR, Mishra RK, Wang J (2017) Continuous minimally-invasive alcohol monitoring using microneedle sensor arrays. Biosens Bioelectron 91:574–579. https:// doi.org/10.1016/j.bios.2017.01.016 29. Gao W et al (2016) Wearable microsensor array for multiplexed heavy metal monitoring of body fluids. ACS Sens 1(7):866–874. https://doi.org/10.1021/acssensors.6b00287 30. Manjakkal L, Sakthivel B, Gopalakrishnan N, Dahiya R (2018) Printed flexible electrochemical pH sensors based on CuO nanorods. Sens Actuators B Chem 263:50–58. https://doi.org/10. 1016/j.snb.2018.02.092 31. Ge Y, Cai K, Wang T, Zhang J (2018) MEMS pressure sensor based on optical Fabry–Perot interference. Optik 165:35–40. https://doi.org/10.1016/j.ijleo.2018.03.112 32. Vilela D, Romeo A, Sánchez S (2016) Flexible sensors for biomedical technology. Sens Actuators A Phys 16(3):402–408. https://doi.org/10.1039/c5lc90136g 33. Abbasnejad B, Thorby W, Razmjou A, Jin D, Asadnia M, Ebrahimi Warkiani M (2018) MEMS piezoresistive flow sensors for sleep apnea therapy. Sens Actuators A Phys 279:577–585. https:// doi.org/10.1016/j.sna.2018.06.038 34. Manikandan N, Muruganand S, Divagar M, Viswanathan C (2019) Design and fabrication of MEMS based intracranial pressure sensor for neurons study. Vacuum 163(Nov 2018):204–209. https://doi.org/10.1016/j.vacuum.2019.02.018 35. Kuwana K, Nakai A, Masamune K, Dohi T (2013) A grasping forceps with a triaxial MEMS tactile sensor for quantification of stresses on organs. In: Proceedings of the annual international conference of the IEEE engineering in medicine and biology society, EMBS, pp 4490–4493. https://doi.org/10.1109/EMBC.2013.6610544 36. Afsarimanesh N, Nag A, Sarkar S, Sabet GS, Han T, Mukhopadhyay SC (2020) A review on fabrication, characterization and implementation of wearable strain sensors. Sens Actuators A Phys 112355. https://doi.org/10.1016/j.sna.2020.112355 37. Tosi D (2015) Advanced interrogation of fiber-optic Bragg grating and Fabry–Perot sensors with KLT analysis. Sensors (Switzerland) 15(11):27470–27492. https://doi.org/10.3390/s15 1127470 38. Bremer K, Lewis E, Leen G, Moss B, Lochmann S, Mueller IAR (2012) Feedback stabilized interrogation technique for EFPI/FBG hybrid fiber-optic pressure and temperature sensors. IEEE Sens J 12(1):133–138. https://doi.org/10.1109/JSEN.2011.2140104 39. Liao N et al (2020) A sensitivity-enhanced micro-cavity extrinsic Fabry–Perot interferometric fiber-optic curvature sensor. Optik 221(May):165310. https://doi.org/10.1016/j.ijleo. 2020.165310 40. Xu F et al (2012) High-sensitivity Fabry–Perot interferometric pressure sensor based on a nanothick silver diaphragm. Opt Lett 37(2):133. https://doi.org/10.1364/ol.37.000133

Recent Advances in Sensor Technology for Biomedical Applications: …

55

41. Arkwright JW, Underhill I (2017) Fibre Bragg grating manometry catheters for in vivo monitoring of peristalsis, 1005410. https://doi.org/10.1117/12.2255645 42. Stephens AF, Busch A, Salamonsen RF, Gregory SD, Tansley GD (2019) A novel fibre Bragg grating pressure sensor for rotary ventricular assist devices. Sens Actuators A Phys 295:474– 482. https://doi.org/10.1016/j.sna.2019.06.028 43. Hosseindokht Z, Mohammadpour R, Asadian E, Paryavi M, Rafii-Tabar H, Sasanpour P (2020) Low cost flexible pressure sensor using laser scribed GO/RGO periodic structure for electronic skin applications. Superlattices Microstruct 140(Feb):106470. https://doi.org/10.1016/j.spmi. 2020.106470 44. Han Y, Fang Y, Ding X, Liu J, Jin Z, Xu Y (2020) A simple and effective flexible electrochemiluminescence sensor for lidocaine detection. Electrochem Commun 116:106760. https://doi. org/10.1016/j.elecom.2020.106760 45. Pourbeyram S, Mehdizadeh K (2016) Nonenzymatic glucose sensor based on disposable pencil graphite electrode modified by copper nanoparticles. J Food Drug Anal 24(4):894–902. https:// doi.org/10.1016/j.jfda.2016.02.010 46. Pan MC, Zhuang HB, Chen CS, Wu JW, Lee SY (2013) A noncontact resonance frequency detection technique for the assessment of the interfacial bone defect around a dental implant. Med Eng Phys 35(12):1825–1830. https://doi.org/10.1016/j.medengphy.2013.05.006 47. He X, Handa J, Gehlbach P, Taylor R, Iordachita I (2014) A submillimetric 3-DOF force sensing instrument with integrated fiber Bragg grating for retinal microsurgery. IEEE Trans Biomed Eng 61(2):522–534. https://doi.org/10.1109/TBME.2013.2283501 48. Padma S, Umesh S, Asokan S, Srinivas T (2017) Bite force measurement based on fiber Bragg grating sensor. J Biomed Opt 22(10):1. https://doi.org/10.1117/1.jbo.22.10.107002 49. Mannoor MS et al (2012) Graphene-based wireless bacteria detection on tooth enamel. Nat Commun 3. https://doi.org/10.1038/ncomms1767 50. Tatlisoz MM, Canpolat C (2017) Mechanical testing strategies for dental implants. https://doi. org/10.1007/978-981-10-4166-2 51. Carvalho L, Silva JCC, Nogueira RN, Pinto JL, Kalinowski HJ, Simões JA (2006) Application of Bragg grating sensors in dental biomechanics. J Strain Anal Eng Des 41(6):411–416. https:// doi.org/10.1243/03093247JSA191 52. Cong TM, Mou RZ, Chen CS, Pan MC (2015) Resonance frequency confirmation for osseointegration of dental implantation in vitro models with varied cortical thickness. J Med Devices Trans ASME 9(3):2015–2017. https://doi.org/10.1115/1.4030599 53. Kittur N, Oak R, Dekate D, Jadhav S, Dhatrak P (2020) Dental implant stability and its measurements to improve osseointegration at the bone-implant interface: a review. Mater Today Proc. https://doi.org/10.1016/j.matpr.2020.08.243 54. Hamed A, Masouleh MT, Kalhor A (2019) Design and characterization of a bio-inspired 3-DOF tactile/force sensor and implementation on a 3-DOF decoupled parallel mechanism for humanrobot interaction purposes. Mechatronics 66(May 2019):102325. https://doi.org/10.1016/j.mec hatronics.2020.102325 55. Schmitt, Segert (2008) Towards robot-assisted vitreoretinal surgery: force-sensing microforceps integrated with a handheld micromanipulator. Bone 23(1):1–7. https://doi.org/10.1109/ ICRA.2014.6907035 56. Oliva Uribe D, Schoukens J, Stroop R (2018) Improved tactile resonance sensor for robotic assisted surgery. Mech Syst Signal Process 99:600–610. https://doi.org/10.1016/j.ymssp.2017. 07.007 57. Banerjee H, Ren H (2018) Electromagnetically responsive soft-flexible robots and sensors for biomedical applications and impending challenges, 43–72. https://doi.org/10.1007/978-98110-6035-9_3 58. Bolat G, Vural OA, Yaman YT, Abaci S (2021) Polydopamine nanoparticles-assisted impedimetric sensor towards label-free lung cancer cell detection. Mater Sci Eng C 119(May 2020):111549. https://doi.org/10.1016/j.msec.2020.111549 59. Zhao X et al (2020) Aptamer-based fluorescent sensors for the detection of cancer biomarkers. Spectrochim Acta Part A: Mol Biomol Spectrosc 119038. https://doi.org/10.1016/j.saa.2020. 119038

56

N. Karnik et al.

60. Upadhyaya AM, Srivastava MC, Sharan P (2020) Integrated MOEMS based cantilever sensor for early detection of cancer. Optik 165321. https://doi.org/10.1016/j.ijleo.2020.165321 61. Moreddu R, Wolffsohn JS, Vigolo D, Yetisen AK (2020) Laser-inscribed contact lens sensors for the detection of analytes in the tear fluid. Sens Actuators B: Chem 317(March):128183. https://doi.org/10.1016/j.snb.2020.128183 62. Ramašauskas L, Meškys R, Ratautas D (2020) Real-time glucose monitoring system containing enzymatic sensor and enzymatic reference electrodes. Biosens Bioelectron 164(May 2020). https://doi.org/10.1016/j.bios.2020.112338 63. Dave PK, Rojas-Cessa R, Dong Z, Umpaichitra V (2021) Survey of saliva components and virus sensors for prevention of COVID-19 and infectious diseases. Biosensors 11(1). https:// doi.org/10.3390/bios11010014 64. Behera S et al (2020) Biosensors in diagnosing COVID-19 and recent development. Sens Int 1(October):100054. https://doi.org/10.1016/j.sintl.2020.100054 65. Nag P, Sadani K, Mukherji S (2020) Optical fiber sensors for rapid screening of COVID-19. Trans Indian Natl Acad Eng 5(2):233–236. https://doi.org/10.1007/s41403-020-00128-4 66. Drobysh M, Ramanaviciene A, Viter R, Ramanavicius A (2021) Affinity sensors for the diagnosis of COVID-19. Micromachines 12(4):390. https://doi.org/10.3390/mi12040390 67. Maghded HS, Ghafoor KZ, Sadiq AS, Curran K, Rawat DB, Rabie K (2020) A novel AIenabled framework to diagnose coronavirus COVID-19 using smartphone embedded sensors: design study. In: Proceedings—2020 IEEE 21st international conference on information reuse and integration for data science, IRI 2020, pp 180–187. https://doi.org/10.1109/IRI49571.2020. 00033 68. Hassantabar S et al (2020) CovidDeep: SARS-CoV-2/COVID-19 test based on wearable medical sensors and efficient neural networks. arXiv 1–11 69. Bhalla N, Pan Y, Yang Z, Payam AF (2020) Opportunities and challenges for biosensors and nanoscale analytical tools for pandemics: COVID-19. ACS Nano 14(7):7783–7807. https:// doi.org/10.1021/acsnano.0c04421 70. Xiao N, Yu W, Han X (2020) Wearable heart rate monitoring intelligent sports bracelet based on internet of things. Meas: J Int Meas Confederation 164:108102. https://doi.org/10.1016/j. measurement.2020.108102 71. Samartkit P, Pullteap S, Seat HC (2021) Validation of fiber optic-based Fabry–Perot interferometer for simultaneous heart rate and pulse pressure measurements. IEEE Sens J 21(5):6195–6201. https://doi.org/10.1109/JSEN.2020.3041782 72. Coffen B, Scott P, Mahmud MDS (2020) Real-time wireless health monitoring: an ultralow power biosensor ring for heart disease monitoring. In: 2020 International conference on computing, networking and communications, ICNC 2020, pp 626–630. https://doi.org/10.1109/ ICNC47757.2020.9049814 73. Kaidarova A et al (2020) Laser-printed, flexible graphene pressure sensors. Glob Challenges 4(4):2000001. https://doi.org/10.1002/gch2.202000001 74. Wang TW, Lin SF (2020) Wearable piezoelectric-based system for continuous beat-to-beat blood pressure measurement. Sensors (Switzerland) 20(3):1–12. https://doi.org/10.3390/s20 030851 75. Poduval RK, Coote JM, Mosse CA, Finlay MC, Desjardins AE, Papakonstantinou I (2021) Precision-microfabricated fiber-optic probe for intravascular pressure and temperature sensing. IEEE J Sel Top Quantum Electron 27(4). https://doi.org/10.1109/JSTQE.2021.3054727 76. Jegan R, Jose PSH, Rajalakshmy P, Raj PAC, Kanmani HJ, Nimi WS (2021) Methodological role of mathematics to estimate human blood pressure through biosensors. Int J Integr Eng 13(1):240–248. https://doi.org/10.30880/ijie.2021.13.01.021 77. Zhao L, Wen Z, Jiang F, Zheng Z, Lu S (2020) Silk/polyols/GOD microneedle based electrochemical biosensor for continuous glucose monitoring. RSC Adv 10(11):6163–6171. https:// doi.org/10.1039/c9ra10374k 78. Felix F et al (2020) Objective sensor-based gait measures reflect motor impairment in multiple sclerosis patients: reliability and clinical validation of a wearable sensor device. Multiple Sclerosis Relat Disord 39. https://doi.org/10.1016/j.msard.2019.101903

Recent Advances in Sensor Technology for Biomedical Applications: …

57

79. Monge J, Postolache O, Alexandre R, De Fatima Domingues M, Antunes P, Viegas V (2020) Fiber bragg gratings solution for gait assessement. In: I2MTC 2020—international instrumentation and measurement technology conference, proceedings, pp 1–6. https://doi.org/10.1109/ I2MTC43012.2020.9128421 80. Carbonaro N, Arcarisi L, Di Rienzo F, Virdis A, Vallati C, Tognetti A (2020) A preliminary study on a new lightweight and flexible sensing sock for gait analysis. In: Proceedings of IEEE sensors, vol 2020-Oct, pp 14–17. https://doi.org/10.1109/SENSORS47125.2020.9278682

Performance Analysis of Diode Clamped and Flying Capacitor Multilevel Matrix Converter Used for DFIG-Based Wind System G. Pandu Ranga Reddy, D. Mahesh Kumar, K. Rajesh, Y. Chintu Sagar, and J. Nageswara Rao Abstract It is impossible to meet the increasing demand for electrical energy with the available deposits of fossils fuels that are already at the edge of depletion. Harnessing energy by means of alternative energy sources has been initiated in order to meet increasing energy demand in the future. Wind energy is found out to be more potential resource than all other renewable sources. The converters that are presented in the wind system play a major role for improving the quality of power. The test system comprises matrix converter attached to DFIG machine and the rotor of generator is coupled to the grid. The conventional matrix converter is then replaced with the proposed multilevel matrix converters. THDs of the voltages and currents at generating and load points are evaluated for the DFIG machine and it can run by diode clamped and flying capacitor multilevel matrix converters. For managing the switching devices in the converter circuit, the space vector pulse width modulation approach is devised. MATLAB/Simulink software is used to graphically show and evaluate the stator voltages and currents of the DFIG machine. Keywords DFIG—double fed induction generator · DCMMC—diode clamped multilevel matrix converter · FCMMC—flying capacitor multilevel matrix converter · FFT—fast Fourier transformation · THD—total harmonic distortion · WECS—wind energy conversion system G. P. R. Reddy (B) Deparment of EEE, G. Pullaiah College of Engineering and Technology (Autonomous), Kurnool, India e-mail: [email protected] D. M. Kumar Deparment of EEE, PVKK Institute of Technology, Anantapur, India K. Rajesh Deparment of EEE, RGM College of Engineering and Technology, Nandyal, India Y. C. Sagar Deparment of EEE, Ashoka Women’s Engineering College, Kurnool, India J. N. Rao Deparment of Electrical and Computer Engineering, Mizan-Tepi University, Teppi, Ethiopia © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_4

59

60

G. P. R. Reddy et al.

1 Introduction A renewable energy sources are now used as alternatives for conventional sources of energy and employed into power grid. Distributed generation also known as decentralized generation involves generation in lower capacities by conventional sources in assistance with renewable sources [1]. Wind energy systems are widely used in distributed generation in addition to conventional sources. Wind energy generation systems are more advantageous than other renewable sources in terms of efficiency and generational capabilities [2]. Static elements such as solar power systems result in PWM output voltage. But wind energy system employed with dynamic machines such as synchronous and induction machines result in nearly sinusoidal waveform. DFIG machine is usually more preferable for a wind energy conversion system as it provides constant speed irrespective of variations in the wind speed [1]. By controlling power electronic converters are regulation of rotor input is controlled. Figure 1 shows the interconnection of DFIG machine with grid and it can operate by power electronics converter. Rotor of DFIG is usually fed by suitable two-stage VSIs having a common DC link capacitor. Cycloconverters are another choice for a wind energy system based on DFIG. Pulse width modulation is used to manage power electronic converters (PWM). Conventional two or three level PWM for power electronic converters leads to increased THD in rotor currents. This in turn results in increased harmonics in stator currents thus injecting harmonics into the grid. Thus, harmonics at source side has adverse affects on loads connected to grid and DFIG [3]. Performance of DFIG can be improved by reducing the harmonic content which can be achieved by replacement of 2-level and 3-level voltage source inverters by

Fig. 1 DFIG interconnection to grid with power electronic converter

Performance Analysis of Diode Clamped and Flying Capacitor …

61

multilevel matrix converters. Several configurations of multilevel matrix converters are available. It is important to choose the right configuration of converter that results in lower harmonic distortion and is feasible with DFIG-based wind system. When compared to other topologies, the diode clamped multilevel matrix converter and the flying capacitor multilevel matrix converter have reduced harmonic distortion.

2 Concept of Wind Turbine and Wind Generator 2.1 Modeling of Wind Turbine The wind turbine generates mechanical energy by converting into kinetic energy [1]. Equation (1) signifies mechanical power of the turbine [3]. Pm =

1 ρCp (λ, β)Ar Vw3 2

(1)

where ρ Vw Cp λ β

is the density of air (Kgm3 ) is the speed of the wind (M/S) is the coefficient of power is the ratio of tip-speed is the pitch angle of the rotor blades

and Ar is the rotor swept area of a wind turbine (m2 ).Here,   −16.5 116 Pm Cp (λ, β) = 0.5 − 0.4β − 5 e λi = λi PW

(2)

where λi =

1 1 λ+0.089β

−

0.035 β 3 +1

(3)

Theoretically, C p value is set to 0.59. But according to Betz’s Law, the practical value is set to 0.2–0.4 and rotor pitch angle is a constant [4].

2.2 Modeling of DIF Wind Generator See Fig. 2.

62

G. P. R. Reddy et al.

Fig. 2 Equivalent circuit diagram of DFIG

Their corresponding mathematical equations are [5] Vds = Rs i ds +

dψds − ωs ψqs dt

(4)

Vqs = Rs i qs +

dψqs + ωs ψds dt

(5)

Vdr = Rs i dr +

dψdr − (ωs − ωr )ψqr dt

(6)

Vqr = Rs i qr +

dψqr + (ωs − ωr )ψdr dt

(7)

Here, ψ denotes the flux linkage and it can be expressed as: ψds = L ss i ds + L m i dr

(8)

ψqs = L ss i qs + L m i qr

(9)

ψdr = L rr i dr + L m i ds

(10)

ψqr = L rr i qr + L m i qs

(11)

L ss = L s + L m

(12)

L rr = L r + L m

(13)

and also

Performance Analysis of Diode Clamped and Flying Capacitor …

63

Rotor swing equation expressed in terms of inertia constant J and damping factor F is J

dωrm + Fωrm = (Te − Tm ) dt

(14)

3 Wind Converter Back-to-back, VSC converters are the most common converters used in technologies for converting wind energy. This research paper focuses on performance of two multilevel matrix converters: the flying capacitor multilevel matrix converter and the diode clamped multilevel matrix converter, which were utilized in place of a back-to-back converter [6].

3.1 Concept of Diode Clamped Multilevel Matrix Converter Figure 3 depicts the conventional topology of diode clamped multilevel matrix converter [7]. This topology is complex and voltage regulation in a 3-bus system would complicate the modulation strategy. For obtaining higher levels of output voltage, practical implementation of this configuration is not a promising solution [8]. A modified DCMMC topology which does not exhibit aforestated problems is introduced in this paper. As exposed in Fig. 4, this architecture consists of a traditional 2-level converter on the input side for AC–DC conversion and a multilevel converter on the output side for DC–AC conversion.

3.2 Control Method for Diode Clamped Multilevel Matrix Converter Both output voltage and input current are adjusted via space vector pulse width modulation. This strategy is efficient and beneficial in allowing input current and output voltage switching vectors to be selected. This technique is helpful for decreasing the harmonics presented in the system. For an adequate small interval of time, the reference voltage vector may be anything you want it to be calculated with help of stationary vectors that are produced by the matrix converter. This period of time is called the converter sample time [9].When the reference voltage vector is rotated to a different angular point at the next sampling moment, a novel set of voltage vectors that are stationary is formed. The average output voltage would safely follow

64

G. P. R. Reddy et al.

Fig. 3 Conventional topology of DCMMC

the reference voltage if this procedure was extended beyond sampling the whole waveform of the required voltage vector being produced in sequence. On the other hand, the phase angle difference between input voltage and current is calculated using specified stationary vectors [10]. The reference signal is a mix of nonzero active switching vectors that are contiguous to the reference vector and one or more zero vectors with proper sampling time in this SVPWM approach. Figure 5 depicts the vector representation with space vectors, whereas Fig. 6 depicts the reference vector switching duration (T Z ) calculation as a combination of adjacent vector durations (T 1 , T 2 ) [11, 12]. Time duration of switching vector is calculated with the following equations

T1 =

 _  √   3.TZ .Vref  

T2 =

Vdc

 sin

 _  √   3.TZ .Vref   Vdc

π n−1 −α+ π 3 3



  n−1 sin α − π 3

T0 = Tz − T1 − T2

(15)

(16) (17)

Performance Analysis of Diode Clamped and Flying Capacitor …

Fig. 4 Proposed diode clamped multilevel matrix converter

Fig. 5 Switching vector diagram of SVPWM technique

65

66

G. P. R. Reddy et al.

Fig. 6 Arrangement of adjacent vectors as a reference vector

3.3 Concept of Flying Capacitor Multilevel Matrix Converter Conventional matrix converter with nine switches can be replaced by flying capacitor multilevel matrix converter [13]. Figure 7 shows the flying capacitor multilevel matrix. As shown in Fig. 7 for achieving freewheeling operation, two MOSFETs are connected antiparallel to diodes connected in antiparallel. Also, the switch count of this topology is doubled as match up to usual matrix converter [14]. For attaining proper voltage division and to generate multilevel output voltage, it is necessary to choose six capacitors having equal rating. PWM signals required for the operation of proposed flying capacitor multilevel matrix converter are generated by providing identical PWM signal to all the bidirectional switches [15, 16].

3.4 Control Method for Flying Capacitor Multilevel Matrix Converter The suggested converter uses a space vector pulse width modulation (SVPWM) control approach to produce PWM signals for the 18-bidirectional switches [9]. Clarke’s transformation is used to the sinusoidal signal as a reference fundamental signal to create space vector control signal via sector selection beginning with the alpha and beta waveforms [17, 18]. ⎡ i (t) ⎤  2 1 −√21 −√21 ⎣ a ⎦ i αβ (t) = i b (t) 3 0 − 23 − 23 i c (t)

(18)

Performance Analysis of Diode Clamped and Flying Capacitor …

67

Fig. 7 Flying capacitor multilevel matrix converter topology

Alpha beta waveform represents the sectors to be taken for space vector trajectory. Figure 8 shows the sector selection by alpha and beta components. A zero vector is to be considered in addition to the six vectors for suitable pulse generation [13, 19]. Proper selection of sector aids to generation of suitable gating signals. The proposed converter switching states are presented in Table 1. It corresponds to nine switching states generating of Phase U resulting in multilevel inversion operation [20].

4 Simulation Results and Discussion In this paper, the test system for evaluation and comparison of DCMMC and FCMMC is given in Fig. 9. The test system consists of DFIG-based WECS employing matrix converter and it is supplying power to load. Up to t = 0.2 s, the wind energy conversion equipment was not linked to the source. After t = 0.2 s, the source is linked to the wind plant’s DFIG machine-based power electronic converter. The load is powered by a three-phase source up to t = 0.2 s. WECS uses DFIG with a power electronic converter to supply power to the RL load after t = 0.2 s. In the test system, it can be observed that conventional matrix converter is used to induce voltage into the rotor circuit. This converter will be replaced with the proposed multilevel matrix converters and THD calculations of the DFIG are carried out with FFT analysis tool in MATLAB/Simulink Power GUI software. By running the simulation with the space vector PWM technique, the THD values of voltages and currents in percent at different places such as B1, B2 and B3 are determined

68

G. P. R. Reddy et al.

Fig. 8 Switching sector calculation with alpha and beta components

Table 1 Switching states of one of the phase in flying capacitor multilevel matrix converter Changing mode (On: 1, Off: 0)

Voltage value at phase ‘A’

SAa1

SAa2

SBa1

SBa2

SCa1

SCa2

1

1

0

0

0

0

VA

1

0

0

1

0

0

(VA + VB )/2

1

0

0

0

0

1

(VA + Vc )/2

0

1

1

0

0

0

(VA + VB )/2

0

0

1

1

0

1

VB

0

0

1

0

0

1

(VB + Vc )/2

0

1

0

0

1

0

(VA + Vc )/2

0

0

0

1

1

0

(VB + Vc )/2

0

0

0

0

1

1

Vc

using MATLAB/Simulink software. Here, B1 is the three-phase source point or grid point, B2 is the DFIG machine connected point and B3 is the load point. In the test system, initially DCMMC is replaced in place of conventional matrix converter after FCMMC is replaced and compared the performance of the converters. Simulink model of the test system obtained by replacing of the basic matrix converter with diode clamped multilevel matrix is as exposed in Fig. 10. Figure 11 shows a test setup for a DIFG-based wind system developed by substituting the basic matrix converter with an FCMMC.

Performance Analysis of Diode Clamped and Flying Capacitor …

69

Fig. 9 Test system with DFIG fed with conventional matrix converter topology

Fig. 10 Diode clamped multilevel matrix converter fed DFIG wind generator

4.1 Simulation Results of Diode Clamped MMC Based on SVPWM for the DFIG The subsequent figures demonstrate the Simulink model results of DFIG-based wind system using SVPWM by diode clamped multilevel matrix converter. The %THD of DFIG source currents and voltages in WECS driven by DCMMC is shown in Fig. 12. It gives information of THD values in percent of source currents and voltages as 2.55 and 0.46 correspondingly. The percent THD values of DFIG voltages and currents under load in WECS operated by DCMMC are shown in Fig. 13. It gives

70

G. P. R. Reddy et al.

Fig. 11 Flying capacitor multilevel matrix converter fed DFIG wind generator

information of THD values in percent of load currents and voltages as 0.54 and 0.48 correspondingly. Figure 14 demonstrates the THD values of DFIG source currents and voltages in WECS employing FCMMC. Figure 14 gives the %THD values of DFIG source currents and voltages as 1.83 and 0.40 correspondingly. Figure 15 depicts the %THD rates of load currents and voltages of DFIG-based WECS by means of proposed

Fig. 12 Source voltage and current harmonic spectrum

Fig. 13 Load voltage and current harmonic spectrum

Performance Analysis of Diode Clamped and Flying Capacitor …

71

Fig. 14 Source voltage and current harmonic spectrum

Fig. 15 Load voltage and current harmonic spectrum

converter FCMMC. From Fig. 15, the load currents and voltages %THD are 0.41 and 0.39 correspondingly.

4.2 Simulation Results of a Flying Capacitor MMC Based on SVPWM for the DFIG THDs of voltages and currents in percentage are shown in Table 2 at various points in the system where suggested converters are used. Table 2 %THDs of DCMMC and FCMMC with SVPWM

Parameter

DCMMC

FCMMC

Point B1: source voltage

0.46

0.40

Points B1: source current

2.55

1.83

Point B2: integrated voltage

5.89

5.66

Point B2:integrated current

2.72

1.76

Point B3:load voltage

0.48

0.39

Point B3:load current

0.54

0.41

72

G. P. R. Reddy et al.

5 Conclusion The converters performance can be observed by modeling the system using MATLAB/Simulink software. In this paper, the diode clamped and flying capacitor multilevel matrix converter models are designed for DFIG wind power production systems, and the converters’ performance is evaluated. The voltage and current %THD values at different points in wind energy conversion system operated by DFIG machine are calculated with the help of FFT analysis. The flying capacitor multilevel matrix converter architecture is thought to be more competent than the diode clamped multilevel matrix converter, according to the FFT analysis tool.

References 1. Casadei D, Serra G, Tani A, Zarri L (2020) Matrix converter modulation strategies; a new general approach based on space-vector representation of the switch state. IEEE Trans Ind Electron 49(2) 2. Panwar V, Kaur T (2018) Overview of renewable energy resources of India. Int J Adv Res Electr Electron Instrum Eng 3(2):7118–7125 3. Reddy GPR, Kumar MV (2015) Analysis of wind energy conversion system employing DFIG with SPWM and SVPWM type converters. J Electr Eng 5(4):95–106 4. Reddy GPR, Kumar MV (2017) Two level versus matrix converters performance in wind energy conversion system employing DFIG. J Inst Eng (India): Ser B. Springer Publications, ISSN: 2250-2106 5. Pena R, Clare JC, Asher GM (1996) Doubly fed induction generator using back-to-back PWM converters and its application to variable speed wind-energy generation. IEE Proc-Electr Power Appl 143(3):231–241 6. Jacob JT (2014) Review on high power multilevel-matrix converters. Int J Adv Res Electr Electron Eng 3(1) 7. Khajehoddin SA, Bakhshai A, Jain P (2017) A sparse multilevel matrix converter based on diode-clamped topology. IEEE Trans 0197-2618/07/$25.00 © 2017 8. Reddy GPR, Kumar MV Diode clamped multilevel matrix converter for DFIG based wind energy conversion system. IOSR J Electr Electron Eng 44–52. ISSN: 2320-3331 9. Lee MY, Wheeler P, Klumpner C (2015) Space-vector modulated multilevel matrix converter. IEEE Trans Ind Electron 57(10):3385–3394 10. Sun Y, Xiong W, Su M, Li X, Dan H, Yang J (2014) Topology and modulation for a new multilevel diode-clamped matrix converter. IEEE Trans Power Electron 29(12) 11. Ghoni R, Abdalla AN (2014) Analysis and mathematical modeling of space vector modulated direct controlled matrix converter. J Theor Appl Inf Technol 27–35 12. Gruson F, Le Moigne P, Delarue P, Videt A, Cimetiére X, Arpillière M (2013) A simple carrier-based modulation for the SVM of the matrix converter. IEEE Trans Ind Inform 9(2) 13. Wheeler P, Xu L, Lee MY, Lee E, KIumpner C, Clare J (2019) A review of multi-level matrix converter topologies. In: Conference on power electronics, machines and control group, University of Nottingham, pp 286–290. Published in IEEE Digital Explore 14. Shi Y, Xu Y, He Q, Wang Z (2005) Research on a novel capacitor clamped multilevel matrix converter. IEEE Trans Power Electron 15. Miura Y, Inubushi K, Ito M, Ise T (2014) Multilevel modular matrix converter for high voltage applications. IEEE Digital Explore, 2014, pp 4690–4696 16. Shukla RD, Singh A, Singh SP (2012) Generators for variable speed wind energy conversion systems: a comparative study. IJESCC 3(2):103– 117

Performance Analysis of Diode Clamped and Flying Capacitor …

73

17. Xu L, Clare JC, Wheeler PW, Lee E, Li Y (2012) Capacitor clamped multilevel matrix converter space vector modulation. IEEE Trans Ind Electron 59(1):105–115 18. Odaka A, Sato I, Ohgushi H, Tamai Y, Mine H, Ito J (2005) A PAM: control method for matrix converter. In: Proceedings of the 2005 Japan industry applications society conference 19. Hojabri H, Mokhtari H, Chang L (2011) A generalized technique of modeling, analysis, and control of a matrix converter using SVD. IEEE Trans Ind Electron 58(3):949–959 20. Nguyen TD, Lee H-H (2014) Multilevel indirect matrix converter with carrier-based pulse width modulation. Published in IEEE Digital Explore, 2014, pp 3318–3323

Real Time Feedback System for Speech Dysfluency in Children Jennifer C. Saldanha and Rohan Pinto

Abstract Stuttering and Cluttering is a speech dysfluency disorder found in people which makes them to speak in a disordered way leading to prolongations, repetitions, pauses, and blocks. The main aim of this work is to develop a real time system that will help people with this disorder to improve their speech. The first phase of this work deals with the three methods that can treat the stutterers, i.e., delayed auditory feedback, frequency altered feedback, and metronome generation. The second phase deals with stuttered speech analysis to find the severity of stuttering using prolongation, repetition, and silence blocks. A model is developed using SoX audio processing toolbox. The processor used in the system is Raspberry Pi B. For the benefit of the user a GUI is implemented using Tkinter. A database is created with a given Kannada passage with 80 samples which includes both the stuttered and normal speech samples. Keywords Delayed auditory feedback · Frequency altered feedback · k-nearest neighbor · Mel frequency Cepstral coefficients

1 Introduction Stuttering and cluttering is commonly known as stammering. It is a speech dysfluency disorder. Here the flow of speech is interrupted by repetitions and prolongations of sounds, syllables, words, and phrases without conscious control. There are even involuntary silent pauses or blocks. In this disorder the person who has the problem of stuttering is unable to produce sounds. The impact of this disorder on a person’s emotional state of mind and normal functioning can be severe [1]. This may include J. C. Saldanha · R. Pinto (B) Department of Electronics and Communication, St Joseph Engineering College, Mangaluru, India e-mail: [email protected] J. C. Saldanha e-mail: [email protected] Affiliated to Visvesvaraya Technological University, Belagavi, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_5

75

76

J. C. Saldanha and R. Pinto

fears of enunciating specific vowels or consonants, there may also be a fear of being caught stuttering in a social gathering, isolating one-self, not being able to express ones feelings, shame, low self-esteem, etc. This disorder can be basically classified into developmental, psychogenic, and neurogenic stuttering. Developmental stuttering is found in all people in their early ages which reduces on its own in most cases but stuttering retains in some cases [2]. This is the most common type of stuttering. A normal person may start stuttering after being subjected to a shock or fear or such situations which is called psychogenic stuttering. A person may start stuttering after a neurological trauma which is called neurogenic stuttering. Out of these, developmental and psychogenic stuttering can be improved effectively by employing different methods whereas neurogenic stuttering cannot be improved. The objective of this work is to develop a real time device which facilitates the people suffering from stuttering and cluttering to reduce the disorder and improve their speech by using three techniques namely Delayed Auditory Feedback (DAF), Frequency Altered Feedback (FAF), and metronome generation. DAF is a technique used in cases where the speech of the stutterer is fed to his ears after a delay. In FAF method the person is made to hear his own pitch shifted sound while speaking and in metronome generation, a continuous impulse signal is generated at regular intervals according to which the person is made to speak. The paper is organized as follows: Sect. 2 gives a brief review of the previous work with the benefits and short comings of the different methods used in the work. The methodology of this work is described in Sect. 3. Section 4 gives a detailed analysis of the obtained results. The paper is concluded in the Sect. 5.

2 Related Work Various methods have been tried in the literature to implement DAF and FAF. Bahadorinejad and Almasganj [3] discussed the implementation of anti-stuttering device using AVR ATmega128 and hearing aid technology. In this work FAF is used to feed once own pitch shifted voice to the user as an auditory feedback. Ramteke et al. [4] focuses on detection of the repetition in the stuttered speech using shimmer, MFCC, and formant features. These features are extracted from the isolated words of 27 s speech data. The dynamic time warping technique with a suitable threshold value is used for the detection of the repetitions in the stuttered speech. Macleod et al. [5] discussed about the implementation of the combination of DAF and FAF in stuttered speech therapy and found that the combination obtained better results than DAF or FAF when used separately. DAF with 50 ms delay and FAF with a downward shift of one half octave was used in this study. Kadam et al. [6] proposed a human voice recognition system that uses Mel frequency cepstral coefficients (MFCC) with artificial neural networks (ANN) to analyze and detect human voice in different applications such as medical, military, and telecommunication with Raspberry Pi as the processor. MFCC features are extracted from the recorded voice samples. Built-in ANN is trained with the training data-table created using MFCC feature. Finally, the

Real Time Feedback System for Speech Dysfluency in Children

77

test data are given to ANN which compares the test data with the training data and recognizes the speaker. Chee et al. [7] discussed about the classification of stuttered speech from the normal speech by acoustically modeling frequently occurring repetitions and prolongations in stuttered speech using MFCC. The performance of MFCC in modeling repetitions and prolongations in stuttered speech was examined using linear discriminant analysis and k-NN classifiers. Czyzewski et al. [8] discusses about the detection of the stop gaps, repetition, and prolongation of a syllable for counting the stuttering events. ANN is used as a classifier for the automatic recognition of stuttering disorder in the extracted feature set. Sathya and Chandra [9] discussed about using a feature extraction algorithm known as MFCC for the detection of repetitions and prolongations in a stuttered speech. Here Linear Discriminant Analysis (LDA) classifier and k-NN classifier are used to measure the effectiveness in recognition of a stuttered speech. An average accuracy of 90% is obtained using these classifiers. Ai and Yunus [10] introduced four different types of anti-stuttering devices which are based on DAF, FAF, masked auditory feedback (MAF), and combined/multiple feedback. These algorithms were developed to assist a speech language pathologist for Malay speech therapy in Malaysia. Saranya and Ravi [11] proposed a low cost portable device based on digital signal processing to alleviate the problem of stuttering. Authors implemented DAF with the delays in the range of 50–200 ms. FAF was implemented with varying pitches such as a quarter, half or full octave shift resulting in 55–74% decrease of stuttering in the short reading tasks. Yaruss [12] discusses about the online real time data collection to analyze the frequency of the different types of speech disfluencies produced by the people who stutter and also proposes the frequency count values to be used by the clinicians in stuttered speech therapy. Borsel et al. [13] investigated the effect of DAF on the people with the stuttering disorder. The patients with the stuttering disorder were exposed to DAF for a period of 3 months. Significant improvement was observed in the stuttering after 3 months as the percentage of stuttered words in the people who were exposed to DAF was reduced drastically. Metz et al. [14] discusses about the modified Van Riperian procedure in the analysis of the stuttered speech to investigate its efficiency in the post and pre stuttering therapy. The experiment was carried out by involving nine adult stutterers. It was found that the stuttering frequency is reduced significantly over the course of therapy using the modified Van Riperian procedure. Riley [15] proposes a device for assessment of the stuttering severity by analyzing the stuttered speech samples of 109 children and 28 adults. The assessment of severity was mainly based on the frequency and duration of the prolongation and repetition of the speech segments in a given passage. Leith and Chmiel [16] verified the effect of DAF in improving the speech rate of a stutterer. The authors also focuses on two major programs: rate control program and fluency program to increase the fluency and rate of the speech in a stutterer. Chee et al. [17] investigated the performance of artificial neural networks (ANN), hidden Markov models (HMM), and support vector machine (SVM) classifiers for the detection of the stuttering. The highest classification accuracy of 96% is obtained with HMM, where as SVM and ANN yielded almost similar accuracies with 94.9% and 94.35%, respectively.

78

J. C. Saldanha and R. Pinto

The usefulness of DAF and FAF in treating the stutter and clutter speech disorder can be seen in the literature. Hence, in this study the effect of DAF and FAF are observed for a Kannada passage in the age group of 10–12 years including both male and female gender.

3 Proposed Method to Treat Stuttering This work involves two phases in detection and treatment of the stuttering. The first phase discusses speech therapy where three feedback methods for assisting the stutterer to improve fluency are discussed. The second phase is on different techniques used for speech analysis for assessment of severity of stuttering.

3.1 Speech Therapy Speech therapy can help patients learn to speak clearly. This helps them to feel confident and less frustrated while communicating. The general block diagram of the speech therapy system is shown in Fig. 1. Raspberry Pi B is used as the processor in this system. The three techniques namely DAF, FAF and metronome generation are used in this work for speech therapy. The user is intended to choose one among the three therapy techniques and follow the instructions as provided by the therapist. Delayed Auditory Feedback: The speech input is given to the processor via a microphone with a good quality reception. With the help of a library called the SoX, the signal is shifted in the time domain causing a delay. The delay is kept variable which will be different for different people. Frequency Altered Feedback: The noise free signal is Fourier transformed to obtain a frequency domain signal. The frequency can be up-shifted or down-shifted. The output is then inverse Fourier transformed to get back the time domain signal. The time domain audio signal is then sent back to the user as a feedback.

Fig. 1 Block diagram of speech therapy

Real Time Feedback System for Speech Dysfluency in Children

79

Metronome Generation: A sine wave of a particular frequency is generated for 100 ms and the same sine wave is repeated after a specified interval of time. The user is made to utter a word according to the metronome generated.

3.2 Speech Analysis The analysis of the stutterer’s speech is carried out to classify the patients according to their criticality and provide the treatment accordingly. The analysis is also done to measure the improvement during the course of the therapy. The categorization in presently done manually by the therapists by listening to the audio recorded during the therapy sessions. The judgements may be different by different therapists and lead to the lack of standardization. Moreover, listening to the recordings of the patients and classifying them manually is a hectic task. The introduction of a standard analysis method can help in getting the same results even if the patient is treated by different therapists with the added advantage of reducing the work of the therapists in listening to the audio sample and classifying. The block diagram of the speech analysis is shown in the Fig. 2.

3.2.1

Preprocessing

Framing and Windowing: The time period for a frame is fixed and the number of samples in a frame is obtained according to the time period and the sampling rate. Each frame is multiplied with a window function of the length of frame size, giving a finite length weighted version of the original signal to avoid a discontinuity between any two frames. This work uses hamming window with 50% overlap and the audio signals have sampling rate of 16 kHz. The time period of the frame is 25 ms. So the frame length is equal to 400 samples. Hamming Window weight equation is given by, 

2π n w(n) = 0.54 − 0.46 cos M

 (1)

where M is the length of the window. Pre-Emphasis: In speech processing, the original speech signal has a greater amount of low frequency energy content compared to the high frequency energy content.

Fig. 2 Block diagram of speech analysis

80

J. C. Saldanha and R. Pinto

Pre-emphasis is used to emphasize the high frequency content of a signal thereby improving the performance of the system. From the spectrum of the speech sounds, one can notice that less energy is present in the highest frequencies, with an overall decreasing slope. The goal of the pre-emphasis filter is to counterbalance this, and flatten the spectrum. It is also used to increase the high frequency formants in the speech. Pre-emphasis can be implemented using a simple filter as shown in Eq. 2. y(n) = x(n) − αx(n − 1)

(2)

where α is a non-zero constant close to 1. (In speech processing application α is commonly in the range of 0.95–0.97).

3.2.2

Feature Extraction

Mel Frequency Cepstral Coefficients: Fast Fourier Transform (FFT) is applied to transform every frame with 400 samples from the time domain to the frequency domain. Since there are 400 samples, 512 point discrete Fourier transform (DFT) is applied. FFT is a fast and efficient method to calculate the DFT. The power spectrum of each frame is calculated. Next step in the computation of MFCC is Mel frequency warping. This step involves the creation of a filter bank with triangular band pass filters placed at a specified intervals. It involves converting the Hertz scale to the Mel scale because Mel scale represents the human ear perception [18]. Equally spaced 13 bandpass filters are used in this work. The filter bank energies are obtained by multiplying each filter bank with the power spectrum and then adding up the coefficients. The 13 values obtained represent the energy in each bandpass filter. The logarithm of all the 13 filter bank energies is calculated to obtain the log filter bank energies. This is done so that the multiplication gets converted to the addition in the logarithmic scale and thus, the envelope can be detected easily by applying a filter. The discrete cosine transform is applied instead of the inverse Fourier transform due to its computational efficiency. It reduces the steps by not calculating for the redundant values and uses only real numbers instead of the complex numbers unlike the inverse discrete Fourier transform (IDFT) which uses complex numbers. It is thus, a better replacement for the inverse Fourier transform. Prolongations: It is basically defined as involuntary lengthening of speech sounds. It is also one of the types of speech dysfluencies [19]. Prolongation is found out by finding the power of each frame and comparing the power with the next consecutive frame. If the power difference between consecutive frames is less than a threshold value for a certain number of frames, prolongation can be easily found out. The number of prolongations occurred in the sentence as well as the duration of the prolongations are noted for reference. The algorithm to analyze and determine the prolongation is as follows:

Real Time Feedback System for Speech Dysfluency in Children

81

Algorithm: Step 1: Start Step 2: Take the Pre-Processed output as an input. Step 3: Compute the power of the individual frames and store the output in an array a[]. Step 4: Find the power difference between consecutive frames (a[i + 1]) − a[i]) and store the result in another array b[]. Step 5: If the power difference is less than 1 mW for 25 consecutive frames or more, consider it as a prolongation else do not consider it as prolongation. Step 6: Store and output the result. Step 7: Stop. Silence Blocks: They are said to occur when sound or air stops in the lungs, throat or mouth, lips or tongue. This is said to occur when a person is repeating the same sound and is unable to utter the first or the next sound, the person is said to have been stuck. The word ‘stuck’ here basically defines the meaning of a ‘block’. The silences are found at the low power in the periodogram for more than a specific number of consecutive frames. The number of silence blocks as well as the duration of the silence blocks are noted for further reference. The algorithm to analyze and determine silence blocks is as follows: Algorithm: Step 1: Start Step 2: Take the Pre-Processed signal as an input. Step 3: Compute the power of the individual frames and store the output in an array a[]. Step 4: If the power is less than 10 mW for 60 consecutive frames or more, consider it as a silence block else do not consider it as silence block. Step 6: Store and output the result. Step 7: Stop. Repetitions: A stutterer tends to repeat certain syllables or words or part of phrases. The criticality of his stuttering depends on the number of repetitions he makes [19]. To count the number of repetitions, syllable counting method is used. The patient is made to read a prescribed passage for which the number of syllables are known. By segmenting the syllables in the person’s speech and counting them allows to evaluate the severity of the stuttering. The number of syllables in the predefined passage which is already known is compared with the number of syllables in the speech sample. The results obtained can indicate the number of repetitions done. The syllable segmentation is done by a sinusoidal signal modeling technique. The

82

J. C. Saldanha and R. Pinto

position of the maximum of the amplitude trajectory is the same for each syllable. This makes comparison between the different syllables easy.

3.2.3

Classifier

Classification is performed using k-NN classifier. This classifier works on the principle of finding the nearest sample in the training dataset. In this work Euclidean distance metric is used for the comparison between the test feature vectors with the training feature vectors. The value of k is chosen to be 1, due to which the test sample is simply assigned to the class of that single nearest neighbor. 50% of the total data set is used for the training, and the remaining 50% is used for testing the classifier accuracy.

3.3 Database Stuttered speech and normal speech samples of the children of the age group 10– 12 years were recorded using a Sony ICDX140 voice recorder. A total of 40 samples were collected of children suffering from dysfluency due to stuttering and 40 samples were collected of children without any speech disorder. The speech samples were collected for the given Kannada passage.

.

(Bengaluru is one of the biggest cities in our state. This city is called Mumbai of Karnataka. This is one of the biggest cities in India. People come from different states and different places to see this city. Other than this, people come to visit Bellur, Jog, and Nandi in our state. Ragi is grown in this state).

4 Results and Analysis 4.1 Speech Therapy Implementation In this work DAF was implemented with the delay of 100 ms. FAF was implemented with 300 Hz frequency up-shift and metronome generation was carried out using a sinusoidal signal with the time period of 1 ms using an audio processing library

Real Time Feedback System for Speech Dysfluency in Children

83

called the SoX. The working interface was set up using shell script. Figures 3, 4, and 5 show the implementation of delayed auditory feedback, frequency altered feedback and metronome generation using the shell script in SoX. The user is intended to select one among the different therapy techniques as suggested by the therapists along with the prescribed specifications. The delay, frequency shift, and metronome speed is decided by the therapist. Figure 6 shows the GUI implementation of these techniques using Tkinter. For example, if the user selects ‘Metronome–0.8 s’ using the touch enabled TFT display, the metronome is generated every 800 ms and the person is intended to utter a word once in every 800 ms. Similarly, if the user selects ‘DAF—200 ms’, the delay between the users speech and his perception is set to 200 ms.

Fig. 3 DAF with a delay of 100 ms

Fig. 4 FAF with 300 Hz frequency up-shift

84

J. C. Saldanha and R. Pinto

Fig. 5 Metronome generation with time period of 1 ms

Fig. 6 GUI implemented using Tkinter

4.2 Speech Analysis During the preprocessing stage the speech samples are recorded at a sampling frequency of 16 kHz and quantized at 16 bit pulse code modulation (PCM). The samples are then passed through the feature extraction stage and later used as the inputs for the classifier to train the model and test the samples. Figures 7a and b show the stuttered speech and normal speech for the first sentence in the passage given in the Sect. 3.3. From the figures it is observed that the stuttered speech sample has many silence intervals when compared with the normal speech sample. The speech samples are converted into small segments of speech for further processing using hamming window of 25 ms duration or 400 samples. Figure 8 shows the windowed speech segment with 400 samples.

Real Time Feedback System for Speech Dysfluency in Children

85

Fig. 7 a Speech signal of a normal person. b Speech signal of a stutterer

Fig. 8 Windowed speech frame having 400 samples

The plot of the pre-emphasis of the windowed frame is as shown in Fig. 9. The plot clearly shows that the amplitude of the lower frequency signals being attenuated and that of the higher frequencies being amplified. In order to compute MFCC, the power spectrum of the windowed speech segments is obtained. Figure 10 shows the power spectrum of a frame of a stuttered speaker. Figure 11 shows the Mel filter bank with 13 triangular filters used in the computation of MFCC. The filter bank has number of filters placed at the lower frequencies and very few filters at the higher frequencies to mimic the human auditory response

86

Fig. 9 Emphasis plot

Fig. 10 Power spectrum of a frame of a stuttered speaker

J. C. Saldanha and R. Pinto

Real Time Feedback System for Speech Dysfluency in Children

87

which is more sensitive to the lower frequencies below 1 kHz than for the higher frequencies. The log energies of these filters constitute MFCC’s. Figure 12 shows plot of Mel frequency cepstral coefficients of a speech frame.

Fig. 11 Mel filter bank with 13 triangular filters

Fig. 12 MFCC of a frame

88

J. C. Saldanha and R. Pinto

Fig. 13 Euclidean distance

The average of the MFCC is taken by calculating the average of each individual coefficients with the corresponding coefficients in the other frames throughout the audio which is used as the feature vector for the classification. These feature vectors obtained are used during the classification stage to compute the Euclidean distance. The Euclidean distance of the average MFCC of the test sample from the training samples are calculated and is assigned with the label of the nearest neighbor. The plot of the Euclidean distance of the test sample from the other training samples is as shown in Fig. 13. The results obtained above were not very satisfactory as they did not yield the expected results. The accuracy of the classification using k-NN classifier was 22.23% which is found to be very low to be used in the real time applications. Mel frequency cepstral coefficients are found to be effective in speech/speaker recognition but these features did not obtain significant results in the recognition of stuttering. Hence the effectiveness of the alternate features such as prolongation, repetition and silence intervals were investigated. Prolongation is found out by finding the power of each frame and comparing the power with the next consecutive frame. The detection of prolongation is as shown in Fig. 14. The silences are found at the low power in the periodogram for more than a specific number of consecutive frames. Figure 15 depicts the detection of the silence blocks in a sample. By segmenting the syllables in the person’s speech and counting them allows to evaluate the severity of the stuttering. The number of syllables in the predefined passage which is already known is compared with the number of syllables in the speech sample. The results obtained can indicate the number of repetitions done. Figure 16 shows the spectrogram of the speech sample as well as the segmented syllables represented in a red color.

Real Time Feedback System for Speech Dysfluency in Children

89

Fig. 14 Detection of prolongation

Fig. 15 Silence block detection

It was evident from the analysis of the stuttered speech that counting the number of prolongations, silence intervals, and repetitions in a given passage can be very useful in the recognition of stuttering, and to determine the severity level of stuttering.

90

J. C. Saldanha and R. Pinto

Fig. 16 Detection of repetitions

5 Conclusion The three techniques namely DAF, FAF, and metronome generation were implemented using Raspberry Pi as the processor and audio processing library called the SoX. Graphical User Interface (GUI) for the above application was implemented using Tkinter–a Python-based GUI Toolkit. To measure the improvement in the speech, the feature vectors of the speech samples were extracted using MFCC. KNN classifier was used to classify stuttered and normal speech using the MFCC feature vector. Using only MFCC as a feature did not prove to yield the expected results. Hence silence blocks, repetition, and prolongation were used as additional features. The speech input was analyzed for the occurrence of prolongation, silence interval and repetition and for the number of times these events occur in speech utterances. These features found to be effective in the detection and assessment of the severity level of stuttering.

References 1. Maguire GA, Yeh CY, Ito BS (2012) Overview of the diagnosis and treatment of stuttering. J Exp Clin Med 4(2):92–97 2. Herder C, Howard C, Nye C, Vanryckeghem M (2006) Effectiveness of behavioral stuttering treatment: a systematic review and meta-analysis. Contemp Issues Commun Sci Disord 33:61– 73 3. Bahadorinejad A, Almasganj F (2012) Delayed auditory feedback for speech disorders. In: IEEE international conference on biomedical engineering (ICoBE), 2012, pp 585–588 4. Ramteke PB, Koolagudi SG, Afroz F (2016) Repetition detection in stuttered speech. In: Proceedings of 3rd international conference on advanced computing, networking and

Real Time Feedback System for Speech Dysfluency in Children

91

informatics, New Delhi. Springer, pp 611–617 5. Macleod J, Kalinowski J, Stuart A, Armson J (1995) Effect of single and combined altered auditory feedback on stuttering frequency at two speech rates. J Commun Disord 28(3):217–228 6. Kadam M, Rao G, Sakshi AP (2018) MFCC feature extraction for speech recognition with hybrid application. Int J Adv Res Ideas Innov Technol 4(3):1380–1384 7. Chee LS, Ai OC, Hariharan M, Yaacob S (2009) MFCC based recognition of repetitions and prolongations in stuttered speech using k-NN and LDA. In: 2009 IEEE student conference on research and development (SCOReD), Nov 2009, pp 146–149 8. Czyzewski A, Kaczmarek A, Kostek B (2003) Intelligent processing of stuttered speech. J Intell Inf Syst 21(2):143–171 9. Sathya MAJ, Chandra E (2015) Types and tools available for fluency disorder—speech therapy. Int J Adv Res Comput Sci Softw Eng 5(2) 10. Ai OC, Yunus J (2006) Overview of a computer-based stuttering therapy. In: Regional postgraduate conference on engineering and science (RPCES 2006), Johore, 26–27 July 2006, pp 207–211 11. Saranya R, Ravi N (2016) A low cost portable device to alleviate the problem of stuttering. Int J Emerg Technol Comput Sci Electron (IJETCSE) 21(3) 12. Yaruss JS (1998) Real-time analysis of speech fluency. Am J Speech-Lang Pathol 7(2):25–37 13. Van Borsel J, Reunes G, Van den Bergh N (2003) Delayed auditory feedback in the treatment of stuttering: clients as consumers. Int J Lang Commun Disord 28(2):119–129 14. Metz DE, Onufrak JA, Ogburn RS (1979) An acoustical analysis of stutterer’s speech prior to and at the termination of speech therapy. J Fluency Disord 4(4):249–254 15. Riley GD (1972) A stuttering severity instrument for children and adults. J Speech Hear Disord 37(3):314–322 16. Leith WR, Chmiel CC (1980) Delayed auditory feedback and stuttering: theoretical and clinical implications. Speech Lang 3:243–281 17. Chee LS, Ai OC, Yaacob S (2009) Overview of automatic stuttering recognition system. In: Proceedings of the international conference on man-machine systems (ICoMMS), Batu Ferringhi, Penang, Malaysia, 11–13 Oct 2009, pp 1–6 18. Jhawar G, Nagraj P, Mahalakshmi P (2016) Speech disorder recognition using MFCC. In: IEEE international conference on communication and signal processing (ICCSP), Melmaruvathur, India, 24 Nov 2016, pp 0246–0250 19. Chee LS, Ai OC, Hariharan M, Yaacob S (2009) Automatic detection of prolongations and repetitions using LPCC. In: IEEE international conference for technical postgraduates (TECHPOS), Dec 14 2009, pp 1–4

Nonlinear Model-Predictive Control Using First-Principles Models R. Russell Rhinehart

Abstract The common attribute in horizon-predictive control is the use of a dynamic model of the process and optimization to plan a sequence of future control actions to best make the model match a desired future path, while avoiding constraints. This paper describes a structure to use first-principles process models in horizonpredictive control, which contrasts the conventional use of linear empirical models derived from step-testing the process. Keywords Process control · Nonlinear · Horizon predictive

1 Introduction Model-predictive control (MPC) has many alternate names and acronyms. Some are DMC (dynamic matrix control), HPC (horizon-predictive control), IdCom (Identify and Command), and APC (advanced process control). The common attribute is that they use a dynamic model of the process and use optimization to plan a sequence of future control actions to best make the model match a desired future path, while avoiding constraints. The first control action (manipulated variable value, MV), in the plan, is implemented. The entire plan is not executed, just the first step. Then after one control interval, the future MV sequence plan is recalculated to accommodate changes in set points, disturbances, and process-model mismatch. And, again, the entire plan is not executed, just the first step. The models in MPC can reveal both disturbance and interaction, as well as deadtime and other ill-behaved dynamics. The optimization in MPC compensates for the model features and also has override action and can choose best MVs to manage the process when there are extra options. Although such features could be implemented in classic ARC (advanced regulatory control) with feedforward, decouplers, ratio, override, etc., once an implementation exceeds about 2 inputs and 2 outputs, the R. R. Rhinehart (B) Oklahoma State University, Stillwater, OK 74074, USA e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_6

93

94

R. R. Rhinehart

complexity and tuning of the number of model coefficients become very complicated to manage. MPC provides a unified structure, with only one tuning coefficient for each controlled variable (CV). A rule of thumb is that MPC (or any control strategy with equivalent features) will halve the CV deviations associated with conventional regulatory control. This reduction in variance permits assigning set points that are closer to specification limits, which translates to any of many economic benefits—higher throughput, less waste, lower energy and raw material use, etc. Although relatively expensive to license an MPC product, and although the process step testing to develop empirical models takes time and moderately upsets the process, MPC has become accepted as a best control approach in continuous chemical processing. Normally, the MPC models of the process response are Finite Impulse Response (FIR) models, a vector of normalized responses to the steps in both control action and disturbances representing a linear response. Alternately, models could be secondorder plus deadtime (SOPDT), also linear. And commonly, the optimization used to plan a future sequence of manipulated variable (MV) moves is Linear Programming (LP). Although many commercial products offer empirical models with nonlinear gain (such as neural networks based on steady-state responses, or even heuristically set gains from interviewing operators and engineers), most models remain linear in the dynamics (lags and delays). Because LP seeks to place the control action on a constraint, an extreme boundary, and requires linear objective function and constraint models, many products use alternate optimizers, such as Successive Quadratic. One issue is that chemical processes are nonlinear and non-stationary. Both gains (sensitivities and interactions) and dynamics (time constants and delays) change with operating conditions, product specification, internal tank levels, piping reconfigurations, equipment switching, etc. When the process operating conditions or process attributes change, linear models become significantly mismatched to the process and require model recalibration (recurring process step testing). It would be desirable to use models that remain consistent with the process. Another issue is that the mathematics and style of the classic controller models are substantially different from the phenomenological models that process engineers commonly used for process design, analysis, troubleshooting, constraint forecasting, optimization, and training. The mathematics of the linear models obscures process understanding. It would be desirable if the models in control were consistent with those used in all the process management actions. This article is about the use of first-principles models in MPC. Since the models are nonlinear, it is nonlinear model-predictive control (NMPC). Phenomenological models are mechanistic, cause-and-effect models. For process control, start by applying fundamental mass, momentum, and energy balances on a control volume, to get a model of an element of the process. Then integrate (sequentially combine) the individual models to describe the unit. If the models are elementary in their description of the process, if they use ideal relationships, they are termed first-principles models. If the models seek to fully describe all phenomena, they are rigorous models. First-principles models are characteristic of undergraduate courses,

Nonlinear Model-Predictive Control Using First-Principles Models

95

product bulletins, training courses, etc. First-principles models are relatively easy to develop, understand, and adapt and are fully adequate for control. Since about 1980, there have been many approaches to nonlinear process control. It seems that the industrial acceptance of digital control devices and distributed control systems (DCS) of the 70s, revealed a vision to many investigators that modelbased control could be practicable. There have been many concepts as to how best to implement it. For example in increasing complexity: Generic process control (GMC) provides a structure to use steady-state models in an output characterization of a PI controller. If the process is SISO and the problem is nonlinear gains, this is an excellent solution. Process model-based control (PMBC) is a single step-ahead controller (it does not forecast a future plan or handle ill-behaved dynamics), but can handle MIMO processes with constraints, disturbances, and nonlinear and non-stationary features. It was developed to adapt the model to changing process attributes such as catalyst reactivity, tray efficiency, heat exchanger fouling, and drag coefficients. If delay, high-order response, or inverse action is strong, PMBC would not be a good solution. Predictive functional control (PFC) was developed to handle both nonlinearity and ill-behaved dynamics (delays, high order, inverse action). It uses a process model to calculate a step-and-hold MV action, if implemented now, that would make the model match its set point at some future time (the coincidence point) well past the influence of the ill-behaved dynamics. These are all practical, industrially proven techniques. Complying with the K.I.S.S. principle (Keep it Simple and Safe) that guides applications, those strategies should be considered prior to the topic of this article, first-principles models for multivariable, constraint handling, and model-predictive control.

2 What Are First-Principles Models? First-Principles models are the process engineer’s models which are used for process design, analysis, optimization, data reconciliation, and troubleshooting. They are based on elementary material and energy balances and ideal constitutive relations and are characteristic of college classroom teaching models for heat exchange, fluid flow, reaction engineering, separations, etc. Representing process mechanisms, they are typically nonlinear. Alternatively, one could use rigorous models attempting to perfectly explain every nuance. But balancing sufficiency with perfection, first-principles models are fully adequate.

2.1 How Can Elementary Models Be Adequate for Control? Contrasting modeling perfection, consider that humans can catch a fly ball or steer a car through a curve, without rigorous models, with just intuitive continual correction.

96

R. R. Rhinehart

And consider that PI controllers, based on linear, stationary, simplistic (FOPDT), inexact models, are also adequate for control. Here is a qualitative explanation: If there are 30 or so control intervals during a process transient, then even if the model is only 70% correct, leaving a 30% error in its first action, at the next control action, the remaining error is 70% corrected, leaving a 0.3 * 0.3 = 0.09 error. By the time a few control actions have been taken, the error is smaller than the discrimination interval for digital values.

2.2 Automobile Speed Illustration As an example of first-principles models, I will choose an automobile speed illustration. Although not a process model (such as reaction of distillation or heat exchange), the concepts are more universally familiar. Start with Newton’s Law of motion and then expand the acceleration term to the rate of change of velocity. F=

ma m dv = gc gc dt

(1)

An elementary representation of the forces on the car is from: (1) The engine, represented by the accelerator pedal position, u, which will be modeled as a linear response, (2) aerodynamic drag, ideally represented as proportional to velocity squared, and (3) gravitational impact of going up or down hill with an angle, θ , from the horizontal. F = ku − bv 2 − sin(θ )mg/gc

(2)

Combine Eqs. (1) and (2), to obtain the elementary model. m dv = ku − bv 2 − sin(θ )mg/gc gc dt

(3)

This can be converted into the classic control form of a 1st-order process by dividing by bv then rearranging terms. k sin(θ )mg m dv +v = u− bvgc dt bv bvgc

(4)

And, now, it can be represented in generic terms for a first-order process with a locally linearized disturbance model. τ

dv + v = Kpu + Kdθ dt

(5)

Nonlinear Model-Predictive Control Using First-Principles Models

97

Compare Eqs. (4) and (5) and notice that the time constant, τ , is nonlinear, depending on the velocity. The time constant is also non-stationary because the drag coefficient, b, will depend on air density; and because the mass, m, will depend on fuel in the tank and other loading changes. The process gain is also nonlinear and non-stationary in spite of the linear force model in Eq. (2). And the sine function nonlinearity of the road angle impact is now also confounded by speed, mass, and drag coefficient. For model-based control, we will use an alternate rearrangement of Eq. (3) which isolates the rate of change.  gc  dv = ku − av 2 − sin(θ )mg/gc dt m

(6a)

In generic terms, with y as the state variable, u as the MV value, d the vector of disturbance variables, and p as the vector of model parameter values, Eq. (6a) can be represented as   dy = f y, u, d, p dt

(6b)

This represents a first-order process, a process stage with one inventory location of mass, momentum, or energy. It could represent how fluid flow rate responds to the controller signal to the valve actuator, how temperature in a well-mixed tank depends on a heat source, how composition in a well-mixed tank depends on influent flow rate or composition, how pressure in a vessel depends on in- or out-flow or internal generation of gas, etc. Some one-inventory-stage processes have more than one state variable. For instance, in a continuous flow reactor with well-mixed contents, temperature, composition, and volume would be several interacting states, which would be dependent on several influences input flow rates, heating/cooling rates, and discharge rates. Here, each CV would have a first-order model such as Eq. (6a), but each state variable would have an influence on the rate of change of the others. In generic terms for a MIMO set of coupled first-order processes, with y as the vector of state variables, u as the vector of MV values, d the vector of disturbance variables, and p as the vector of model parameter values: dy dt

  = f y, u, d, p

(7)

However, the process might have several stages that sequentially influence each other. For example, in a tray-to-tray distillation column, the liquid contents on each tray have several interacting states (temperature and compositions), which are influenced by the trays above and below. In a 10-tray binary distillation, there would be 20 states on the trays (temperature and x1 ) (the other liquid and vapor compositions would be related to the states). But there would also be column pressure and reboiler

98

R. R. Rhinehart

and condensate receiver as states. Although there may be 23 states, only 5 would be controlled (top and bottom composition, column pressure, and reboiler and receiver levels). Equation (7) could represent 23 first-order, nonlinear, interactive models, but there may only be 5 MVs (reboiler heat, bottom’s flow rate, cooling flow rate to the condenser, reflux flow rate, and distillate flow rate). As a note, regardless of the form of the model presentation, do not solve the model with calculus for several reasons. The nonlinearity and non-stationarities may preclude analytical solution. Any solution will be predicated on an idealized timedependent behavior of the MV and disturbances. And any solution will be predicated on fixed values (or at least elementary models) of how the model coefficients might change in time. Alternately, to determine what the model predicts, use a simple numerical solution. Here is Euler’s explicit method, by converting the differential to a forward finite difference.   (8) yt = yt−Δt + Δt · f y, u, d, p t−Δt

This is a recursive relation, the new values of the states, yt , are calculated from the   prior values of the states, and of all other variables, yt−Δt + Δt · f y, u, d, p . t−Δt

Initialize this calculation with the prior modeled y-value. In this car speed example, the model was a single variable and could be coded on a single line. It appears that Eq. (8) could be coded on N lines, one for each of the N state variables. Alternately, a model may have IF–THEN conditionals such as representing disparate mechanisms such as due to the transition from laminar to turbulent flow, or from turbulent to choked flow. And there may be many other constitutive relations, such as thermodynamic models that relate state variables, even requiring root-finding methods to solve them. Or steady-state or equilibrium relations for actions that have very fast dynamics. Also, for numerical stability the model Δt might be 1/10th of the control interval, requiring that Eq. (8) be recursively calculated 10 times within each controller interval. Alternately, a preferred numerical solution might be a 4th-order Runge–Kutta approach. For many reasons then, the future modeled values might be coded as a separate procedure (a function or a subroutine). There is no need to have a single line of code (or N single lines) as suggested by Eq. (8). Note: The model is not the truth about the process. This model presumed that the engine force is linear with pedal position and independent of engine speed, gear, and intake air composition. This model presumed that the engine force instantaneously responds to pedal position and that there are no lags in getting fuel–air mixture into the engine or back pressure on exhaust as engine speed increases. This model ignores rolling friction of the tires on the pavement. This model uses the ideal Bernoulli exponent of 2 for the aerodynamic drag force, and the model also ignores that the drag force would be a function of the differential velocity between wind and car speeds, not just car speed. First-principles models are fully adequate for control. Use ideal relations. Do not seek fully rigorous models.

Nonlinear Model-Predictive Control Using First-Principles Models

99

3 MPC Controller Structure There are three functions within model-based control—predict, correct, and act. The structure can be represented by the operations in Fig. 1.

3.1 Predict The first MBC function is to use the model to predict or mimic what the process output is expected to do. This is the lower right box in Fig. 1. This updates the past model value (from the prediction) with past disturbance and control action, to predict the updated (current or new) modeled output. With conventional practice of 30 control samplings in an open loop transient (or 10 samplings within a time constant), the control interval, Δt, is usually small enough to permit Euler’s explicit finite difference method to solve the model numerically. Using Eq. (8) to represent a SISO dynamic model, this model prediction calculation becomes: ym,i = ym,i−1 + Δt · f (ym , u, d, p)i−1

(9)

where “i” is the time increment counter, representing the present value, and “i − 1” the immediate past values for each item in the function list. The subscript “m” refers to the modeled value, not the measured process value. However, “u” refers to the actually implemented MV value, after any overrides change what the controller initially desired. The symbol “d” refers to measured disturbances or feedforward variables, and “ p” represents parameter values in the model. Biased SP SP

Act

y-process

Override

Process

MV bias Predict

Correct pmm

Fig. 1 MBC representation

ymodel

100

R. R. Rhinehart

Note: The right-hand side of Eq. (8) uses the prior modeled output value, ym,i−1 , not the prior process value, to update the modeled value, ym,i . Because of processmodel mismatch, noise, and such, the process measurement is not the right value to initiate the model update. However, since computers use assignment statements, the i-subscript is not necessary, and there is no need to store all of the past model values. The new modeled y-value is based on the most recent prior ym and possibly delayed input conditions. Here, the symbol “:=” will indicate a computer assignment statement. A delay array with put and retrieve pointers would be needed to obtain the u θ and dθ values. ym := ym + Δt · f (ym , u θ , dθ , p)

(10)

If the process model is high order, then internally to a procedure representing “Predict” there would be a set of equations similar to Eq. (10). Also, if the model has conditionals, or root-finding algorithms, then again, the model function would be a procedure, not just one equation.

3.2 Correct The second function within model-based control is a correction for process-model mismatch. Process-model mismatch, pmm, is the difference between modeled prediction and process measurement, as represented by the circle difference operation on the lower right side of Fig. 1. The modeled value will not exactly match the process value for a variety of reasons including model error (as discussed in Sect. 2.2), noise, or unmeasured disturbances. A simple correction approach is to bias the set point for the model with the process-model mismatch as follows: pmm = yp − ym

(11)

ySPbias = ySP − pmm

(12)

The logic is illustrated in Fig. 2. “If you aim at the target, but hit 3 cm low. Then, next time aim 3 cm above the target.” The box labeled “correct” in the lower left of Fig. 1 could provide other functions. One related to model coefficient adjustment has some managerial information benefits, but in this simple correction approach, it simply is a pass-through of the pmm value to bias the set point for the model. Note: The biased set point is the set point for the model, not the set point for the process. Note: pmm is not the actuating error. It is not the process deviation from the target. If at first you aim at the bull’s eye, but hit 3 cm low. Next aim 3 cm high. If that shot (modeled to hit 3 cm above) has a process response that is 0.5 cm above the bull’s

Nonlinear Model-Predictive Control Using First-Principles Models

101

3) So aim here instead

Aim higher by pmm. Bias the target by -pmm

1) Modeled value, you aimed here 2) Actual process value that results

Fig. 2 The concept of biasing setpoint for model by pmm

eye, the pmm is the difference between where you aimed (+3 cm) and what happened (+0.5 cm). So, the pmm is 2.5 cm, next aim 2.5 cm above (not 0.5 cm below).

3.3 Act—An Elementary Choice—Simple PMBC The third model-based control function is to calculate the control action, to determine the controller MV value plan. This is represented as the box labeled “Act” in Fig. 1. This starts with a user-defined, desired performance objective for the controller to achieve. A simple desire, in the strategy termed process-model based control (PMBC), is to calculate a u-value that would push the model toward the biased set point at a rate proportional to the deviation from biased set point. (This is an effective strategy for low-order models with inconsequential delay. Section 3.5 discusses how to include constraints and ill-behaved dynamics in the MV plan.)

3.4 A Caution About the Triple Model Use The control model is used in two places. In both the Predict and Act functions. The Predict function updates the past modeled value with the actual implemented MV to predict what the CV is doing now. This is a one-time-step update using the first step of past values. I use the name Past-to-Now which has the acronym P2N. Initialize the one-time-step update with the prior P2N modeled value. The Action (Act) function uses a model to forecast what might happen in the future, associated with a trial solution for the sequence of MV values. Guess at an MV plan and see what the modeled CV does. Re-guess at an MV plan until the future modeled CV matches the desired rate path. I call this the Now-to-Future model, which has the acronym N2F. Initialize the N2F prediction with the current P2N modeled value. After each MV guess, the N2F model needs to be re-initialized with the current P2N value.

102

R. R. Rhinehart

Note: Both the P2N and N2F models use the same Eq. (10). But they need to be initialized with the correct ym values. Use ym,i−1 for the P2N model. Use ym,i for the N2F model. Note: The P2N and N2F models must be consistent with each other. Once when I was clever, seeking a way to avoid the iterative search for the MV, I modeled the inverse with empirical approaches. But at steady state, the inverse surrogate for the N2F prediction did not exactly match the P2N model, and accordingly, control had −1 a steady-state offset. The inverse could be represented as u = MN2F (y ), and the  SPbias −1 model could be represented as ym = MP2N (u). Then, ym = MP2N MN2F (ySPbias )  −1  and the models must ensure that at steady state, ySPbias = MP2N MN2F (ySPbias ) . In control simulations, the process function will be simulated by a model. If the process model in the simulator is the same as the controller model, then there will be no steady-state offset, zero pmm, and perfect control. This is cheating. I do not know of any process that could possibly be exactly modeled. Be sure that the controller models are not identical to the process model in the process function. This includes both functionality and coefficient values. For example: If you use the Ideal Gas law in the controller model, use a different Equation of State in the process. If you use the ideal Bernoulli relation for fluid behavior in the controller model, use 1.8 rather than 2 as the exponent in the process. If you use a catalyst reactivity of 0.8 in the controller model, use a different and ever-changing value in the process simulator. There should be many differences between the controller models and the process simulator model.

3.5 The Act Function in Horizon-Predictive Constraint Handling Control In horizon-predictive control (HPC), the Act function forecasts several future MV moves (3 will be used in this article) to: 1. Shape the future modeled CV response to best fit a reference trajectory (here, first-order, from current model P2N value to the biased setpoint for the model). 2. Steer variables clear of constraints or limits, or balance violations when a constraint is unavoidable. 3. Economically optimize which MVs to use when there are extra Degrees of Freedom (extra MVs). Calculating the future MV sequence is an optimization application. HPC Objective 1—Following the Reference Trajectory—Figure 3 illustrates Item 1 in the list above. The vertical axis represents a PV or MV value, and the horizontal axis represents time. The time of 40, marked by the vertical line, is the current time. To the left are past events, and to the right extending beyond a time of 50 is a possible future plan of three MV steps. The plan has not happened and might not happen. Like a trip plan, which is revised when you hit a detour or decide to take a side visit,

Nonlinear Model-Predictive Control Using First-Principles Models

103

Fig. 3 Forecast MV moves to make the model best match a desired path

the original might never exactly happen. In this example, the process started at an initial steady state. At the left of the time maker, the black line (the lower of the three lines) has been the CV, which is steady and at the set point. The red line (the upper of the three) has been the MV. The turquoise line (the middle line) has been the model predictions. The difference between the modeled and actual CV values is the pmm. To the right of the time marker are the possible future MV outcomes. The lower horizontal line is the biased setpoint, the current target for the model. The black line, which makes a first-order transition from the current modeled CV to the biased set point, is the reference trajectory, the desired path for the controller MV plan to move the modeled CV value. This was chosen to be a first-order path. The plan is the three steps in the red line, the possible MV sequence. The green line, the upper of the lines early in the future, just after the now marker, is the modeled CV value response to the three MV steps. It starts at the present modeled value and shows a bit of a delay, then a higher order lag, and a subsequent path which substantially follows the reference trajectory. To get this modeled CV path, the controller MV plan is to first make a large change to get the modeled CV to move quickly; but, if held there, the modeled CV would fall substantially lower than desired. Continuing the plan, raise the MV to prevent overshoot from the reference trajectory, but this would cause the modeled CV to remain above the reference trajectory longer into the future, so the third step in the MV plan is a bit lower. Note: The future forecast time is termed the control horizon. Hence, the name is horizon-predictive control. The horizon does not extend to the settling time, to steady state for either the modeled or process CV or for the reference trajectory, but typically to 70–90% of the settling time. Accordingly, the reference trajectory does not get to the biased set point, as Fig. 3 illustrates. This does not matter, because only the first MV step will be implemented. Then at the next sampling, the future plan is recalculated with information about new disturbance values, new pmm evidence, and new set points. Note: The control interval is not the length of the first MV step-and-hold. In this simulation, there are about 30 control intervals in the future plan. The first MV step-and-hold in the current plan lasts for about 15% of the horizon (about 5 control

104

R. R. Rhinehart

intervals), the second step-and-hold for about 25% (about 7 control intervals), and the last for about 60% (about 24 intervals). Note: This three-step forecast MV pattern is not essential. More steps would give more precise matching of model to reference trajectory, but in my experience three is fully adequate and minimizes the optimization burden. One could use alternate spacing patterns for the MV transitions. Since the first step is what will be implemented and it shapes the early CV response and since subsequent MV steps trim the plan and are farther out into the future, I like a plan that has more MV moves earlier. This plan roughly spaces MV moves according to the golden ratio, but any of many options works just as well. You might be more comfortable with 1/4, 1/4, and 1/2. The optimization objective is not to force the model to match the reference trajectory exactly, but to minimize the forecast CV deviations from r , the reference trajectory. If the model dynamics are not simply first order, then it would be impossible to make model match the reference exactly. The objective function is a least squares deviation from reference trajectory over the control horizon. Min {u 1 , u 2 , u 3 }

J=

H  2  r1i − Y˜1i

(13)

i=1

Here, i is a counter for future intervals, and H is the number of intervals in the control horizon. If the reference trajectory is first order, the time constant, τ , is the controller tuning factor: τ

dr + r = SP, , tt=now ≤ t ≤ H, rt=now = Y˜P2N,now dt

(14)

If Δt, the control interval is small relative to τ , the time constant for the reference trajectory, as it should be, then using Euler’s explicit method to determine r(t): rt+Δt =

  Δt Δt , SP + 1 − rt τ τ

(15)

For a multivariable process, each CV will have its own reference trajectory, and each MV will have several future values. Note: Some MPC algorithms have as the objective to make the model match the biased set point, which would create a very aggressive control action. For these, the control action is tempered by a move suppression factor, a penalty for large MV moves. This is termed MV damping. What I have described with a reference trajectory target is termed CV damping. In a MIMO process, the move suppression factor on any one MV affects all CV responses. Since tuning with MV damping is interactive in its effect on CV responses and since it does not have an explicit relation to any one of the CV dynamics, I prefer CV damping. CV damping choice of τ for each CV has a direct relation to that CV response speed.

Nonlinear Model-Predictive Control Using First-Principles Models

105

HPC Objective 2—Avoiding Constraints—Another desirable, Item 2 in the list above, is to plan the future MV sequence to avoid constraints. The constraint might be on a CV, or any other process variable (PV), often termed an Auxiliary Variable (AuxV). The constraint might represent a safety or specification limit, or a desired resting value, or a rate of change. There are diverse ways to penalize either a constraint violation or proximity to the constraint. The “Soft” penalty method for a deviation penalizes a violation, proportional to the square of the violation. Penalty1 =

0, if AuxV1 is within the limit (PV1 − L 1 )2 , if AuxV1 exceeds the limit

(16)

But constraints could happen at any future time, so sum all of the penalties. H  Min Penalty1,i {u 1 , u 2 , u 3 }

(17)

i=1

If several variables are subject to constraints, calculate the penalties for each. Constraints on the MV, such as rate-of-change limits or values beyond the 0–100% range of feasible values, could be handled as either soft constraints in Eq. (17) or as hard constraints, limiting MV choices in the optimization. I prefer to treat MV constraints as hard, limiting feasible values for the optimizer. CV or AuxV constraints should be soft. Near a constraint, a disturbance, noise, or such could move the AuxV beyond the constraint. Then, being beyond now, it has happened, and no future control action can undo that. In such a case, all attempts to find the MV sequence will fail. To prevent such infeasible corrections, treat CV and AuxV constraints as soft. HPC Objective 3—Minimizing Utility Cost—The third objective in the list, Item 3, is to minimize costs of using MVs, when there are extra MVs available. If there are extra MVs, then an infinite number of MV actions can provide equivalent control. But usually, some MVs are more expensive than others. For instance, in heating, you may have available electrical resistance heating or low-pressure steam. You would prefer to use the less expensive steam until supplemental heat is necessary. To minimize the cost of MV use over the future horizon calculates the cost for each MV in the plan. A simple model for cost rate for an MV is C˙ = c · MV. Summed over each future time interval, for one MV the objective is: t=H t=H   Min Δt · C˙ t = Δt c · MVt J= {u 1 , u 2 , u 3 } t=now t=now

(18)

For a process with multiple MVs, the projected cost over the control horizon is:

106

R. R. Rhinehart L t=H   Min cl · MVl,t J = Δt {u 1 , u 2 , u 3 } t=now

(19)

l=1

Note: if there are extra MVs then, instead of this economic cost method, you could define a resting value of zero for the high cost MVs and use a deviation from the resting value as a penalty as described next. Combining all Optimization Considerations—Each of the three considerations represents undesirable aspects. The controller should determine the control action that minimizes all. Now that the three considerations are quantified, a standard way of combining all is to sum all the measures of badness, and seek the MV sequences that minimize the sum of all undesirables. 

Min J= u1, u2, u3, . . .

2

H  ˜ i=1 r 1i − Y1i

+ ··· + + ··· +

EC21

H i=1

Δt

+

2

H  ˜ i=1 r 2i − Y2i

Penalty1,i

EC22

H +

EC23

L H l=1 i=1 cl ECcost

i=1

· MVl,i

Penalty2,i EC24 (20)

Note: The decision variables, the u i factors, are underscored, representing vectors. In a multivariable process, there are several MVs, and each would have its several future possible values. (Three MV values for one MV have been illustrated here.) Because each term in Eq. (20) will have unique units, and because some aspects will be more meaningful than others, weighting factors are required to unify units, and to balance relative importance. This representation uses Equal Concern (EC) factors. They scale each quantity by a factor that has the same units as the numerator and represent equal concerns for the violation of the desired value. Note: If there are no constraint violations, then the penalty terms in Eq. (20) are all zero. If the model CVs can exactly follow the reference trajectories (or if the model is at steady state at the biased setpoint), then the first set of terms is also zero. The MV cost term, however, will not be zero. If the optimizer changes MV values so that the MV cost term is reduced, the new MV values will cause the model CVs to deviate from the desired path or from the biased setpoint value. This would make the first set of terms non-zero. It may be possible to reduce cost by a greater amount than the penalty for not following the reference trajectory. This would mean that the process CVs will have a steady-state offset. So, if the cost term is included in the optimization objective function, it is important that the weighting factors make the MV cost term much less important than either the CV deviation or the violation penalty terms. Choose EC factors so that any steady-state offset is imperceptible. See the next subsection. Alternately consider other ways to manage the economic factors. See the second-next subsection.

Nonlinear Model-Predictive Control Using First-Principles Models

107

Note: The EC weighting factors could be used for tuning aggressiveness, but that is not their role. The time constants for the reference trajectories are the tuning factors. One for each CV. HPC 2—Setting the EC Factor Values—Here is a procedure for setting the EC factors. 2

H  r1i − Y˜1i , maybe 1. Choose the most critical term. One term to minimize is i=1 this is of greatest concern. 2. Choose a moderate Y˜1 deviation from perfection. It might be the deviation from the reference trajectory along the future time or the deviation from the biased set point at steady state. This value is an EC factor, and EC1 is equal to the Y˜1 deviation considered to be moderate. 3. Feel the concern this deviation creates. Use both qualitative and quantitative aspects. Perhaps rate the concern level on a 1–10 basis. Concern aspects could include your boss’s opinion, customer reaction consequences, the generation of waste product, etc. 4. Assign EC values to each term in Eq. (20) representing each deviation from perfection that has the same level of concern. As an example: • Perhaps EC1 = 0.5 wt% product deviation from SP has a 1–10 concern rating of a 3. • To have the same concern rating of 3, perhaps the deviation from a target flow rate is 15 lb/min. Then, EC2 = 15 lb/min. • Also, to have the same concern rating of 3, perhaps the deviation from a process temperature limit is 10 F. Then, EC3 = 10 F. The EC values are situation-dependent, and they change with the situation. Note: Traditional optimization uses Lagrange-type multipliers for the terms in a sum. Here, those values are the reciprocal of the denominator values. I think that the weighting method using EC factors is intuitive and easier to assign weighting values, which provide the relative importance. Choosing the economic EC factor may be a bit difficult because it must also balance the consequential steady-state offset. To help see the way to include this aspect, consider a single CV. Ideally, the deviation from set point is zero with some particular MV value. If the optimizer chooses a less expensive MV value perturbed by Δu, there will be a corresponding Δ y˜ = K m Δu, and the CV deviation factor  2 Δ y˜ will be n EC , where n is the number of samplings in the horizon. The cost term y y˜ /K m l ·Δu improvement will be Δt·n·c = Δt·n·cECl ·Δ . We would like the CV deviation to ECcost cost be imperceptible, perhaps small relative to CV noise or relative to the width of the image on the display. We desire Δ y˜ ≤ ε. Equating the two factors and solving for

ECcost , ECcost =

Δt·cl ·EC2y . ε·K m

Alternate Methods to Account for MV Costs—One option is to not include the last term in Eq. (20) objective. It is not in standard linear empirical model-based controllers.

108

R. R. Rhinehart

Another option is to recognize that two MVs may provide the same benefit, but one is more expensive than the other. These may be two possible heating mechanisms steam or the more expensive electrical resistance. In such a case, use the steam until its capacity is inadequate and then add the electrical resistance to provide the extra needed. As another example, there may be two sources of a reagent, dilute and the more expensive concentrate. Use the dilute first, until a flow rate limit is reached or until the dilution causes alternate issues, then trim with the concentrate. There are several ways to handle this logic, and one is to split the 200% range of the optimizer MV. If 0–100% use the less expensive source, then 101–200% also use the more expensive source. Another method is to have a resting value for the more expensive source (it actually may be non-zero to ensure it is immediately available) and add a penalty for deviations from the resting value. A third case might be one in which the extra MV does not have a cost, but affects the values of the other MVs. As an example, pressure on a distillation column can be adjusted by a throttling valve on the vapor leaving the column or on adjusting the flow rate of coolant (process water, or air) in the distillate condenser. Pressure is normally specified in an optimization exercise to stay within limits, maximize column efficiency, and balance performance over a range of expected duties, then not included in the column composition control. However, pressure could be considered an MV for control, and the optimization may determine the pressure that results in minimizing the cost of reboiler power and distillate condenser cooling. Optimization to Determine the Possible Future MV Values—The optimization statement is

min J = {U l,m }

N 

2

H  ˜ − Y r j j i i i=1

j=1

EC2j

+

Δt

L H l=1

i=1 cl

ECcost

+

M  k=1

˙ l,i · MV

H i=1

Penaltyk,i EC2k (21)

S.T: hard constraints on the MVs (Value or Rate-of-Change). The U Subscripts are l = # MVs, m = # future moves for each MV N = # CVs, M = # soft constraints, L = # MVs H = control horizon—the number of Δt calculations in the future time projection. This should go to about 85% of open loop settling time. One could include MV constraints, such as values or rate-of-change (RoC) in the soft penalty terms. These are input variables to the process. However, one cannot include process response variables (CVs and AuxVs) in the list of hard constraints, because a disturbance may cause a response to exceed a limit, and the controller will be unable to immediately remove it from the constraint. Classic MPC products use Linear Programming or Successive Quadratic. However, there are a number of other constraint-handling gradient-based techniques that could be used. Contrasting these, my preference is direct search, multiplayer optimization methods. Multi-player methods are “global,” and direct search

Nonlinear Model-Predictive Control Using First-Principles Models

109

methods do not get confounded by surface aberrations (non-idealities, discontinuities, numerical striations, etc.) and can handle hard constraints, and nonlinear responses. Leapfrogging is my choice, considering effectiveness, speed, simplicity, and robustness [1]. DDC or Not?—Direct digital control (DDC) sends signals from the controller directly to the final control element (valves, heaters, etc.). If you have a single loop controller, SISO, and the model-based control calculation is within the PLC or DCS, then this may be permissible. However, model-based controllers are typically implemented external to the PLC or DCS where greater computational power is available. Although the model-based controller decisions regarding the signal to the final control element could be transferred to the DCS to be relayed to the process, experience has shown that it is best to transfer set points for the process primitive variables (flow rates, temperatures, levels, etc.) and let the PID algorithms in the DCS or PLC determine the signals to the final elements. This is a cascade type of implementation. The primitive or lower-level loops typically respond faster than the primary variables that need to be managed by constraint-handling horizon-predictive control. So, this transfer of set points does not realistically slow control response. The advantages are that the DCS or PLC has many auxiliary functions related to signal processing and validation that should be used, and it is designed to be robust to power and communication failures. If for any reason the computer in which MPC is implemented gets confounded or fails, the DCS or PLC will continue process operation with the last set of set point values. The process will not shut down, if the MPC computer does.

4 MPC Example This example will discuss nonlinear, horizon-predictive, constraint-avoiding modelbased control for the car example in Sect. 2.2. The process, representing the real car, is a simulation. To make the illustration legitimate, the simulation equations for the process are not the same as those for the controller model. The simulation is more complex and has alternate functionalities for effects, and coefficient values that are different from the controller model. One cannot perfectly model the reality of a process. Simulations of control should reflect that reality. For control, this is a SISO process with the controller output to the final control element. However, the process also has an auxiliary variable, intended to represent cabin noise, which has an upper limit constraint. Figure 4 represents a deterministic and disturbance-free trend in simulated process and controller model, subject to ideal step changes in the MV. The red trace that makes step changes is the car accelerator pedal position, the MV. It is attached to right-hand vertical axis. Each step is an increase of 20%. The upper curve is both the car speed and the set point, which are attached to the left-hand vertical axis. In the MAN mode, the set point tracks the CV for bumpless transfer

110

R. R. Rhinehart

Fig. 4 Trends in MAN mode

to AUTO. Note, the first MV step increases the CV by about 42 kph, but the last equivalent MV step is only by about 6 kph. The gain change ratio is about 7:1, being one evidence of nonlinearity. Also note that at the higher speed the settling time is about 15 s, while at the first MV step at the lower speed, even after 25 s, the process is still rising. This nearly 2:1 change in settling time is a second evidence of nonlinearity of the process. The model is represented by the blue dashed line, also attached to the left-hand vertical axis. It too, is nonlinear, with dynamics that change with operating speed. However, the model features do not match the process, exactly. The model is usually lower than the process and the pmm progressively increases with speed. But because the dynamics are different at the first MV step, the model actually rises, temporarily, above the process. The lowest curve, the blue dots, is an Auxiliary Variable; in this case, it is a very simple representation of noise volume associated with the engine and car movement. Figure 5 illustrates several features in AUTO mode. Like Fig. 3, the vertical line toward the right of the graph marks the time now, separating the past and future. Initially, at time = 0 the controller is in MAN mode and placed in AUTO at a time of 2 s. Because of set point tracking and P2N model calculation in MAN mode, there is no bump in the transfer to AUTO. Set point changes occur at times of 5, 40, 60, and 80 s indicated by the black line that makes steps. The solid blue curve represents the process CV (car speed); and in spite of the 7:1 gain nonlinearity, the 2:1 settling time ratio, and the changing pmm, the overshoot magnitude and rise time in the first and third set point changes are very similar. The second and fourth set point changes represent constraint conditions. The AuxV (blue dots) has a soft constraint of 50 (right vertical axis) so even though Fig. 4 shows that the speed set point of 120 kph is feasible at MV = 85%; at that MV value, the AuxV has a value of about 70, exceeding the constraint. In Fig. 5, the EC factors make constraint violations more important than meeting the speed set point, and the optimizer limits the MV to about 60% preventing the speed from reaching

Nonlinear Model-Predictive Control Using First-Principles Models

111

Fig. 5 Events in AUTO mode—1

its set point. When the setpoint is dropped to a speed of 100 kph, there is no delay in the MV action (there is no wind-up in the controller). That was a soft constraint. A second constraint happens at the final set point change, where the controller would like to place the MV below zero. The MV stays at the constrained value limit, until responding (again without wind-up). That is a hard constraint. The third constraint is a rate-of-change constraint on the controller, another hard constraint, which is evidenced by the linear ramp change in the MV immediately after each SP change. Figure 6 illustrates the same events as in Fig. 5, except that the rate-of-change constraint is removed, and the EC factor for the AuxV violation is increased to place greater importance on the CV to SP response. Without the rate-of-change constraints, note the spikes in the MV action just after the set point changes and the MV limit of 100% at a time of 40 s. Also, notice the lesser deviation of the CV to the set point and associated larger deviation of the AuxV from the limit of 50 during the 40–60 s period. There is only one tuning factor for each CV, the time constant for the reference trajectory. In Fig. 7, the time constant is significantly larger, making the controller less aggressive. Compared to Fig. 6, also with no rate-of-change constraint and the duplicate EC factors. Notice the smaller jumps in the MV action and nearly overdamped response of the CV to the SP. Finally, in Fig. 8 both measurement noise and random disturbance patterns are included in the process. Compare this to Fig. 6, which has the same tuning and EC values. The noise is evident in the perturbations on the CV value throughout. The disturbance is evident on the trend in the MV during the 20–40 s period and the upand-down action in the MV during the 60–80 s period, even though the CV remains about at the set point. The disturbance is unmeasured, not included in the controller model, but affecting the process only. This additional source of the process-model mismatch is also evident in the 20–40 s period where the pmm changes during the 20–40 time period.

112

Fig. 6 Events in AUTO mode—2

Fig. 7 Events in AUTO mode—3

Fig. 8 Events in AUTO mode—4

R. R. Rhinehart

Nonlinear Model-Predictive Control Using First-Principles Models

113

The pmm is filtered to temper the impact of measurement noise. The noisy CV is not filtered, which has a benefit of being able to display the noise pattern and the real state of the process. If the process CV were filtered creating a significant lag, then the modeled CV should be equivalently filtered also.

5 End Notes The model does not have to be perfect, great, very good, or even good. A fair model works as this simulation reveals. As additional evidence that imperfection works: We use linear constant FOPDT models for nonlinear high-order processes. We use linear FIR models in APC (MPC, DMC). We steer cars through curves and catch fly balls by continually correcting our mental intuitive “models.” Implement advanced control in stages. Start with the model shadowing the process to affirm model validity. Test bumpless transfer. Then observe the controller suggesting action. If confident that it is making good decisions, let the controller take action while being supervised. Finally, implement autonomous control, but displaying goodness of control messages. Implement HPC showing the operator the horizon-predictive constraint-avoidance predictions. Otherwise, the decoupling and predictive consideration might lead to unexpected decisions and operator discomfort. There may be MVs removed for temporary maintenance. These should be removed from optimization. Include on–off switches so that the controller understands the temporary condition. Be sure that measurements are valid. Faulty inputs lead to bad decisions. This is not different for any level of complexity in control. You may need sensor redundancy or data reconciliation for critical measurements. When the process changes (perhaps piping or units are reconfigured), you may need to change the model. Provide a trap for any possible infeasible calculations and a Plan B action. Cascading MPC outputs as set points to lower-level PLC or DCS controllers is one strategy to keep the process running even if the supervisor MPC gets stuck. K.I.S.S. Keep it Simple and Safe. If gain-scheduled PI works use it, if classic advanced regulatory control solves the problem use it, if GMC-SS solves the nonlinearity use it. Consider incremental model parameter adjustment online to make the model adapt to the process. Observation of the adapting model coefficient can be a clue for predictive maintenance, and the model will be better at forecasting the constrained limits of process operation. The justifications for first-principles models in control include: • One-model consistency for many uses. • Preserving process mechanistic knowledge. • Process analysis, monitoring, and constraint forecasting.

114

R. R. Rhinehart

• Handling nonlinearity, non-stationarity, and disturbances that invalidate linear models. • Ease of tuning. • Ease and reduced process testing for model calibration (adjustment). • Wider valid modeling range than with linear models. How to economically justify using a nonlinear model? APC (linear multivariable control) experience shows a halving (or better) of process variance, which permits operating closer to constraints and specifications. This permits higher throughput, reduced waste, higher quality, lower operating costs, etc., which provides the economic benefit. In my assessment, any control scheme (ARC, Fuzzy, APC, NMBC) that has equivalent disturbance-rejecting, feedforward, and decoupling features has the same benefits as any other controller with those same features. The nonlinear model-based controllers will have those benefits. As well, it will work over a wider operating range, require fewer recalibrations, and be less expensive to recalibrate than controllers with empirical models.

Reference 1. Rhinehart RR (2018) Engineering optimization: applications, methods, and analysis. Wiley, New York, NY

DC Motor System Identification and Speed Control Using dSPACE Tools S. Menaka and S. Patilkulkarni

Abstract This paper presents the system identification and speed control technique of DC motor using dSPACE tools. The two different methods for the speed measurement of the DC motor using dSPACE encoder and the hardware interrupt blocks are discussed. The system identification of DC motor is done using the dSPACE Control Desk 5.3 and system identification toolbox. The DC motor speed controller is implemented using dSPACE hardware tools by PWM technique, where the speed of the DC motor is controlled by controlling the motor’s terminal voltage with the help of duty cycle of PWM. The control algorithm is developed using MATLABSimulink and the Real time interfacing (RTI) is done by using dSPACE DS1104 controller board and CP1104 controller panel. The data capturing, visualization and layout designing is done by using Control Desk 5.3. The developed model can be used for Rapid control prototyping (RCP) and testing purposes for DC motor related applications. Keywords DC motor · CP1104 · MATLAB Simulink · Control Desk 5.3 · System identification · RTI · PWM · dSPACE · RCP

1 Introduction DC motors in industrial control systems have many applications as they are straightforward to operate and model. Sometimes an accurate DC motor model utilized in a control system may be necessary for analytical analysis and optimization of the control system design. In this scenario, the reference values of the DC motor characteristics as specified in the motor specification which is generally provided by the manufacturer of the motor might not be regarded sufficient particularly for S. Menaka (B) · S. Patilkulkarni JSS Science and Technology University, Mysuru, India e-mail: [email protected] S. Patilkulkarni e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_7

115

116

S. Menaka and S. Patilkulkarni

the more affordable DC motors, whose electric and mechanical parameters tend to be quite high. DC motor model identification may be accomplished using general system identification methods [1–3]. In the absence of costly testing equipment and a lengthy testing cycle, a rapid and effective technique of system identification [4– 7] based on the input and output values (i.e., the voltage and speed values) of the motor is desirable and advantageous, particularly for field applications and quick controller development [8]. It is necessary to load a specific input signal and identify the system’s output [9] in order to make an approximation to obtain the mathematical model of the system. This permits the design and insertion of PID controller to a closed-loop system model [10, 11]. By the auto tuning process of the PID controller, PID parameters are created [12]. The modified PID controller can then integrated into the hardware in the loop (HIL) simulation of the control system [13], and the performance of the system’s response can be assessed. From the literature review for the system identification process different models such as Process model, ARX, ARMAX, OE model, State space model [14], non-linear models were used. Among all of the models used the process model is considered in our work since it is less complex and is sufficient for second order system and the model coefficients have an easy interpretation as poles and zeros. Other models can also be considered for estimating transfer function to achieve more accuracy which is the future scope of the work. In this paper, a DC motor speed controller system using PWM technique is proposed and implemented, and its performance is analyzed. The system identification process is also discussed. The controller is developed using MATLABSimulink and implemented in real time using dSPACE Control Desk and hardware tools (DS1104 controller board and CP1104 controller panel). The dSPACE tool helps in developing a virtual controller for easy prototyping with making use of less hardware components. The proposed system uses voltage, current and speed sensor for measuring the input and output signal of the DC motor.

2 Methodology Figure 1 describes the proposed block diagram of the experiment. The voltage sensor is used to detect the voltage drawn by the motor. Output of voltage sensor is given to ADC of ds1104 controller board and is then calibrated using the gain blocks to obtain the motor voltage output. The current sensor is used to detect the current drawn by the motor. The output of current sensor is given to ADC of ds1104 controller board and is then calibrated using the gain blocks to obtain the motor current output. The quadrature incremental magnetic encoder is used to obtain the speed feedback of the motor. The encoder channel outputs is given to the encoder or interrupt pin of ds1104 controller board based on the speed technique used and is calibrated using the gain blocks to obtain the output in terms of rpm value of the motor. The experimental data of voltage, current and speed values are captured using Control Desk and are used to obtain transfer function by system identification method. The speed control

DC Motor System Identification and Speed Control Using dSPACE Tools

117

Fig. 1 The proposed block diagram

of DC motor is implemented by using the pulse width modulation technique (PWM), by applying variable duty cycle PWM signal to the motor driver circuit which also controls the motor terminal voltage.

2.1 Voltage and Current Measurement RTI Simulink model for voltage sensing is shown in Fig. 2. For analog inputs, the dSPACE unit has an integrated gain of 0.1. Any signal read from a dSPACE ADC block is multiplied by 0.1 in Simulink. Hence a gain value of 10 is applied on the output of an ADC block to compensate for this. The 5 V range obtained across the analog pin of voltage sensor lets recognizing of voltages up to 25 V with the help of controller. Hence a gain value of 5 is multiplied to obtain the voltage value drawn by the motor. RTI Simulink model for current sensing is shown in Fig. 3. The dSPACE device features an inbuilt gain of 0.1 for analog inputs. In Simulink, each signal received from a dSPACE ADC block is multiplied by 0.1. Therefore, a gain value of 10 is used to compensate for this in the output of an ADC block. The voltage observed across VCC of ASC712 is around 4.86 V. When there is no current flow through the terminals of the module, the output voltage will be half of the supply voltage, i.e., 2.4325 V. This implies subtracting 2.4325 V from the voltage recorded at the ADC pin. The current value is calculated by dividing this value by the sensitivity value of the sensor. The below equation is used to calculate current drawn by the motor Currentvalue = (adcvoltage − offsetvoltage)/Sensitivity

Fig. 2 The RTI Simulink model for voltage sensing

(1)

118

S. Menaka and S. Patilkulkarni

Fig. 3 The RTI Simulink model for current sensing

2.2 Speed Measurement of DC Motor Using Incremental Magnetic Encoder 2.2.1

Speed Measurement Techniques

There are two techniques for measuring the speed using an incremental encoder they are – Pulse counting (Frequency measurement) – Pulse timing (Period measurement). Pulse Counting (Frequency Measurement) The conventional and perhaps easiest way to determine rotor speed is by measuring the frequency of the pulses of the encoder. In a given and constant time window, the number of pulses observed is counted. Pulse counting method determines the average pulse time. For the calculation of the angular velocity ω, say the count is n pulse for a sample period T. Thus, the average pulse time is T /n. If N windows are available on the disc, NT /n is the average time of one rotation. Therefore, the angular velocity is given by ω = 2π n/N T

(2)

where ω n T N

angular velocity pulse count sample period pulse per revolution.

Pulse Timing (Period Measurement) The high frequency clock signal is measured with the pulse timing method during a single encoder cycle. Assume the frequency of the clock is f Hz. If m clock cycles are counted over an encoder cycle, m/f will be the time taken to complete one encoder

DC Motor System Identification and Speed Control Using dSPACE Tools

119

cycle. The average time for one disc revolution is Nm/f when there are a total of N windows on the track. Therefore, the angular velocity is given by ω = 2π f /N m

(3)

where ω f m N

angular velocity clock frequency number of clock cycles pulse per rotation.

2.2.2

The Implementation of Simulink RTI Model for the Speed Measurement of DC Motor

Method 1: Speed measurement using encoder of ds1104 board (Pulse counting method). In this method speed is calculated based on the number of pulses obtained during a certain amount of time. The block “DS1104ENC_POS_C1” gives read access to two encoder interface channel’s position and delta position. The encoder master configuration block defines the global characteristics for the encoder interface channels. The encoder signal type is set to single ended (TTL) in the “Encoder master setup” block. The encoder has a disc of 51.42° for each line, with 7 lines per revolution. The line count is multiplied by the fixed number 51.42 to convert it to degrees. The “Encoder delta position” output is used to compute the angular velocity of the motor in rad/s by dividing it by the sampling time as in Fig. 4. The angular velocity is converted to rpm by multiplying the rad/s value with 60/2π. The rpm values obtained in this method was not accurate because in a fixed sample time the sufficient information of the pulses is not obtained as a low-speed motor is used. This method is best suited for high-speed motors. Method 2: Speed measurement using hardware interrupt of ds1104 controller board (pulse timing method). The implementation of time measurement is being carried out by the function call subsystem which is triggered by the DS1104MASTER_HWINT_I1 external hardware interrupt block as shown in Fig. 5. The channel A of the incremental magnetic encoder is connected to the external hardware interrupt pin of the ds1104 controller board. The input port of the function call subsystem, i.e., “Angle” denotes the fixed angle between two successive pulses of the sensor which is calculated using the equation 2π /(impulses per revolution). Figure 6 depicts the proposed contents of the function call subsystem for the time measurement. The previous time and the time difference value between two pulses of the encoder are being stored in the data store memory and data store read blocks. The system outputs block has been taken from the custom code library.

120

S. Menaka and S. Patilkulkarni

Fig. 4 The RTI Simulink model for speed measurement using encoder

Fig. 5 The RTI Simulink model for speed measurement using hardware interrupt

The following code is entered in the System Outputs block // declare local variable double currentTime; // read timestamp currentTime = ts_time_read(); // calculate difference to previous step deltaT = currentTime – oldTime; // save new timestamp oldTime = currentTime;

The ts_time_read() function reads the absolute simulation time from the hardware through which the currentTime and the oldTime values are obtained. The time difference deltaT and the previous time of the encoder pulse oldTime is written to the corresponding data store memory blocks. Thus the speed of the motor is calculated by dividing the fixed angle with the measured time difference deltaT, i.e., (2 * pi)/(impulses per revolution * deltaT).

DC Motor System Identification and Speed Control Using dSPACE Tools

121

Fig. 6 The Simulink model of function call subsystem block

3 System Identification of DC Motor The experimental voltage and speed values of the DC motor captured using the dSPACE control desk are exported to the MATLAB workspace. The transfer function of the DC motor is identified by system identification toolbox as shown in Fig. 7. Several models such as Process model, Polynomial models transfer function model, state equation model, non-linear models are available to estimate the plant model. In this work the process model structure was incorporated for the identification process because the order of the model was easier to anticipate as the DC motor model is a second order system. The reason why the ARX model is being ignored is that the estimated model is a prediction model and is not directly dependent on the input and output values which might cause difficulty in building the controller. The obtained transfer function model has the best curve fit of 84% between the model response and the measured output. Figure 8 shows the flowchart of the system identification process which depicts the steps involved during the identification process. The obtained transfer function from the system identification process is given below

122

S. Menaka and S. Patilkulkarni

Fig. 7 System identification toolbox

G(s) =

83.2 (1 + 0.01534s)(1 + 0.00234s)

(4)

The Transient response, frequency response and the pole-zero plots of the obtained transfer function from the system identification process is as shown in Figs. 9, 10 and 11, respectively.

4 Controller Implementation and Result The Pulse Width Modulation (PWM) method can regulate the efficient armature voltage using ON-OFF control technique. The greater duty cycle pulse turns ‘ON’ more long than the lower one. The duty cycle δ, is regulated by the following equation Duty cycle (δ) =

Pulse on time or pulse width (ton ) Pulse time period (T )

(5)

The ratio of ON to OFF time is known as the duty cycle, and it influences the motor’s speed. By varying the duty cycle the desired speed can be obtained. PWM pulses are used to control the duty cycle of a motor driver circuit. A square wave with a constant voltage but variable pulse width or duty cycle is used to power the motor. The controller implementation using pulse width modulation technique is as shown in Fig. 12. The PWM pulse is generated using DS1104SL_DSP_PWM3 block. Since

DC Motor System Identification and Speed Control Using dSPACE Tools Fig. 8 Flowchart of system identification process

Fig. 9 Step response

123

124

S. Menaka and S. Patilkulkarni

Fig. 10 Bode plot

Fig. 11 Pole zero map

the frequency of the PWM pulse is kept constant to 1 kHz but the on-off duration varies, the determination of duty cycle of PWM is based on the pulse width of the pulse. As pulse width increases, the duty cycle of PWM also increases which in turn increases the power delivered to the motor driver circuit. The duration of ON period of PWM is longer at 100% duty cycle than at 50% duty cycle. This result in a faster motor speed at 100% duty cycle compared to 50% duty cycle. The following equation expresses the relationship between average voltage, supply voltage, and duty cycle.

DC Motor System Identification and Speed Control Using dSPACE Tools

V(average) =

δ

125

(6)

V(supply)

Table 1 shows experimental results of the DC motor speed and voltage values obtained at different values of duty cycle. The duty cycle was varied from 0 to 1 and the corresponding voltage and speed values were noted down. As the duty cycle increases the DC motor’s speed also increases. The motor speed is 295 RPM at 30% duty cycle, and the corresponding output voltage of the converter is 3.62 V. When the duty cycle is set to 50%, the motor’s speed is 498 RPM, and the resulting output voltage of the converter is 6.14 V. As the duty cycle approaches 70 and 90%, the motor’s speed and the converter output voltage increases. The graphs shown in Figs. 13 and 14 depicts the relation between duty cycle, voltage and speed.

Fig. 12 The controller implementation model of DC motor

Table 1 Experimental results Duty cycle (%)

Voltage (V)

0

0

Speed (RPM) 0

0.3

3.62

295

0.4

4.85

393

0.5

6.14

498

0.6

7.25

595

0.7

8.47

698

0.8

9.64

794

0.9

10.82

896

1

12.3

997

126

S. Menaka and S. Patilkulkarni

Fig. 13 Duty cycle versus speed

Fig. 14 Duty cycle versus voltage

Figure 15 shows the GUI developed using control desk for capturing the voltage, current and speed data’s of the motor.

5 Conclusion and Future Scope The two different methods for speed measurement of motor were implemented. It was found that the pulse counting method using encoder block is best suited for high-speed motors and the pulse timing method using hardware interrupt block is best suited for low-speed motor. The process of system identification is discussed, and the transfer function of the motor was achieved. The PWM method was used to create a motor speed controller system. The relation of duty cycle against voltage and speed parameters was investigated. It has been discovered that the greater the pulse width, the higher the average voltage supplied to the motor terminals. As a result, the magnetic flux inside the armature windings becomes greater. As a result, the motor will revolve faster. By varying the duty cycle of the PWM pulse, the DS1104 controller gives flexibility in managing the speed. The impact of PWM pulse width on DC motor voltage and speed values has been investigated. The accuracy of the obtained transfer function from system identification process can be increased by considering other models for estimation. The closed-loop controller can be implemented for the transfer function obtained from the system

DC Motor System Identification and Speed Control Using dSPACE Tools

127

Fig. 15 The control desk layout designed for the experiment

identification process and can be implemented on hardware. The cascade control of the motor can also be incorporated by considering current also as feedback along with speed. The same model can be worked on and implemented using dSPACE MicroAutobox II.

References 1. Awoda M, Ali R (2019) Parameter estimation of a permanent magnetic DC motor. Iraqi J Electr Electron Eng 15(1):28–36 2. Sankardoss V, Geethanjali P (2017) Parameter estimation and speed control of a PMDC motor used in wheelchair. Energy Procedia 117:345–352 3. Ramasubramanian D (2016) Identification and control of DC motors. Master’s thesis in automatic control and robotics, Barcelona School of Industrial Engineering, ETSEIB 4. Bature A (2013) Identification and real time control of a DC motor. IOSR J Electr Electron Eng 7(4):54–58 5. Becedas J, Mamani G, Feliu V (2010) Algebraic parameters identification of DC motors: methodology and analysis. Int J Syst Sci 41(10):1241–1255 6. Othman K, Kamal M (2009) System identification of discrete model for DC motor positioning. In: Recent advances in circuits, systems, electronics, control and signal processing, pp 212–216 7. Hassan MM, Aly AA, Rashwan AF (2007) Different identification methods with application to a DC motor. J Eng Sci 35(6):1481–1493 8. Tang WJ, Liu ZT, Wang Q (2017) DC motor speed control based on system identification and PID auto tuning. In: Chinese control conference, CCC, 61403422, pp 6420–6423 9. Ljung L (1998) System identification. In: Procházka A, Uhlíˇr J, Rayner PWJ, Kingsbury NG (eds) Signal analysis and prediction. Applied and numerical harmonic analysis. Birkhauser, Boston, MA

128

S. Menaka and S. Patilkulkarni

10. Joshi B, Shrestha R, Chaudhary R (2014) Modeling, simulation and implementation of brushed DC motor speed control using optical incremental encoder feedback. In: Proceeding of IOE graduate conference, Nov 2014, pp 497–505 11. Fliess M, Sira-Ramirez H (2008) Closed-loop parametric identification for continuous-time linear systems via new algebraic techniques. In: Identification of continuous-time models from sampled data. Springer, pp 363–391 12. Dewangan AK, Chakraborty N, Shukla S, Yadu V (2012) PWM based automatic closed loop speed control of DC motor. Int J Eng Trends Technol V3(2):110–112 13. Wu W (2010) DC motor parameter identification using speed step responses. In: Proceedings of the American control conference 2012, no 3, pp 1937–1941 14. Basilio JC, Moreira MV (2004) State-space parameter identification in a second control laboratory. IEEE Trans Educ 47(2):204–210 15. Rahmani F, Quispe D, Agarwal T, Barzegaran M (2020) Speed control of brushless DC motor by DC-DC boost and buck converters using GaN and SiC transistors for implementing the electric vehicles. Comput Res Prog Appl Sci Eng 06(02):70–75 16. Amalrajan R, Gunabalan R, Tewari N (2020) DSPACE1103 controller for PWM control of power electronic converters. In: Lecture notes in electrical engineering, vol 687, pp 307–322 17. Fatma MW, Hamid MI (2019) PWM speed control of DC permanent magnet motor using a PIC18F4550 microcontroller. IOP Conf Ser Mater Sci Eng 602(1) 18. Hamzi B, Owhadi H, Koch J, Naung Y, Schagin A, Oo HL, Ye KZ, Khaing ZM, Simani S, Alvisi S, Venturini M (2021) Implementation of data driven control system of DC motor by using system identification process. Electronics (Switzerland) 8(2):1801–1804 19. dSPACE DS1104 control workstation & Simulink tutorial (2018) Utah University, pp 1–38 20. Tatenda Katsambe C, Luckose V, Shahabuddin NS (2017) Effect of pulse width modulation on DC motor speed. Int J Students’ Res Technol Manag 5(2):42 21. Agung IGAPR, Huda S, Wijaya IWA (2014) Speed control for DC motor with pulse width modulation (PWM) method using infrared remote control based on ATmega16 microcontroller. In: Proceedings—2014 international conference on smart green technology in electrical and information systems: towards greener globe through smart technology, ICSGTEIS, pp 108–112 22. Gunda KK (2008) Adjustable speed drives laboratory based on dSpace controller. MSc thesis, Louisiana State University 23. Mamani G, Becedas J, Feliu-Batlle V, Sira-Ramirez H (2007) Open-loop algebraic identification method for a DC motor. In: 2007 European control conference (ECC), pp 3430–3436 24. Petrella R, Tursini M, Peretti L, Zigliotto M (2007) Speed measurement algorithms for lowresolution incremental encoder equipped drives: a comparative analysis. In: International Aegean conference on electrical machines and power electronics and electromotion ACEMP’07 and electromotion’07 joint conference, pp 780–787 25. Azli NA, Bakar MS (2004) A DSP-based regular sampled pulse width modulation (PWM) technique for a multilevel inverter. In: International conference on power system technology, POWERCON, vol 2, pp 1613–1618 26. Lord W, Hwang JH (1977) DC servomotors-modeling and parameter determination. IEEE Trans Ind Appl 3:234–243

Oil Quality Analysis Using Image Processing Nivedita Daimiwal, Revati Shriram, Harish Shinde, Radhika Kulkarni, and Apeksha Galewad

Abstract Now-a-days, people are demanding good quality of cooking oil (edible oil). For a healthy heart, quality of cooking oil we consume plays a very important role. As the cost of good quality cooking oil is more, there is a possibility of adulteration of oil by the traders. FSSAI also decided to crack down on the sale of adulterated edible oil. The campaign was made by FSSAI all over India in August 2020. But, the test results are expected in a month’s time. The aim of the project is to determine the change in various parameters of an oil based on the presence of adulterants or some other oil residues. Determination of the presence of adulterants would be helpful in avoiding many health conditions. Changing the ratios of the oils in the mixture can show a change in the parameters that are being measured as the chemical contents of the oils are changed. The objective is to design a system to detect the adulteration and its percentage using machine learning method in various types of oil using oil image. So, the initiative is made to develop a machine learning image processing system. In this method, oil images were used to detect the adulteration in the oil. Features are collected for different ratios of adulteration, and these features obtained by image processing are used for detection of adulteration and its percentage. Variation in the features like RGB, mean, variance and entropy is measured and plotted in Minitab for various percentage of adulterations. Keywords Pure groundnut (wooden pressed edible oil) · Refined oil (chemically treated) · Quality · Adulterants · RGB · Entropy · Mean · Variance

1 Introduction Oil and its quality-related issues have been dealt with by researchers worldwide for more than 50 years. Contact-based and non-contact-based many techniques are tried and tested till date. Quality of edible oil is checked mainly for the reasons like to look for adulteration in the oil, to decide its suitability from the point of N. Daimiwal (B) · R. Shriram · H. Shinde · R. Kulkarni · A. Galewad Cummins College of Engineering for Women, Pune 411052, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_8

129

130

N. Daimiwal et al.

view of cardiovascular health of a person, then use of reduced quality edible oil after multiple frying sessions for the vehicles as a biodiesel, etc. Depending on the objective to check the oil quality, various methods are tried and tested by the researchers internationally and nationally. The oil is image is capture by using digital camera and various statistical parameters are extracted from the oil image by using image processing techniques. Results show the variation in the statistical parameters for different percentages of adulteration. These features can be used for adulation detection and checking oil quality from oil images [1, 2].

1.1 Methodology The oils that were used for observing the results were—pure groundnut, pure palm, refined groundnut and a 50–50 mixture of refined groundnut and pure palm. A total of 10 ml was considered as 100%. Refined groundnut, pure groundnut and pure palm were taken in a 10-ml measurer. For a 50–50 ratio of the oils, 5 ml of the measurer was filled with refined groundnut oil, and the rest 5 ml was filled with pure palm oil. A clear picture of the 100% oil and the 50–50 mixture was taken with the help of a camera of resolution 12 megapixel (f /1.8) + 12 megapixel (f /2.4) + 12 megapixel (f /2.0). The parameters that are being measured are—RGB, entropy, mean, variance.

1.2 Review of Literature Babatunde proposed image processing based on a chemical method. Babatunde proposed image processing based on chemical method for identification of sunflower and groundnut oil. Using this method of inspection for oil is too long and hard to satisfy the oil circulation demands. Also, application of this method for inspection is expensive. Although this method has many limitations, it can give more exact results [3]. Shariff performed analysis on Agilent 1260 Infinity Analytical SFC System combined with an Agilent 1260 Infinity ELSD Evaporative Light Scattering Detector and proposed an image processing technique based on that. The ELSD was coupled to the SFC module. To obtain good sensitivity and reproducibility, the addition of a make up flow before the back pressure regulator with additional heating at the entrance of ELSD is necessary [4]. Ariffin proposed the estimation of quality analysis on oil on the basis of color and texture with the help of image processing. A digital camera having high resolution is used to capture an image. After getting the image, the noise from the image is removed by applying preprocessing technique. From that source image, the color and texture characteristics of oil are extracted. The color characteristics are analyzed using RGB and HSI analysis. Texture features like entropy are extracted [5].

Oil Quality Analysis Using Image Processing

131

Pearson proposed a technique of supercritical fluid chromatography (SFC) in combination with evaporative light scattering detection (ELSD) for image processing which is important for the determination of triglyceride composition of vegetable oils. SFC separates triglycerides at much lower temperatures, and SFC permits greater selectivity with shorter analysis times as compared to GC and HPLC, respectively [6]. Heinemann proposed a capacitive oil sensor technique for image processing. The sensor from Testo can make measurements in hot oil because of its ergonomic design. The user is protected from heat, protected against water jets without a protective cover, and there is also a visual alarm indication. Frying with cooking oil is not healthy after couple of turns of use. Using this sensor when the TPM (Total Polar Material)/TPC (Total Polar Compound) value is reached, the cooking oil can be changed [7, 8]. Work carried out for the oil quality checking at national level in the last decade is as follows: Choudhary et al., India: Study presented the conventional and non-conventional edible oils in India with the objective of how difficult it is to choose the correct edible oil [9]. Prema, Department of Commerce, Coimbatore, India: Empirical study of the oil brand preference in rural India was presented by the author. Study did not present any oil property-related analysis [10].

1.3 Digital Image Processing Image processing involves transforming an image into digital form and performing various operations on it in order to extract desired information or to get an enhanced image. The input in this process is the image itself, but the output is the desired part of the image. The images are treated as a two-dimensional signal, and then, signal processing methods are applied on it by any image processing system. Image processing can be used for pattern recognition, to define various objects in an image, to detect objects in an image, to retrieve high quality images, to improve noisy images in order to perform image sharpening and also to detect and visualize objects that are difficult to see.

2 System Specifications The tool used for image processing is MATLAB. MATLAB is a multi-paradigm programming language and a numeric computing environment. It has a built-in toolbox for image processing which helps in maintaining the efficiency of the code. The advantages of using MATLAB for image processing are as follows:

132

N. Daimiwal et al.

• MATLAB is a scientific language and hence helps implementation of advanced algorithms through mathematical and numerical support. • The final result has maximum numerical precision. • The image processing steps can be documented and replicated at any given time. • Image processing algorithms are more advanced in MATLAB than any other language. Version—MATLAB R2021a Software—Windows 10 Hardware—Intel core i5 processor, minimum 2 GB RAM.

3 Observations 3.1 RGB Analysis RGB—values of red, green and blue hues in a particular cropped portion of an image. Entropy—measures the degree of randomness of an image. Shown in Eq. 1. Mean—gives the contribution of individual pixel intensity for the entire image. Shown in Eq. 2. Variance—variance tells us how each pixel is different from the neighboring pixel and is used to classify different regions of the image.   Entropy = sum p.∗ log 2( p)

(1)

where p contains the normalized histogram counts Mean = Sum of pixel values/Total number of pixel values

(2)

3.2 Methodology A cropped portion measuring 10 mm * 10 mm having the same coordinates was selected from the image. The analysis was performed on that cropped portion for extracting statistical parameters like mean, variance and entropy. Analysis of pure groundnut (Wooden pressed) and 50–50 mixture of pure palm and pure groundnut oil. Figure 1 shows steps incorporated in the oil image analysis. Steps of Image Processing: Technique used for feature extraction: • Read the image from the location • Display the read image • Select the cropped portion from the image

Oil Quality Analysis Using Image Processing

133

Fig. 1 Flowchart of methodology

• • • • • • • • • • • •

Display the cropped portion Get the information about individual pixels in the image Calculate the entropy of the cropped image Get red hues for RGB analysis Get green hues for RGB analysis Get blue hues for RGB analysis Find mean of the 2-dimensional cropped image Find average of the red pixels Find average of the green pixels Find average of the blue pixels Convert the type of cropped image into double Calculate the variance.

For calculating the variance, it is mandatory that the image is of the type double. This same code was used for the analysis of both pure groundnut and the 50–50 mixture of two oils. Oil image of pure groundnut oil and 50% adulteration oil (wooden pressed groundnut and palm oil) captured using mobile camera are shown in Figs. 2 and 3, respectively. Table 1 shows the statistical parameter values. Analysis of pure palm and refined groundnut oil (chemically treated) • • • • •

Assume two inputs and take image of jpeg format through location If filename is not of suitable format then it will return null Assign filename and path to variable file path Read file and storing it in variable Show image on display

134

N. Daimiwal et al.

Fig. 2 Pure groundnut oil (wooden pressed)

Fig. 3 50–50 mixture of pure palm and pure groundnut oil

Table 1 Observations table of pure groundnut oil (wooden pressed) and 50% adulteration with palm oil Parameter

Pure groundnut

50–50 mixture

RGB

145.06, 135.48, 57.20

155.07, 130.37, 61.59

Mean

112.7033

115.5109

Variance

1.6397e + 03

1.5669e + 03

Entropy

5.4382

4.9049

• • • • • • •

Crop given image with specified pixel values Calculate total entropy of given image Get RGB values from given image Calculate mean RGB values Calculate average mean Convert image to double values Show above details.

Oil Quality Analysis Using Image Processing

135

Oil image of refined groundnut oil (chemically treated) and pure palm oil is shown in Figs. 4 and 5 respectively. Statistical parameters of refined ground nut oil and pure palm oil are given in Table 2. The work is carried out for different percentage of adulteration of palm oil in refined ground nut oil. Table 3 shows the parameters (mean, variance and entropy) for 10%, 20%, 50% and for pure ground nut oil and palm oil. Fig. 4 Refined groundnut oil (chemically treated)

Fig. 5 Pure palm oil

Table 2 Observations from the analysis Parameter

Refined groundnut

Pure palm

RGB

138.311, 119.68, 6.337

119.290, 110.26, 7.32

Mean

78.1925

149.67

Variance

4.19e + 03

3.208e + 03

Entropy

6.5201

6.245

136

N. Daimiwal et al.

Table 3 Comparison of different percentages of impurities Parameter

10% adulteration

20% adulteration

50% adulteration

100% pure

Mean

123.3436

107.3899

115.5109

149.67

Variance

3.7230e + 03

1.5592e + 03

1.5669e + 03

3.208e + 03

Entropy

6.672

5.9573

4.9049

6.245

4 Conclusion Various adulterations like mixing of two different oils can be detected. System of digital image processing extracts the feature from oil images. The features extracted are RGB, mean, variance and entropy. The variations in the features were observed for oil adulteration. It is observed that the mean of pure groundnut oil (wooden pressed) is low as compared to the 50% adulteration with palm oil. The variance of pure groundnut oil is greater than that of adulterated refined oil. The entropy of pure groundnut oil (wooden pressed) is greater than oil adulterated with palm oil. For the pure 100% palm and refined oil, the observations were made. The mean of pure palm oil is greater than that of refined groundnut oil (chemically treated), and the variance is also greater. The entropy of refined groundnut oil is slightly greater than palm oil. Based on the database of oil adulteration and corresponding features, we can use the system for preliminary detection of oil adulteration using a mobile camera. Mobile application can be developed using the same.

References 1. Kabagambe EK, Baylin A, Ascherio A, Campos H (2005) The type of oil used for cooking is associated with the risk of nonfatal acute myocardial infarction in Costa Rica. J Nutr 135(11):2674–2679 2. Rashvand M, Akbarnis A (2019) The feasibility of using image processing and artificial neural network for detecting the adulteration of sesame oil. AIMS Agric Food 4(2):237–243 3. Babatunde OO, Ige MT, Makanjoula GA (1988) Effect of sterilization on fruit recovery in oil palm fruit processing. J Agric Eng Res 41:75–79 4. Shariff AR, Nor AA, Mispan R, Shattri M, Rohaya H, Goyal R (2002) Correlation between oil content and DN values. In: MAP Asia conference. Geospatial application papers, India 5. Ariffin AA (1991) Chemical changes during sterilisation process affecting strippability and oil quality. In: Seminar on developments in palm oil milling technology and environmental management. Palm Oil Institute of Malaysia (PORIM), Bangi 6. Pearson YT (1996) Machine vision system for automated detection of stained pistachio nuts. J Food Sci Technol 19(3):203–209 7. Heinemann PH, Hughes R, Morrow CT, Sommer HJ, Beelman RB, Wuest PJ (1995) Grading of mushrooms using a machine vision system. Trans ASAE 37(5):1671–1677 8. Zahir E, Rehana S, Mehwish AH, Anjum Y (2014) Study of physicochemical properties of edible oil and evaluation of frying oil quality by Fourier transform-infrared (FT-IR) spectroscopy. Arab J Chem

Oil Quality Analysis Using Image Processing

137

9. Choudhary M, Sangha JK, Grover K (2014) Conventional and non-conventional edible oils: an Indian perspective. J Am Oil Chem Soc 91(2):179–206 10. Prema R (2013) Empirical study of the oil brand preference in rural area. Indian J Appl Res 3(3)

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive Moment for Washing Machine S. Elavaar Kuzhali, Kotha Manvitha, Anisha Singh, Lakshmi Pranathi, Shreya Dhavule, and M. Poorvita

Abstract There are diverse fabric materials available in today’s world, which need various washing methods to make them last longer. However, present technology does not have an efficient mechanism for managing numerous fabrics when it comes to washing these fabrics in the washing machine. Therefore, it is important to classify different fabrics and give the optimized wash according to the quality of the fabrics. So, clothes quality and lifetime are increased. If a neural network model is trained to learn these features that are inherent to the fabric structures, by exhibiting a set of cotton and silk fabric images, then the neural network would be able to correctly distinguish between different fabric images. To implement this a convolutional neural network (CNN) algorithm is used. The hardware part of the project includes a Raspberry Pi, which is connected to a camera sensor, a temperature sensor, a motor, and a water level sensor for controlling the level of water in the washing machine. It is also automated to supply the required amount of soap and water based on different fabrics. Keywords Fabric classification · Adam · Nadam · Automation · CNN · VGG16

1 Introduction Fabric is one of the ancient human inventions [1] that has evolved from handcrafted woven textiles to modern machine-based electronic textiles. In the textile industry, the weave pattern, which is the most important factor for fabrics, plays an important role in design and redesign, analysis of the texture for structure, and appearance of the fabrics [2–5]. Clothes are made from different fabrics, and each fabric has its preferred method of washing technique. Segregating clothes based on their fabric is the first step to ensure that the fabrics are healthy and durable. It is essential to perform recognition of different types of fabric for different wash method for each fabric type. Currently, there is no effective method to identify fabrics and wash them according to S. Elavaar Kuzhali (B) · K. Manvitha · A. Singh · L. Pranathi · S. Dhavule · M. Poorvita Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_9

139

140

S. Elavaar Kuzhali et al.

their required conditions. This long-term practice is detrimental to the durability of the fabric. Presently, fabric-type recognition is relied on the manual operation using human eyes and touch. Traditionally, this manual inspection is performed by experts who require knowledge and experience. However, along with various inconveniences such as heavy workload, low efficiency, and time consuming, it also leads to subjective human factors, such as physical and mental stress, dizziness, and fatigue, which ultimately affect the recognition results. Therefore, it is required to develop a set of fabric-type recognition automation to produce high-quality products that meet customer needs. The traditional washing machine requires the user to manually set all of the criteria required for optimized and efficient washing. Consequently, there is a need for a clever washing device that can not only identify the given fabric type but also set the washing parameters automatically based on the fabric type. Many researchers have used neural networks and artificial intelligence (AI) to build models which identify and classify the type of fabric. Few examples are fabric defect inspection system [6–8] that uses image processing technique to detect good and faulty fabrics using the distribution of the areas of super pixels, image retrieval system using different methods like convolutional neural network [9], sketch-based system and with the help of metadata such as tags [10–12], fabric classification [13–16], yarn-dyed fabric identification [17], erosion and dilation-based system for organic fabric [18], feature extraction of different fabrics has been proposed [19–22], and so on. Identifying faults in a limited number of fabric types does not fulfill the functionality of an automatic washing machine, it also needs to identify the type of fabric. This requires a very efficient technique with better accuracy. Therefore, research continues to create the best model which can identify the fabric type with utmost accuracy. Some researchers, namely Kampouris et al. [23, 24] have used a photometric stereo sensor to collect the database of images and used VGG-M neural network. Based on the findings, they determined that cotton, terry cloth, denim, and fleece can all be correctly identified with nearly no errors. Nylon and viscose, on the other hand, are difficult to classify due to their glossy look and tiny array of samples. Anami et al. [18] have presented a paper that focused mainly on the classification of a variety of plants, animals, and mineral origin fabric through images. Erosion and dilation are the morphological operations used. The classification rates are predicted using the ANN classifier. However, most of the existing techniques involve the classification of yarn-dyed fabric, based on its color and woven pattern. But only by having the knowledge of color for washing, a definite washing method cannot be proposed since two different types of fabric with the same color cannot be washed in a similar way. Drawbacks of the existing papers are as follows: • Most of the existing papers concentrate on finding faults in the fabric but do not identify the type of fabric. • Some paper involves identification of the fabric using metadata like tags and labels associated with the images which are unreliable for images without captions.

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

141

• Almost none of the research papers used the output of the model to create an application for the washing machine. The paper presents an application of implementing the neural network algorithm in the washing machine. An effective classification technique for commonly used fabrics (cotton and silk) is developed. Convolutional neural network (CNN) is a type of neural network built for two-dimensional image processing, although it can be used with one-dimensional and three-dimensional data. VGG is a classical convolutional neural network architecture. It is based on an analysis of how to increase the depth of such networks. VGG16 is one of the earliest CNN models used for large-scale image recognition. The model achieves 92.7% top-5 test accuracy in ImageNet, which is a dataset of over 15 million images belonging to 1000 classes. VGG16 is distinguished by the fact that it focuses on having convolution layers of 3 × 3 filter with stride 1 rather than a large number of hyper-parameters, and it always uses the same padding and max pooling layer of 2 × 2 filter with stride 2. This convolution and max pool layer configuration are maintained throughout the architecture. Finally, two fully connected FC (completely connected layers) and a SoftMax layers are included. VGG16 architecture is implemented from scratch using TensorFlow and Keras for the classification of fabrics. A dataset is created for each type of fabric, such as cotton and silk. Fabric categorization is done using convolutional neural network technique. The developed neural network model is put into Raspberry Pi attached to the camera sensor after training, validating, and testing the performance of the convolutional neural network framework on fabric photos. Images of fabrics are captured by the camera sensor for classification. Using a convolutional neural network, this research offers an excellent classification strategy for regularly used materials (cotton and silk). The proposed method makes use of the VGG16 network convolution model, which has a 92.7% accuracy, and it is trained using Adam, SGD, and Nadam optimizer. Since Nadam optimizer gave the best result, it has been used in the proposed model. After training, validating, and testing the performance of the convolutional neural network model on fabric images, the model is deployed into Raspberry Pi which is connected to a camera sensor. The paper presents an application of neural networks, where it is integrated with the washing machine to make it smarter. Major contributions of the paper are as follows: • The proposed model identifies the type of fabric between cotton and silk with Nadam optimizer having a learning rate of 0.00001 to obtain an improved validation accuracy of 98.7% and training accuracy of 99.3%. • The model is deployed with Raspberry Pi to automate the process of washing in the washing machine. The further sections are depicted as follows. Section 2 proposed an approach for the washing machine which consists of two subsections explaining fabric classification using CNN model and circuit design and process flow, respectively. Section 3 discusses the results obtained. Section 4 gives the conclusion of this paper.

142

S. Elavaar Kuzhali et al.

2 Proposed Approach for Washing Machine The proposed method aims at developing a prototype for classification of fabrics and a suitable washing technique based on the load of fabric clothes given. As stated before, the approach is composed of two main steps. In the first step, a CNN model is developed for the classification of given fabric images. The second step consists of design of washing techniques for different sets of loads and setting up of the hardware prototype using Raspberry Pi and other required hardware components. In the following sections, each of these steps has been explained in detail.

2.1 Fabric Classification Using CNN Model To solve the problem of classification of types of fabrics, convolutional neural network is used. Figure 1 shows the flowchart of CNN model. VGG16 is a network convolution model for neural networks that can achieve a test accuracy of 92.7% among the top five on ImageNet. This is a collection of over 15 million photos divided into 1000 categories. The datasets created by image augmentation can be separated into two types: training datasets and validation datasets. Image augmentation is used to enlarge the training dataset and improve the model’s performance and generalization capabilities by creating a new dataset from the input photos. The validation set is only used to evaluate the model’s performance, while the training set is utilized to train it. A dataset of 4331 images of cotton and silk was collected out of which a training dataset comprised 3466 images and a validation dataset comprised 865 images. The amount of data available usually improves the performance of convolutional neural networks. VGG16 model with four convolutional blocks has been proposed by slightly modifying the last three dense layers by including average pooling layer, flatten layer, and ReLU activation layer. VGG16 has a convolutional layer of 3 * 3 filter and a max pool layer of 2 * 2 filter. To down sample the identification of features in feature maps, pooling is required. The pooling methods used by VGG16 include max pooling and average pooling. The maximum value for each patch of the feature map is calculated using max pooling. The average value for each patch on the feature map is calculated using average pooling. Throughout the architecture, VGG16 maintains this arrangement of convolution and pooling layers. It features two fully connected layers at the end. After the convolution, the data is passed on to the dense layer in order to smooth out the convolution from the emerging vector. The images will be flattened to size 224 * 224. The rectified linear activation function (ReLU) activation is used to stop the transmission of negative values through the network. Mathematically, ReLU activation function is expressed as f (x) = max(x, 0)

(1)

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

143

Fig. 1 Flowchart

The SoftMax level generates the value between 0 and 1 based on the class of the image identified. SoftMax is the following function mathematically, where z is a vector of inputs to the output layer and j is an index of output units from 1, 2, 3 … k. ez j σ (z) j = Σ K k=1

ez k

for j = 1

(2)

Different optimizers like stochastic gradient descent (SGD), adaptive moment estimation, and Nesterov-accelerated adaptive moment estimation (Nadam) were used to test the model, and their efficiency was compared. Adam optimizer can be written as

144

S. Elavaar Kuzhali et al.

wt+1 = wt − /

α



Sˆt + ∈

β1 Vˆt−1 +

1 − β1 ∂ L · 1 − β1t ∂wt

 (3)

Nadam employs Nesterov to update the gradient one step ahead by substituting the current in the above equation for the previous: wt+1 = wt − /

α Sˆt + ∈



1 − β1 ∂ L β1 Vˆt−1 + · 1 − β1t ∂wt

 (4)

where Vˆt =

St Vt t , St = 1 − β2t 1 − β1 Δ

(5)

And ∂L ∂wt   ∂L 2 St = β2 St−1 + (1 − β2 ) ∂wt Vt = β1 Vt−1 + (1 − β1 )

(6)

(7)

m and v are both set to zero. The following are the default values (taken from Keras): α = 0.002, β1 = 0.9 β2 = 0.999, ε = 10−7 Out of all the optimizers mentioned, the best results were given by Adam and Nadam. The training and validation accuracy for Adam came out to be 98.25 and 98.96, respectively. The training and validation accuracy for Nadam came out to be 99.3 and 98.7, respectively. Out of these two, Nadam is used to train the VGG16 model since better results were observed using this optimizer. The model is trained until the maximum epochs are reached, and the best model is saved. The dataset ran for 100 epochs with steps per epoch being 70 and a learning rate of 0.00001. The model was successful in classifying the images between cotton and silk.

2.2 Circuit Design and Process Flow The circuit architecture of fabric classification and washing machine prototype is shown in Fig. 2.

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

145

Fig. 2 Circuit architecture of washing machine with fabric classification

The circuit is constructed with two main sections which are input and output. The first section consists of the camera sensor, temperature sensor, float sensor, and IR sensor connected with Raspberry Pi as input devices. The second section consists of a DC motor, liquid crystal display, and water pump connected with Raspberry Pi as output devices. DC motor is operated using a motor driver for which a 9 V supply is given. For temperature sensors also, 9 V supply is used. For water pump, relay is used as a driver or switches to produce the desired voltage. Figure 3 describes about the block diagram of the proposed model. In order to have a longer and safer usage period, the camera sensor in the washing machine that is placed near the rim of the door should be water resistant. The camera sensor used for the prototype is not water resistant, so care was taken to implant the camera sensor to avoid water contact. To get accurate predictions, it is suggested that the camera sensor be placed such that the illumination is uniform and the fabric should be between 4 and 6 cm from the camera sensor lens to capture clear images. The output of the VGG16 model is deployed into the Raspberry Pi microcontroller using TensorFlow Lite framework, where it is programmed to control the operations of a temperature sensor, DC motor, float sensor, and IR sensor. Based on the identified fabrics, the protocol is set for the temperature of the water, washing time, spin rate, and detergent quantity. After the protocol is set, the system will check if the washing machine door is closed or not using an IR sensor. If the door is not closed, it will give an indication using LED light. However, if the door is closed then water will be poured till the set level is reached which will be indicated by the float switch which goes up and the water pump stops. The detergent will be added according to the preset quantity. Simultaneously heater is switched on, the water gets heated till the preset temperature for the particular fabric is reached, and the DC motor will be ON for a specified time having a specified spin rate. The standard specified water temperature, spin rate, and time of wash for cotton are 45 °C, 900–1400 rpm, and 30 min, respectively. The standard specified water temperature, spin rate, and time of wash for silk are 32 °C, 800–1000 rpm, and 35 min, respectively.

146

S. Elavaar Kuzhali et al.

Fig. 3 Block diagram of proposed method

3 Results and Discussion The CNN model for the classification of fabrics is developed using a Python programming language with TensorFlow framework and Keras API for image processing tools. An open-source archive of images was available that was originally prepared for research to acquire a dataset for training the neural network. Dataset consisting of 356 cotton images and 168 silk images is taken from an open-source archive. The images are of 400 × 400 resolution, and the feature size is of dimension 224 × 224. The dataset also consists of 2058 cotton images and 1749 silk images with phone cameras and re-sized all images to 400 × 400. Training dataset consisted of 4331 images and a validation dataset of 153 images belonging to cotton and silk classes. VGG16 network model is used to train the fabric image dataset. VGG16 model consists of 16 layers in the form of five convolutional blocks, and each block has convolution layers of 3 × 3 filter followed by a 2 × 2 filter max pooling layer and last three fully connected layers. Results were observed by using the transfer learning VGG16 convolution neural network model with different optimizers such as stochastic gradient descent (SGD), adaptive moment estimation, Nesterov-accelerated adaptive moment estimation (Nadam), Adamax which is a variation of Adam estimation, and root mean square propagation. Nadam and Adam optimizers are preferred because they provide accurate, quick results, are preferred for sparse data, and decrease noise issues. 94% validation accuracy and 99.60% training accuracy are obtained using all five blocks of the VGG16 model. But better result of 97% validation accuracy is seen by using only four blocks of the VGG16 model. Hence, VGG16 model with four convolutional blocks has been proposed by slightly modifying the last three dense layers by including average pooling layer, flatten layer, and ReLU activation layer. From Table 1, Adam and Nadam show higher accuracy with better results compared to other optimizers. Training the model with Adam optimizer has achieved training accuracy 98.25% and validation accuracy 98.96%. Similarly, with Nadam optimizer there is a slight increase in training accuracy with 99.3% and validation accuracy

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

147

with 98.7%. Optimizers are used to adjust the characteristics of the convolutional neural network, such as weights and learning rate, to reduce the losses. Optimization algorithms or techniques are in charge of lowering losses and delivering the most reliable performance. Table 2 shows the summary of the VGG16 model used, model summary includes complete information on types of layers in the model, output shape for each layer, and parameters or weights of each layer which is updated during training the model using optimization techniques to minimize the error. It is also observed from Table 2 that there are four convolutional blocks. Figure 4 shows that model accuracy is increased with a decrease in the loss. Test or validation accuracy is close to training accuracy. This represents that the model has achieved good accuracy with silk and cotton datasets. Table 1 Model accuracy with Adam, Nadam, and SGD optimizer—classification of cotton and silk Optimizer

Learning rate

Epochs

Training accuracy %

Validation accuracy %

Adam

0.00001

100

98.25

98.96

Nadam

0.00001

100

99.3

98.7

SGD

0.00001

120

69.97

65

Table 2 Model summary Type of layer

Output shape

Parameter

Blocks

conv2d_13 (Conv2D)

(None, 224, 224, 64)

1792

Block1

conv2d_14 (Conv2D)

(None, 224, 224, 64)

36,928

max_pooling2d_5 (MaxPooling2D)

(None, 112, 112, 64)

0

conv2d_15 (Conv2D)

(None, 112, 112, 128)

73,856

conv2d_16 (Conv2D)

(None, 112, 112, 128)

147,584

max_pooling2d_6 (MaxPooling2D)

(None, 56, 56, 128)

0

conv2d_17 (Conv2D)

(None, 56, 56, 256)

295,168

conv2d_18 (Conv2D)

(None, 56, 56, 256)

590,080

conv2d_19 (Conv2D)

(None, 56, 56, 256)

590,080

max_pooling2d_7 (MaxPooling2D)

(None, 28, 28, 256)

0

conv2d_20 (Conv2D)

(None, 28, 28, 512)

1,180,160

conv2d_21 (Conv2D)

(None, 28, 28, 512)

2,359,808

conv2d_22 (Conv2D)

(None, 28, 28, 512)

2,359,808

max_pooling2d_8 (MaxPooling2D)

(None, 14, 14, 512)

0

average_pooing2d_1 (Average)

(None, 7, 7, 512)

0

flatten_1 (Flatten)

(None, 25,088)

0

dense_2 (Dense)

(None, 128)

3,211,392

dense_3 (Dense)

(None, 2)

387

Block2

Block3

Block4

Output layers

148

S. Elavaar Kuzhali et al.

Fig. 4 Plot of model accuracy and loss

Figure 5 shows the confusion matrix for the predictions made by the trained model. 336 cotton images and 248 silk images which are not included in the training dataset are taken, out of which 305 and 225 images are correctly predicted as cotton and silk, respectively, thus it can be stated that the model is predicted with 90% accuracy for non-trained images and it has an error rate of 10%. After the model is trained, it is saved in ‘.h5’ format. In the ‘.h5’ file, it is easy to store structured data; hence, all the weight parameters are stored in this file. This file is used to make predictions on the input image with a confidence rate mentioned. Figure 6 represents the prediction and confidence rate on input images. For estimating the accuracy of the model for non-trained images in the existing system, 91 cotton fabrics and 70 silk fabrics are taken. These fabrics are captured Fig. 5 Confusion matrix

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

149

Fig. 6 Prediction on silk and cotton images

Fig. 7 Predictions on fabric images using Raspberry Pi

by the camera sensor connected to the Raspberry Pi. The results obtained from the prediction are that 83 cotton fabrics are predicted correctly as cotton and 51 silk fabrics are correctly predicted as silk. Thus, it can be stated that the accuracy of non-trained fabric images is 82.025% and the error rate obtained is 17.97%. Figure 7 is the output of predicted images using Raspberry Pi and a camera module displayed on a 16 × 2 liquid crystal display. Based on the results of fabric images, predicted spin rate of the motor, the temperature of the water, and time of washing are set.

4 Conclusion A convolutional neural network model is used for classifying fabrics of cotton and silk images. A VGG16 architecture model with network weights is used for training the model. Four blocks of the VGG16 model with an average pooling layer, a flattening

150

S. Elavaar Kuzhali et al.

layer, and a ReLU activation layer have been used for the proposed model. Adam, SGD, and Nadam optimizers are used to evaluate the model for better performance. 69.97% and 98.25% training accuracy have been achieved using SGD and Adam optimizers, respectively. The Nadam optimizer is used for the final model as it has higher performance and is efficient in improving validation accuracy. 99.3% training accuracy and 98.7% validation accuracy are obtained with the final model using the Nadam optimizer with a learning rate of 0.00001. For non-trained data, the accuracy is 82.025% in the existing system. The proposed model helps in improving the quality of the wash and the longevity of the cloth since the amount of detergent and water level are automatically set based on the load of the fabric and the type of fabric detected. This AI model can also be applied to the textile industries, where automatic classification of various fabrics is required instead of manual labor.

5 Future Work Future work is to ensure that neural network model can classify more than two classes, i.e., nylon, polyester, wool, linen, denim, etc., and to identify multiple images simultaneously. Further experiments can be conducted using different network models for image classification like ResNet50, Inception v3, and EfficientNet since they have an important influence on the results obtained at the end. To extend the result of fabric classification to different applications like laundries and textile industries, this AI model can be applied to the textile industry, where automatic classification of various fabrics is required instead of manual labor.

References 1. Riello G, Roy T (2009) How India clothed the world: the world of South Asian textiles 1500– 1850. Brill 2. Osborne EF (2010) Weaving process for production of a full-fashioned woven stretch garment with load carriage capability. U.S. patent 7,841,369 3. Guarnera GC, Hall P, Chesnais A, Glencross M (2017) Woven fabric model creation from a single image. ACM Trans Graph (TOG) 36(5):1–13 4. Shih C-Y, Kuo C-FJ, Cheng J-H (2016) A study of automated color, shape and texture analysis of Tatami embroidery fabrics. Text Res J 86(17):1791–1802 5. Jing J, Xu M, Li P, Li Q, Liu S (2014) Automatic classification of woven fabric structure based on texture feature and PNN. Fibers Polym 15(5):1092–1098 6. Ouyang W, Xu B, Hou J, Yuan X (2019) Fabric defect detection using activation layer embedded convolutional neural network. IEEE Access 7:70130–70140 7. Hussain I, Ather M, Khan B, Wang Z, Ding S (2020) Woven fabric pattern recognition and classification based on deep convolutional neural networks. Electronics 9(6):1048 8. Garg M, Dhiman G (2020) Deep convolution neural network approach for defect inspection of textured surfaces. J Inst Electron Comput 2(1):28–38 9. Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification with deep convolutional neural networks. Commun ACM 60(6):84–90

Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive …

151

10. Xiang J, Zhang N, Pan R, Gao W (2019) Fabric image retrieval system using hierarchical search based on deep convolutional neural network. IEEE Access 7:35405–35417 11. Lei H, Yi Y, Li Y, Luo G, Wang M (2018) A new clothing image retrieval algorithm based on sketch component segmentation in mobile visual sensors. Int J Distrib Sens Netw 14(11) 12. Joshi KD, Bhavsar SN, Sanghvi RC (2014) Image retrieval system using intuitive descriptors. In: ICIAME 2014. Elsevier, India. Procedia Technol 14:535–542 13. Chae YK (2020) Fabric identifying method, apparatus, and system. U.S. patent application 16/599,960 14. Da Silva Barros MAC, Ohata EF, da Silva SPP, Almeida JS, Filho PPR (2020) An innovative approach of textile fabrics identification from mobile images using computer vision based on deep transfer learning. In: International joint conference on neural networks (IJCNN) 2020. IEEE, Glasgow, UK, pp 1–8 15. Varma M, Zisserman A (2008) A statistical approach to material classification using image patch exemplars. IEEE Trans Pattern Anal Mach Intell 31(11):2032–2047 16. Xie T, Grossman JC (2018) Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys Rev Lett 120(14) 17. Zhang H, Zhang L, Li P, Gu D (2018) Yarn-dyed fabric defect detection with YOLOV2 based on deep convolution neural networks. In: IEEE 7th data driven control and learning systems conference (DDCLS) 2018. IEEE, China, pp 170–174 18. Anami BS, Elemmi MC (2019) ANN approach for classification of different origin fabric images. Int J Image Graph Signal Process 10(12):29–38 19. Makantasis K, Karantzalos K, Doulamis A, Doulamis N (2015) Deep supervised learning for hyperspectral data classification through convolutional neural networks. In: IEEE international geoscience and remote sensing symposium (IGARSS) 2015. IEEE, Italy, pp 4959–4962 20. Li Z, Meng S, Wang L, Zhang N, Gao W (2019) Intelligent recognition of the patterns of yarn-dyed fabric based on LSRT images. J Eng Fibers Fabr 14 21. Benco M, Hudec R, Kamencay P, Zachariasova M, Matuska S (2014) An advanced approach to extraction of colour texture features based on GLCM. Int J Adv Robot Syst 11(7):104 22. Wang X, Wu G, Zhong Y (2018) Fabric identification using convolutional neural network. In: International conference on artificial intelligence on textile and apparel. Springer, Cham, pp 93–100 23. Kampouris C, Zafeiriou S, Ghosh A, Malassiotis S (2016) Fine-grained material classification using micro-geometry and reflectance. In: European conference on computer vision. Springer, Cham, pp 778–792 24. Sparavigna AC (2017) Image segmentation applied to the analysis of fabric textures. Philica. HAL Id: hal-01633061

System Identification, Stability Analysis and PID Controller Design for PEM Electrolyzer Aruna Rajaiah and Jaya Christa Sargunar Thomas

Abstract In a hybrid renewable energy system (RES), different types of energy sources are integrated for meeting the continuous power demand. To overcome the problem of intermittent availability of energy RES, suitable energy storage system is required. The energy storage as hydrogen gas is more efficient and suitable for long-term storage applications. Proton exchange membrane (PEM) electrolyzer is used to store the surplus amount of energy from RES in the form of hydrogen gas. In this paper, initially the mathematical modeling of PEM electrolyzer is developed to know the influence of various parameters on the performance. It is noted that the cell voltage and efficiency of the electrolyzer depend on its temperature. Based on this observation, temperature is taken as input parameter, and the system model is identified for PEM electrolyzer using system identification technique. The different black box approaches are used to identify the suitable system by comparing the response on the basis of percentage of fitness, cost function and final prediction error. It is found that Box–Jenkins method is appropriate for PEM electrolyzer with percentage fitness of 90.5%. To check the system stability, Lyapunov stability analysis is applied and it is found to be stable. For the stable system, conventional PID controller based on Ziegler–Nichols method is used to maintain constant cell voltage by varying the temperature of PEM electrolyzer. For the system with the PID controller, the rise time is 3.8 s, settling time is 75.3 s, peak time is 9.3 s and overshoot is 57.4%. Keywords PEM electrolyzer · System identification · Lyapunov stability analysis · PID controller

A. Rajaiah (B) P.S.R. Engineering College, Sivakasi, Tamil Nadu, India e-mail: [email protected]; [email protected] J. C. S. Thomas Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_10

153

154

A. Rajaiah and J. C. S. Thomas

1 Introduction The utilization of renewable energy sources is increased nowadays, to reduce the usage of fossil fuels. While using renewable energy sources, the main limitation is their irregular energy. There are several alternatives available to overcome this drawback. Among those, saving the excess energy as hydrogen gas using electrolysis method is considered as one of the options. Since hydrogen is an abundantly available chemical element, environment friendly, effective and not dependent on hydrocarbon-based fuels. Various electrolysis techniques are available, among them PEM electrolyzer has the following merits such as compact design, high working current densities, pure gas generation and production at high pressure. Many research works are carried out in the modeling and analysis of PEM electrolyzer. A review on recent development in hydrogen production using PEM electrolyzer is presented in [1]. The challenges related to the components of PEM cell and electro-catalyst are addressed. A case study on integrated system with solar PV panel, hydrogen storage system and a geothermal heat pump is carried in [2]. On analyzing the integrated system, it is found that the system is able to meet the requirement in agriculture field as greenhouse heating. Reference [3] provides a review on modeling of PEM electrolyzer for the beginners including the basic principles of this technology. In [4], the hydrogen generator system based on the PEM electrolysis is developed using MATLAB Simulink tool. The cell and stack behavior under various temperatures and pressures are presented. The photovoltaic panel with electrolyzer and fuel cell is given in [5]. The system is simulated to verify the effectiveness. It provides solution for energy storage problem and is able to meet the power demand. In a standalone renewable hybrid power system, the capacity of PEM electrolyzer for hydrogen production is evaluated in [6]. A comprehensive analysis for PEM electrolyzer is done [7] for improving the performance, reliability and to reduce cost. The literature [8] provides modeling of electrolyzer with the controller for hydrogen production. In [9], a simplified model of PEM electrolyzer is presented with a comparison of experimental data for various temperatures. It is concluded that the electrolyzer cell voltage and efficiency depend on the temperature of the system. For controlling the system, a suitable system model is required. It can be found by using system identification techniques. In literature [10], the black box and white box approaches in system identification technique for cascade tank are applied and compared to choose the most appropriate model. For a heating system, the best model structure is found among the black box approaches based on the estimated and validated test results [11]. A more realistic model structure is identified using black box approach in system identification techniques for PEM fuel cell [12]. Similarly, the model identification and controller design for metal hydride storage tank during desorption process are presented in [13]. In literature [14], the PEM fuel cell is combined with PEM electrolyzer. A sliding mode control is developed to control the fuel cell system and the rate of hydrogen production in the electrolyzer. A sliding mode pulse width modulation controller is

System Identification, Stability Analysis and PID Controller Design …

Modeling of PEM Electrolyzer

System Identification for PEM Electrolyzer

Lyapunov stability analysis

155

PID Controller Design

Fig. 1 Flow diagram presents the flow of work in this paper

designed to ensure the voltage stability and minimize the ripple current in electrolyzer [15]. The stability analysis and design of conventional PID controller for various linear and nonlinear systems are provided in [16, 17]. Based on the literature survey, it is found that there is a research gap in model identification for PEM electrolyzer. Hence, the novelty of this work is system identification for PEM electrolyzer and design of PID controller to maintain a constant output voltage. The flow of this research work is presented (see Fig. 1).

2 Modeling of PEM Electrolyzer In PEM electrolyzer, water molecules dissociate into hydrogen ions and oxygen on applying electrical energy. The PEM electrolyzer is compact in design and produces pure hydrogen gas at high pressure and high current density [3–6]. It has four modules such as anode, cathode, membrane and voltage modules and are interlinked to each other. On applying the input current, an electrochemical reaction takes places where the hydrogen ions developed from the anode and passed to the cathode through the catalyst layer where the H2 gas is produced. Thus, the rate of hydrogen production is depending on the current density. The mathematical model of PEM electrolyzer is taken from [3]. A single cell with 100 cm2 active area is considered and the maximum current density is assumed as 1.8 A/cm2 . The simulation diagram of PEM electrolyzer is presented (see Fig. 2). The electrolyzer works at a temperature range of 343–373 K and the pressure is considered as 1 bar. Increasing the temperature of PEM electrolyzer from 343 to 373 K reduces the cell voltage and increases the efficiency (see Figs. 3 and 4, respectively). The efficiency is calculated based on the thermoneutral voltage, 1.48 V. Thus, on increasing the temperature, the activation loss and ohmic over potential are decreased due to the increase in reaction kinetics and improved ionic conductivity. As on increasing the current density, the rate of power loss on the stack becomes large and the efficiency starts to decrease. By manipulating the operating temperature, the losses can be reduced and thus higher efficiency of PEM electrolyzer can be maintained.

156

A. Rajaiah and J. C. S. Thomas

Fig. 2 Simulation diagram of PEM electrolyzer

Cell Voltage (V)

2 1.8 1.6

Temperature at 343K Temperature at 353K Temperature at 373K

1.4 1.2 1

0.5

0

1 Current density (A/cm2 )

1.5

2

Fig. 3 J–V characteristic curves obtained for PEM electrolyzer

Efficiency (%)

100 80 60 Temperature at 343K

40

Temperature at 353K Temperature at 373 K

20 0

0

2

6 4 Current density (A/cm2 )

Fig. 4 Efficiency curve of PEM electrolyzer

8

10

System Identification, Stability Analysis and PID Controller Design …

157

3 System Identification System identification is a tool to identify the numerical expression of the system with the help of input and output real-time measurement data or experimental data without a knowledge of what is happening inside the system [11]. The process of model identification is to select the modeling technique first, then to choose the structure of model and finally the model is estimated and validated. On the basis of final prediction error, loss function and percentage of fitness, the best system model is selected. The system identification technique has three common approaches such as white box, gray box and black box modeling methods. The white box and gray box modeling methods need knowledge of internal structure of the system along with large amount of data. They are very complex and consuming time is also high. If the internal model structure and first principle equations are unknown, then the black box technique can be used to identify the model structure only by having measured input and output data [12]. In this research work, black box techniques such as Auto-Regressive Exogenous (ARX), Auto-Regressive Moving Average Exogenous Model (ARMAX), Output Error (OE) and Box–Jenkins (BJ) are used to find out the suitable model structure for PEM electrolyzer. In discrete time, the general linear time invariant model [10] can be written as y(t) = G(q)u(t) + H (q)e(t)

(1)

where y(t) is signal output of the model and it corresponds to the input u(t) and the disturbance e(t). Based on Eq. (1), the models are able to depict the dynamical relations based on the linear difference equations. The ARX model is expressed [13] as y(t) + a1 y(t − 1) + · · · + ana (t − n a ) = b1 u(t − 1) + · · · + bn b u(t − n b ) + e(t) (2) where e(t) is direct error (white noise term). The variables A and B are stated in terms of delay operator q, thus the expression is A(q) = 1 + a1 q −1 + · · · + ana q −na

(3)

B(q) = b1 q −1 + · · · + bn b q −n b

(4)

By comparing Eqs. (1)–(4), we get G(q) =

B(q) 1 , H (q) = A(q) A(q)

The ARMAX model structure contains the error equation as a moving average of white noise. The model structure is expressed as

158

A. Rajaiah and J. C. S. Thomas

y(t) + a1 y(t − 1) + · · · + ana (t − n a ) = b1 u(t − 1) + · · · + bn b u(t − n b ) + e(t) + c1 e(t − 1) + · · · + cn c u(t − n c ) (5) where ai , bi and ci are model parameters and e(t) is white noise term. In the Output Error (OE) model, the linear difference equation is correspondence between the input and undisturbed output.   w(t) + f 1 w(t − 1) + · · · + f n f w t − n f = b1 u(t − 1) + · · · + bn b u(t − n b ) (6) where f and b are model parameters. In Box–Jenkins (BJ) model, the model structure corresponds to the disturbance properties and is independent of system dynamics. It is expressed as, y(t) =

C(q) B(q) u(t) + e(t) A(q) D(q)

(7)

where B(q), A(q), C(q) and D(q) are polynomials with respect to the delay q.

3.1 System Identification of PEM Electrolyzer

Temperature (K)

Cell Voltage (V)

From the characteristics and efficiency curve of PEM electrolyzer shown in Figs. 3 and 4, it is found that the efficiency of the system can be maintained by controlling the operational temperature. Hence for model identification, the operational temperature is taken as input variable and the variation of cell voltage is taken as output variable (see Fig. 5). The estimated output of PEM electrolyzer (see Fig. 6). Input and output signals

2

1.5 360

0

200

400

600

800

1000

0

200

600 400 Time (seconds)

800

1000

355 350

Fig. 5 Input and output variables of PEM electrolyzer

System Identification, Stability Analysis and PID Controller Design …

159

Measured and simulated model output

2

Cell Voltage (V)

1.9 1.8 BJ - 90.6% OE - 89.05% ARMAX - 77.62% Cell Voltage

1.7 1.85

1.6

500 1.5

0

200

520

540

400 600 Time (Seconds)

800

1000

Fig. 6 Comparison of simulated output and estimated output by black box approach for PEM electrolyzer

Table 1 Comparison of black box models for PEM electrolyzer Model

Final prediction error 10−5

Cost function 1.756 ×

Fitness (%)

10−5

77.62

ARMAX

1.743 ×

Output error

2.687 × 10−5

3.21 × 10−5

89.05

Box–Jenkins

1.826 × 10−5

2.364 × 10−5

90.6

From Fig. 6, it is found that the best fitness is provided by Box–Jenkins model, followed by Output Error model and then by ARMAX model. The ARX model provides less than 20% of fitness and it is neglected. For the above three models, the cost function and final prediction values are listed in Table 1. Based on the comparison table, it is identified as the Box–Jenkins model is best fitted compared to other models.

4 Lyapunov Stability Analysis Lyapunov stability analysis is the most suitable tool for analyzing nonlinear systems [16]. A nonlinear autonomous system is represented by state equation as x˙ = f (x); f (0) = 0

(8)

The sufficient stability condition for nonlinear systems is as referred in [16]. The Lyapunov function is represented as V (x) = x T P x

(9)

where P is a positive definite Hermitian matrix. The time derivative of V (x) along any trajectory is

160

A. Rajaiah and J. C. S. Thomas ·

V (x) = x˙ T P x + x T P x˙ = (x A)T P x + x T P Ax

(10)

·   V (x) = x T A T P + P A x

(11) ·

Since V (x) is selected as positive definite, for asymptotically stable V (x) be negative definite. Therefore, ·

V (x) = −x T Qx

(12)

  Q = − A T P + P A = positive definite.

(13)

where,

4.1 Stability Analysis for PEM Electrolyzer From the Box–Jenkins model of PEM electrolyzer, the system matrices are obtained as     −0.0564 −0.0002 1 A= B= 1 0 0   C = 5.08e−4 8.45e−7 D = [0] For Lyapunov stability method, the Q matrix is assumed as positive definite matrix 

10 Q= 01



 and P =

P11 P12 P12 P22



The Lyapunov function is V (x) = x T P x From Eq. (13), the positive definite P matrix is obtained as,  44.3e3 2.5e3 >0 P= 2.5e3 149.87 

System Identification, Stability Analysis and PID Controller Design …

161

Hence, P is positive definite. The equilibrium state at the origin is asymptotically stable in large and a Lyapunov function is V (x) = 44.3e3 x12 + 5000x1 x2 + 149.87x22 ·

V (x) = −x22 ·

V (x) is negative definite. Hence, the PEM electrolyzer is asymptotically stable.

5 Design of Controller The conventional PID controllers more commonly used are electronic, hydraulic and pneumatic types depending on the needs of industries [17]. In design of PID controller, there are various possible techniques available for identifying the parameters of the controller. It will help to meet up the specification such as transient and steady-state response of the closed loop system. This process of selecting the controller parameters is known as controller tuning [16]. Ziegler–Nichols tuning method is able to find the parameters such as proportional gain, integral time constant and derivative time constant as K p , T i and T d , respectively. The block diagram of PID controller is represented (see Fig. 7).

5.1 PID Controller for PEM Electrolyzer In PEM electrolyzer, the cell voltage is controlled by manipulating the operational temperature of cell to meet the specified cell voltage which is given as reference voltage 1.8 V. The PID gain parameters identified based on the Ziegler–Nichols tuning method are K p = 67.2, T i = 196.75 and T d = 49.18. The simulation result of PID controller for PEM electrolyzer (see Fig. 8). From Fig. 8, the time domain response is obtained with rise time as 3.8 s, settling time as 75.3 s, peak time as 9.3 s and percentage of overshoot as 57.4%. For improving the response, the controller tuning can be done by optimization techniques. Fig. 7 Block diagram of PID controller for a closed loop system

162

A. Rajaiah and J. C. S. Thomas 3 Reference voltage PID controlled output

Cell voltage (V)

2.5 2 1.5

Rise time : 3.8 seconds Settling time : 75.3 seconds Peak time : 9.3 Overshoot :57.4%

1 0.5 0

0

20

60 40 Time (seconds)

80

100

Fig. 8 PID controller responses for PEM electrolyzer

6 Conclusion This paper presents the model structure, Lyapunov stability analysis and design of PID controller for PEM electrolyzer. The system model is identified by using black box approaches of system identification technique. On comparing the estimated and validated output of ARX, ARMAX, Output Error and Box–Jenkins methods, the Box–Jenkins method has obtained better percentage of fitness as 90.8%, final prediction error as 1.826 × 10−5 and cost function as 2.364 × 10−5 . For the identified Box–Jenkins model, the system stability is analyzed using Lyapunov technique and found that the system is stable. PID controller is designed by using Ziegler–Nichols tuning method. The desired cell voltage of 1.8 V is obtained with a rise time of 3.8 s, settling time of 75.3 s, peak time of 9.3 s and overshoot of 57.4%. Further, to improve the controller response, optimization techniques can be adopted for controller tuning.

References 1. Kumar SS, Himabindu V (2019) Hydrogen production by PEM water electrolysis—a review. Mater Sci Energy Technol 2(3):442–454 2. Pascuzzi S, Anifantis AS, Blanco I, Mugnozza GS (2016) Electrolyzer performance analysis of an integrated hydrogen power system for greenhouse heating. A case study. Sustainability 8(7):629 3. Falcão DS, Pinto AMFR (2020) A review on PEM electrolyzer modelling: guidelines for beginners. J Clean Prod 261:1–38 4. Yigit T, Selamet OF (2016) Mathematical modeling and dynamic Simulink simulation of high-pressure PEM electrolyzer system. Int J Hydrogen Energy 41(32):13901–13914 5. Lajnef T, Abid S, Ammous A (2013) Modeling, control, and simulation of a solar hydrogen/fuel cell hybrid energy system for grid connected applications. Adv Power Electron 1–9 6. Dursun E, Acarkan B, Kilic O (2012) Modeling of hydrogen production with a stand-alone renewable hybrid power system. Int J Hydrogen Energy 37(4):3098–3107

System Identification, Stability Analysis and PID Controller Design …

163

7. Ma Z, Witteman L, Wrubel JA, Bender G (2021) A comprehensive modeling method for proton exchange membrane electrolyzer development. Int J Hydrogen Energy 46(34):17627–17643 8. Zhou T, Francois B (2009) Modeling and control design of hydrogen production process for an active hydrogen/wind hybrid power system. Int J Hydrogen Energy 34(1):21–30 9. Liso V, Savoia G, Araya SS, Cinti G, Kær SK (2018) Modelling and experimental analysis of a polymer electrolyte membrane water electrolysis cell at different operating temperatures. Energies 11(12):1–19 10. Giordano G, Sjöberg J (2018) Black and white-box approaches for cascaded tanks benchmark system identification. Mech Syst Signal Process 108:387–397 11. Rabbani MJ, Hussain K, Khan AUR, Ali A (2013) Model identification and validation for a heating system using MATLAB system identification toolbox. IOP Conf Ser Mater Sci Eng 51(1):1–10 12. Chavan SL, Talange DB (2018) System identification black box approach for modeling performance of PEM fuel cell. J Energy Storage 18:327–332 13. Aruna R, Jaya Christa ST (2020) Modeling, system identification and design of fuzzy PID controller for discharge dynamics of metal hydride hydrogen storage bed. Int J Hydrogen Energy 45(7):4703–4719 14. Sankar K, Jana AK (2021) Nonlinear control of a PEM fuel cell integrated system with water electrolyzer. Chem Eng Res Des 171:150–167 15. Koundi M, El Fadil H (2019) Mathematical modeling of PEM electrolyzer and design of a voltage controller by the SMPWM approach. In: 5th international conference on power generation systems and renewable energy technologies. IEEE, Turkey, pp 1–6 16. Ogata K (2002) Modern control engineering, 5th edn. Prentice-Hall, New Jersey 17. Stephanopoulos G (1984) Chemical process control: an introduction to theory and practice, 1st edn. Prentice-Hall, New Jersey

Sliding Mode Hybrid Control of PMSM for Electric Vehicle Ajay Pawar and S. V. Jadhav

Abstract Permanent magnet synchronous motor (PMSM) is becoming a favorable motor for electric vehicle. This is because of its well-known features like better torque to weight ratio, larger power factor, high efficiency, and reliability as against its counterpart induction motor (IM). The paradigm of control of drives has been shifted because of vector control (VC) and field-oriented control (FOC). Speed control using proportional-and-integral (PI) controllers and many nonlinear controllers is popular in the heavy performance application of PMSM. Applications like robotics, CNC, and electric vehicle demand for dynamic control over a large operation range including four-quadrant operations. This work envisages control of PMSM drive for electric vehicles. Sliding mode control (SMC) algorithm is used in order to achieve higher energy efficiency, in constant power and constant torque regions. Further, higher order sliding mode control (HOSMC) algorithm is also tested for robustness against disturbances and parameter variations, nonlinearity load patterns, and, importantly, EV. HOSMC not only reduces the ripples caused by SMC but also offer dynamic speed and torque response with optimized control input. Many simulations are carried on MATLAB Simulink platform to prove the effectiveness of SMC and HOSMC. Keywords Permanent magnet synchronous motor (PMSM) · Field-oriented control (FOC) · Field weakening control (FWC) · Sliding mode control (SMC) · Higher order sliding mode control (HOSMC) · Integral time absolute error (ITAE) · Four-quadrant operation · Simulation and result

A. Pawar (B) · S. V. Jadhav Department of Electrical Engineering, College of Engineering, Pune, India e-mail: [email protected] S. V. Jadhav e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_11

165

166

A. Pawar and S. V. Jadhav

1 Introduction Advances in the control of electric drives have led to the replacement of the IC engines by gear-less drive systems where electric motors drive the vehicles. Robust and dynamic control in constant torque and constant power regions replicate the lower and higher gears of a mechanical engine. Advanced control algorithm like sliding mode control (SMC) or intelligent control has made the control operation fast and accurate for a wide operating speed range [1]. PMSM is popularly used in electric vehicles. On account of reduce copper losses and use of high energy permanent magnets, PMSM possesses advantages like high efficiency, high torque to weight ratio, etc. [2]. In past two decades, vector control (VC) and field-oriented control (FOC) have dominated the drive applications [3, 4] due to the well-known feature of decoupled speed and torque control. Thanks to the developments in the high frequency power switches and DSP processors; because of which modern complex control algorithms can be applied in real time. Various intelligent controllers and variable structure control (VSC) algorithms are now preferred over simple PI control algorithms [5]. Several algorithms like SMC, fuzzy control, artificial neural network (ANN)-based control, and fuzzy-SMC control are comparable and offer satisfactory drive performance [6]. The researchers have explored combination of SMC and PI control to achieve required performance in steady-state and transient conditions [7]. In order to reduce the ripples in the response, higher order SMC (HOSMC) [8] is proposed and is applied in drives applications [9–11]. Speed control of PMSM is generally has cascaded control loop structure. It consists of an outer speed control loop and two inner current loops. This gives good disturbance rejection and better steady-state performance. Torque and speed decoupling is achieved by two inner current loops out of which, magnetizing current is regulated to zero to obtain fieldoriented control (FOC) [12] improving the torque per ampere ratio. The EV is operated in constant torque region for normal speeds. High speed operation is obtained using field weakening control (FWC) in constant power mode [13]. Thus, wide speed range along with four-quadrant operation is obtained for EV. In this paper, SMC and HOSMC, both algorithms are attempted separately and the performances are compared. For better settling time and less steady-state error, super twisting algorithm (STA) is employed [14]. For the EV application purpose TATA nano car, which is currently IC engine based, is considered because of its low weight, good volume space, and low price. It is tested for its conversion to EV. The paper is arranged as follows: Modeling of PMSM is illustrated in Sect. 2. FOC and FWC as applied to PMSM is explained in Sect. 3. Design of cascade SMC hybrid control is given in Sect. 4. Simulation and results are shown in Sect. 5, and Sect. 6 concludes the paper.

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

167

2 PMSM Model 2.1 Modeling of PMSM Three-phase voltage model of stator of PMSM is given by following equations [15]. Va (t) = i a R + L s

di a + E a (t) dt

(1)

Vb (t) = i b R + L s

di b + E b (t) dt

(2)

Vc (t) = i c R + L s

di c + E c (t) dt

(3)

2.2 d-q Model of PMSM The model described by Eqs. (1), (2), (3) is converted into time invariant model by applying Clarke and Parke transformations. This results in a d-q-axis model in the synchronously rotating stator flux reference frame. di ds − ωe i qs L dt

(4)

di qs + ωe i ds L + ωe Φ dt

(5)

Vds = Ri ds + L Vqs = Ri qs + L Rearranging the equations

 1 di ds = Vds − Ri ds + ωe Li qs dt L

(6)

 di qs 1 = Vq − Ri qs + ωe Li ds − ωe Φ dt L

(7)

The expression of electromagnetic torque in terms of inductance, direct axis, and quadrature axis current is Te =

   3  P ϕi qs + L d − L q i ds i qs 2

(8)

In surface mounted PMSM, L d = L q and hence electromagnetic torque becomes

168

A. Pawar and S. V. Jadhav

Te =

 3  P ϕi qs 2

(9)

The expression for electromagnetic torque in terms of mechanical torque is Te = TL + Bωm + J Here, ωm =

dωm dt

(10)

P ω. 2 e

3 Field-Oriented and Flux Weakening Control From Eq. (9), electromagnetic torque is represented as Te =

 3  P ϕi qs 2

(11)

1 dωe = Te dt J

(12)

FOC gives better result below base speed when we require speed above the base speed. Field weakening control is come into the picture. In FOC, we make ids current equal to zero but for FWC we make ids current negative so that we can achieve speed more than base speed similar to DC motor. For that purpose, we select i dref as i dref

1 ∅ =− + L L

/

Udc w

2

  − Li qs

(13)

4 Design of Cascade SMC Hybrid Control This is the cascade sliding mode hybrid control with current control as inner loop and voltage control as an outer loop, with current we can control torque and with voltage we can control speed. A switch is used for hybrid control as shown and gives efficient control scheme (FOC and FWC) depending on the reference speed (Fig. 1). As there are two separate current control loops (d-axis and q-axis currents), and one outer speed loop, and three SMC controllers are designed. The surfaces are designed on the basis of error in current and speed as below   ∗ Sd = C T × ed = C T i ds − i ds =0

(14)

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

169

Fig. 1 Block diagram of SMC hybrid control



∗ =0 Sq = C T × eq = C T i qs − i qs

(15)

  Sw = C T × ew = C T ωe − ωe∗ = 0

(16)

4.1 Direct Axis Sliding Mode Control We have −R di ds P 1 = i ds + ωe i qs + u d dt L 2 L where u d = flux control input. ∗ Now sliding surface, sd = i ds − i ds . By taking first derivative of sd , we get s˙d =

di ∗ di ds − ds dt dt

p 1 di ∗ R S˙d = − i ds + ωe i qs + u d − ds L 2 L dt

(17)

The presence of control input in S˙d equation ensures that we choose right sliding surface. Here, u d = u d1 + u d2 where u d1 = equivalent control u d2 = switching control.

170

A. Pawar and S. V. Jadhav

So, u d1 is chosen as u d1 = Ri ds + L

∗ p di ds − ωe Li qs dt 2

(18)

Now, u d2 can be defined with upper bound of uncertainties, i.e., u d2

| | | | di ∗ds p | ≥ |Δ Ri ds + Δ L − ωe Δ Li qs || dt 2

(19)

switching component with Lyapunov stability approach is u d2 = −k sgn(sd )

(20)

Then, total flux control input becomes u d = Ri ds +

p di ∗ ωe Li qs + L ds + −k sgn(sd ) 2 dt

(21)

4.2 Quadrature Axis Sliding Mode Control di qs −R 1 p p φ = i qs − ωe Li ds − ωe + u q dt L 2 2 L L where V q = uq = control input. Sliding surface Sq = i qs − i qs ∗ Taking derivative di ∗qs di qs − dt dt di ∗qs 1 p p φ −R i qs − ωe Li ds − ωe + u q − = L 2 2 L L dt

sq· =

(22)

Then, u q1 is chosen as ∗ di qs p p ωe Li ds + ωe φ + L 2 2 dt   u q2 = −k sgn sq

u q1 = Ri qs +

(23) (24)

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

171

4.3 Speed Sliding Mode Control Similarly, here, control input is i qs1



1 dωe∗ τm + Bωe + I = k dt

(25)

where k = 43 Pφ i qs2 = −k sgn(sw )

(26)

4.4 Super Twisting Algorithm (STA) Chattering phenomenon can be reduced by modifying the switching control with the help of super twisting algorithm (STA) we choose uq as  u s = k1 |S|1/2 sgn(S) − k2

sgn(S)dt

(27)

Replace control input according to selected sliding surface.

5 Simulations and Result The simulations are carried out on MATLAB Simulink platform. The rating of PMSM are 1.3 kW, 300 V, and 4.5 A. The parameters for simulation of PMSM motor is as shown in Table 1. Figure 2 shows transient no load speed response with PI, SMC, and HOSMC (hybrid) controllers. Also, Fig. 3 shows steady-state responses. Settling time for PI Table 1 Parameters of PMSM

Parameter

Value

R

1.3 Ω

Ld = Lq

0.0069 H

φ

0.157 Wb

Pole pair

6

Inertia

0.0027 kg m2

B

0.0001 Nm S

172

A. Pawar and S. V. Jadhav

controller is 16 ms and that for SMC and HOSMC is 10 and 6 ms. No load steady-state errors are 0.4 rpm, 0.002 rpm, and 0.0001 rpm, respectively. Speed response for load of 5 Nm is applied at 0.025 s, and transients are shown in Fig. 4. Figure 4 shows the speed response, and Fig. 5 shows stator current with load (5 Nm) at 0.025 s. Speed response with change in resistance by 150% is shown in Fig. 6, and from the figure, it is observed that SMC and HOSMC transient behavior is unaffected by resistance change, but in case of PI control, settling time is increased from 15 to 19 ms also when load is applied, then transient behavior is changed in case of PI. Speed response when disturbance of 0.25 sin(1000t) is inserted in the system and Fig. 7 shows that HOSMC and SMC are unaffected with disturbance but there is change in speed response with PI controller.

Fig. 2 No load speed transients with PI, SMC, and HOSMC

Fig. 3 Speed at no load (steady-state error)

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

Fig. 4 Speed response for step load (5 Nm) at 0.025 s

Fig. 5 Stator current for condition in Fig. 4

Fig. 6 Speed with 150% change in resistance and load (5 Nm) at 0.025 s

173

174

A. Pawar and S. V. Jadhav

Fig. 7 Speed response with disturbance (0.25 sin(1000t))

Figure 8 shows speed response when nonlinear load applied, and Fig. 9 shows the nonlinear load torque. ITAE error with different values of inertia shown in Fig. 10. Four-quadrant operation of PMSM is shown in Figs. 11 and 12. Quadrant

Mode of operation

Time

Speed

Torque

First quadrant

Forward motoring

0–0.025

525

+5

Fourth quadrant

Forward regeneration

0.025–0.05

525

−5

Third quadrant

Reverse motoring

0.05–0.075

−525

−5

Second quadrant

Reverse regeneration

0.075–0.1

−525

+5

Constant torque and constant power mode of operation with PI, SMC, and HOSMC are compared as in Figs. 13, 14, 15 and 16.

Fig. 8 Speed response when nonlinear load applied

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

175

Fig. 9 Nonlinear load torque with HOSMC

Fig. 10 ITAE for the normal operating conditions with step load applied at 0.025 s with J = 0.027 kg m2

5.1 Application on Electrical Vehicle For an electric vehicle application, we have considered TATA nano vehicle because of its low weight, less price, and sufficient volume space. As it has low weight requires low power to drive the vehicle. For this purpose, we have considered 300 Vdc, 4500 rpm surface mounted PMSM motor (Tables 2, 3 and 4). Figure 17 shows the drive cycle source (FTP75) for 50 s with HOSMC (Figs. 18, 19; Tables 5, 6, 7).

176

A. Pawar and S. V. Jadhav

Fig. 11 Speed in all quadrant

Fig. 12 Torque in all quadrant with HOSMC

6 Conclusion This paper deals with the cascade SMC hybrid control for PMSM. Four-quadrant operation is achieved with this control strategy, and results are satisfactory. SMC deals with parameter uncertainty (e.g., change in stator resistance) and external disturbance. Smooth ripple-free operation of drive (TATA nano car) is a result of HOSMC. It is observed that HOSMC offers better performance with respect to SMC and PI which is proved by the values of ITAE, peak overshoot, and settling time. HOSMC is the most suitable algorithm for EV applications.

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

Fig. 13 Speed in constant torque and constant power mode

Fig. 14 Direct axis current with PI

177

178

A. Pawar and S. V. Jadhav

Fig. 15 Direct axis current with SMC

Fig. 16 Direct axis current with HOSMC (hybrid) Table 2 PMSM parameters

Motor power (P) = 3 kW Stator resistance (Rs) = 0.62 Stator inductance (Ls = Ld = Lq) = 0.002075 Rotor flux linkage (Øm) = 0.08627 Pole pair (P) = 4 Viscous friction coefficient (B) = 0.00009444 Wheel inertia (J) = 0.1

Table 3 TATA nano specifications

Mass of TATA nano = 630 kg Wheel radius of TATA nano = 0.3 m Area of TATA nano = 2.5 m2

Sliding Mode Hybrid Control of PMSM for Electric Vehicle Table 4 Gear ratios

Gear ratio (R1) = 3.45 Gear ratio (R2) = 1.95 Gear ratio (R3) = 1.26 Gear ratio (R4) = 0.838

Fig. 17 Drive cycle source

Fig. 18 Stator current

179

180

A. Pawar and S. V. Jadhav

Fig. 19 Stator current (steady-state error) Table 5 Comparison of PI, SMC, and HOSMC is carried out, on the basis of different performance parameters Property

Controller PI

SMC

Hybrid-HOSMC

J= 0.027 kg m2

J= 0.03 kg m2

J= 0.027 kg m2

J= 0.03 kg m2

J= 0.027 kg m2

J= 0.03 kg m2

0.011

0.012

0.0025

0.0028

0.0022

0.0023

Normal 0.0115 operating conditions with step load applied at 0.5 s

0.013

0.0027

0.003

0.0024

0.0024

Change in stator resistance

0.0134

0.0027

0.003

0.0024

0.0024

Normal operating condition

0.0117

Table 6 THD analysis Controller (J = 0.027 kg m2 )

PI

SMC

Hybrid-HOSMC

Speed overshoot (%)

0.1

0.05

0.0001

Steady-state error in speed

0.09

0.002

0.0001

THD

0.776

0.7751

0.7747

Sliding Mode Hybrid Control of PMSM for Electric Vehicle

181

Table 7 Comparative analysis Property

Controller PI

SMC

Hybrid-HSMC

Starting transient performance

Good

Very good

Excellent

Robustness

Good

Very good

Excellent

Steady-state performance

Good

Very good

Excellent

Computational effort

Medium

Low

High

References 1. Maini C, Gopal K, Prakash R (2013) Making of an ‘all reason’ electric vehicle. In: 2013 world electric vehicle symposium and exhibition (EVS27), pp 1–4. https://doi.org/10.1109/EVS.2013. 6915015 2. Loganayaki A, Kumar RB (2019) Permanent magnet synchronous motor for electric vehicle applications. In: 2019 5th international conference on advanced computing & communication systems (ICACCS), pp 1064–1069. https://doi.org/10.1109/ICACCS.2019.8728442 3. Mishra A, Makwana JA, Agarwal P, Srivastava SP (2012) Modeling and implementation of vector control for PM synchronous motor drive. In: IEEE-international conference on advances in engineering, science and management (ICAESM-2012), pp 582–585 4. Pindoriya RM, Mishra AK, Rajpurohit BS, Kumar R (2016) Performance analysis of control strategies of permanent magnet synchronous motor. In: 2016 IEEE region 10 conference (TENCON), pp 3224–3227. https://doi.org/10.1109/TENCON.2016.7848645 5. Jadhav SV, Srikanth J, Chaudhari BN (2010) Intelligent controllers applied to SVM-DTC based induction motor drives: a comparative study. In: 2010 joint international conference on power electronics, drives and energy systems & 2010 power India, pp 1–8. https://doi.org/10.1109/ PEDES.2010.5712400 6. Jadhav SV, Kirankumar J, Chaudhari BN (2012) ANN based intelligent control of induction motor drive with space vector modulated DTC. In: 2012 IEEE international conference on power electronics, drives and energy systems (PEDES), pp 1–6. https://doi.org/10.1109/PEDES.2012. 6484350 7. Levant A (2003) Higher-order sliding modes, differentiation and output-feedback control. Int J Control 76(9–10):924–941. https://doi.org/10.1080/0020717031000099029 8. Laghrouche S, Harmouche M, Ahmed FS, Chitour Y (2015) Control of PEMFC air-feed system using Lyapunov-based robust and adaptive higher order sliding mode control. IEEE Trans Control Syst Technol 23(4):1594–1601. https://doi.org/10.1109/TCST.2014.2371826 9. Feng Y, Zheng J, Yu X, Truong NV. Hybrid terminal sliding-mode observer design method for a permanent-magnet synchronous motor control system. IEEE Trans Ind Electron

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor M. Bala Abhirami and I. Thirunavukkarasu

Abstract In this article, a PID tuning rule is presented for a lab-scale batch reactor model in order to minimize the impact of load disturbance on the process. The tuning technique is solely determined from the maximum sensitivity (M S ) criteria, which are dependent on the full-state feedback and also on the matching of frequency response. The only tuning parameter for the controller is the maximum sensitivity (M S ) which indicates the level of robustness. Hence, by adjusting the value of (M S ), the values of PID tuning coefficients can be determined precisely. The suggested method is suitable for integrating processes with dead time and also for other complex systems by narrowing it down to FOPDT process. This article provides a controller for a lab-equipped batch reactor using the proposed method for tuning the controller parameters. Keywords Maximum sensitivity (M S ) · PID controller · Load disturbance rejection

1 Introduction Most of the industrial control system applications are based on conventional PID controllers owing to its simplicity, cost-effectiveness, and easy maintenance. It must be noted that about 90% of the industrial applications still choose PID controllers over other modern control methods. Various approaches are employed to tune of the parameters of a PID controller [1–5], some of them which include Ziegler–Nichols method, internal mode control (IMC), pole placement method, etc. The most popular method for tuning PID control parameter is the Ziegler–Nichols method. This method is used widely in most of the industries due to its simplicity. Meshram et al. [6] designed a controller to control the speed of a DC motor. They used M. Bala Abhirami (B) · I. Thirunavukkarasu Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India e-mail: [email protected] I. Thirunavukkarasu e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_12

183

184

M. Bala Abhirami and I. Thirunavukkarasu

Ziegler–Nichols methods to tune the parameters of the PID controller. Li et al. [7] used Ziegler–Nichols method to reduce and validate PID parameters using MATLAB for a model as an example. Ziegler–Nichols method is known for its simplicity, but the tuning the parameters might be difficult in the presence of noise. In case of complex systems, the process output might give oscillatory response with large overshoots. Integral model control (IMC) design of PID is majorly known for achieving robustness in the system design [8]. It is also employed for the tuning of model with delay. It is also used for stable and integrating process models [9, 10]. In spite of their advantages, IMC control design has one major setback. It sets the integral time of the controller equal to that of the process time constant. Due to this, the controller might have very long integral time which leads to slower recovery from the load disturbances. Pole placement method is used to design linear control systems. This method makes use of the dominant poles of the system, and the closed loop poles are placed on the desired location such that all the poles remain to the far left of the poles assigned. Primary setback of this method is that it does not place the zeros at desired locations, and this method requires more skill and knowledge to find the optimum locations to place the poles. Examples of this method of controller design are presented in [11, 12]. In this article, a simple PID controller is developed based upon the maximum sensitivity (M S ) in order to improve the effectiveness of load disturbance rejection [13]. The controller is to be designed for a lab-equipped batch reactor where the system model is approximated to FOPDT model. Simulations have been carried out for the FOPDT model using the proposed tuning method. The article is structured accordingly: Sect. 2 demonstrates the design method of the proposed method. Section 3 presents a controller design for a lab-scale batch reactor on the basis of the proposed method to tune the controller gains, and Sect. 4 summarizes the paper with final remarks.

2 Design Method The expression for FOPDT process can be written as: G(s) =

1.4493 −0.1093s e 5.832s + 1

(1)

where K, T, and L represent the steady-state gain, time constant, and time delay of the FOPDT system. The expression for a PID controller is can be written as: G(s) = K C +

  1 KI + KDs = KC 1 + + Td s s Ti s

(2)

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor

185

Table 1 M S -based PID tuning rule for a general FOPDT process [13] Constants

Controller parameter

ac = −0.3751MS2 + 1.875MS − 1.439 bc = −0.01233MS − 0.9926

K K C = ac (L/T )bc + cc

cc = −0.1847MS2 + 0.9121MS − 0.7085 ai = −0.72721MS−3.083 + 1.587 Ti /T = ai ebi (L/T ) + ci edi (L/T )

bi = −0.5518MS−7.687 + 0.1663 ci = −0.01998MS3.535 − 1.073 di = −51.19MS−8.921 − 1.581

ad = Td /T = ad

ebd (L/T )

+ cd

edd (L/T )

−0.0638MS2 + 0.659MS − 0.6299 MS − 1.143

bd = −1.359MS−7.717 + 0.1459 −0.4453MS + 0.4476 MS − 1.129 −4.573 dd = 2 MS − 3.7MS + 10.41 cd =

where K C , K I , and K D are the proportional, integral, and derivative gains of the controller. K I and K D are defined as K C /T i and K C T d where T i and T d denote the integral time and derivative time constants, respectively. Maximum Sensitivity M S is given by,     1   MS = max  0≤ω≤∞ 1 + G( jω)C( jω) 

(3)

The robustness of the system is measured by the value of M S and smaller value of M S indicates larger stability margin of the control system. The value of M S fixed in the range of 1.2–2.0 [13]. The PID control parameters can be determined directly by choosing the most appropriate value for M S and by calculating the constants KK C , T i /T, T d /T using the equations provided in Table 1.

3 Controller Design for Lab-Scale Batch Reactor Batch reactor is highly nonlinear system with temperature profile as a set point instead of constant set point. Batch reactors are mostly used in chemical process industries having wide application in paint and pharma industries.

186

M. Bala Abhirami and I. Thirunavukkarasu

The objective is to maintain the reactor temperature in-line with the temp. profile. The challenge is the reagents A and B reacts in exothermic way results with constant rise in reactor temperature. The reactor temperature will be measured by the RTD. The reactor will have coolant jacket with continues coolant flow with control valve. The outlet of jacket coolant water will also be measured to control the valve action. Another RTD is used to measure the coolant jacket outlet temperature. The batch reactor setup is displayed in Figs. 1 and 2 shows the schematic diagram of the process. The transfer function of a batch reactor setup is expressed by FOPDT model as: G(s) =

1.4493 −0.1093s e 5.832s + 1

(4)

where is the steady-state gain K = 1.4493, time constant T = 5.832, time delay L = 0.1093. The controller parameters are obtained for several other FOPDT processes with different L/T ratios. Figure 3 shows the controller gains of the batch reactor along with several other FOPDT processes that are considered. The scaled controller gains KK C are plotted against the system parameter L/T. The PID parameters are computed for different M S values as shown in the figure. Table 2 shows the different values of M S that are considered to calculate the constants ac , bc , cc . Figure 4 shows the constants ac , bc , cc plotted against the M S values. Figure 5 shows the scaled integral time constants T i /T plotted against the system parameter L/T for different values of M S . Table 3 gives the values of ai , bi , ci , d i for different M S values. Figure 6 shows the constants ai , bi , ci , d i plotted against different M S values. Figure 7 shows the scaled derivative time constants T d /T plotted against the system parameter L/T for different values of M S . The values of ad , bd , cd , d d for the different Fig. 1 Batch reactor setup

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor

187

Fig. 2 Schematic diagram of a batch reactor

20

Ms=1.2 Ms=1.3 Ms=1.4 Ms=1.5 Ms=1.6 Ms=1.7 Ms=1.8 Ms=1.9 Ms=2.0

18 16

Gain KKc

14 12 10 8 6 4 2 0.1

0.15

0.2

0.25

0.3

0.35

L/T

Fig. 3 Plots of KK C versus system parameter L/T

M S values that are considered are given in Table 4. The constant values of ad , bd , cd , d d , plotted against M S values are shown in Fig. 8. As seen from the plots, the parameters KK C , T i /T, and T d /T are dependent on M S . Hence, by assigning a suitable value for M S , the PID control parameters are calculated as per the equations provided in Table 1. The PID values for M S = 1.2 are calculated from above-mentioned parameters as K p = 9.90, K i = 01.633, and K d = −0.00639, whereas the PID values for lambda method are K p = 0.383, K i = 0.285, and K d = 0.87.

188

M. Bala Abhirami and I. Thirunavukkarasu

Table 2 Constants values for the curves in Fig. 5

MS

ac

bc

cc

1.2

0.3418

−1.0074

0.1201

1.3

0.4489

−1.0086

0.1651

1.4

0.5497

−1.0099

0.2064

1.5

0.6441

−1.0111

0.2441

1.6

0.7322

−1.0123

0.2780

1.7

0.8139

−1.0136

0.3083

1.8

0.8893

−1.0148

0.3349

1.9

0.9583

−1.0160

0.3577

2.0

1.0210

−1.0173

0.3769

1.5

1

ac,bc,cc

0.5 ac vs Ms bc vs Ms cc vs Ms

0

-0.5 -1

-1.5 1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2

Ms

Fig. 4 Constants ac , bc , cc versus maximum sensitivity M S

4 Simulation and Results The open-loop response of the lab-scale batch reactor is provided in Fig. 9. The controller for the batch reactor is designed using the suggested method, and the results are compared with the PID controller implemented using the lambda tuning approach as proposed in [14]. The most suitable value of M S is 1.2 for the suggested controller design. Figures 10 and 11 show the real-time implementation of M S -based PID and lambda PID on a batch reactor setup for trajectory tracking and respective control signal. It is observed with oscillation on the reactor temperature as well on the manipulated variable. In general, PID control performance will not be good for trajectory tracking due to the absence of hard constraints engagement. This can be overcome

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor

189

1

0.9

Ti/T

0.8 Ms=1.2 Ms=1.3 Ms=1.4 Ms=1.5 Ms=1.6 Ms=1.7 Ms=1.8 Ms=1.9 Ms=2.0

0.7

0.6

0.5 0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

L/T

Fig. 5 Plots of T t /T versus system parameter L/T

Table 3 Constants values for the curves in Fig. 6 MS

ai

bi

ci

di

1.2

1.1726

0.0304

−1.111

−11.646

1.3

1.2630

0.0929

−1.1235

−6.5093

1.4

1.3294

0.1248

−1.1386

−4.1254

1.5

1.3787

0.1419

−1.1568

−2.9559

1.6

1.4163

0.1514

−1.1782

−2.3540

1.7

1.4498

0.1570

−1.2034

−2.0311

1.8

1.4683

0.1603

−1.2326

−1.8513

1.9

1.4865

0.1623

−1.2662

−1.7479

2.0

1.5012

0.1636

−1.3046

−1.6866

with prediction-based controller, especially with MPC and NMPC categories with nonlinear Kalman filters.

5 Conclusion This article provides a simple method of PID controller design using maximum sensitivity (M S ) which relies on the full-state feedback and frequency response model matching. This technique is employed on FOPDT process models but is also suitable for other complex models after converting it to FOPDT model. A controller design is implemented on a lab-equipped batch reactor using MATLAB2021® platform.

190

M. Bala Abhirami and I. Thirunavukkarasu 2

0

ai,bi,ci,di

-2 ai vs Ms bi vs Ms ci vs Ms di vs Ms

-4

-6

-8

-10

-12 1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2

Ms

Fig. 6 Constants ai , bi , ci , d i versus maximum sensitivity M S

0.26

0.24

Td/T

0.22 Ms=1.2 Ms=1.3 Ms=1.4 Ms=1.5 Ms=1.6 Ms=1.7 Ms=1.8 Ms=1.9 Ms=2.0

0.2

0.18

0.16

0.45

0.5

0.55

0.6

0.65

L/T

Fig. 7 Plots of T t /T versus system parameter L/T

The output response is compared with the controller designed using lambda tuning approach. The real-time results show the tight tracking with M S -based PID than lambda PID method. Oscillations can be reduced or nullified with the predictivebased nonlinear controllers.

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor

191

Table 4 Constants values for the curves in Fig. 7 MS

ad

bd

cd

dd

1.2

1.2110

−0.1869

−1.222

−0.617

1.3

0.7578

−0.0335

−0.7678

−0.627

1.4

0.6523

0.0446

−0.6488

−0.636

1.5

0.6024

0.0864

−0.5939

−0.643

1.6

0.5715

0.1098

−0.5624

−0.648

1.7

0.5494

0.1233

−0.5419

−0.652

1.8

0.5321

0.1313

−0.5275

−0.654

1.9

0.5577

0.1363

−0.5168

−0.654

2.0

0.5051

0.1394

−0.5086

−0.652

1.5 ad vs Ms bd vs Ms cd vs Ms dd vs Ms

1

ad,bd,cd,dd

0.5

0

-0.5

-1

-1.5 1.2

1.3

1.4

1.5

1.6

1.7

Ms

Fig. 8 Constants ad , bd , cd , d d versus maximum sensitivity M S

1.8

1.9

2

192

M. Bala Abhirami and I. Thirunavukkarasu Time Series Plot:

1.5

Process Output

1

0.5

0 0

10

20

30

40

50

60

70

80

90

100

Time (seconds)

Fig. 9 Open loop response of temperature control in a batch reactor

Fig. 10 Implementation of M S -based PID (pink) and lambda PID (blue) controller on a pilot plant batch reactor for trajectory tracking

Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor

193

Fig. 11 Manipulated variable (coolant control valve) for M S -based PID control

References 1. Hagglund T (2019) The one-third rule for PI controller tuning. Comput Chem Eng J 2. Astrom KJ, Hagglund T (2006) Advanced PID control. ISA-The Instrumentation, Systems and Automation Society, Research Triangle Park, NC 3. Cvejn J (2011) Simple PI/PID controller tuning rules for FOPDT plants with guaranteed closedloop stability margin. Acta Montan Slov 16(1):17–25 4. Hayashi K, Yamamoto T (2011) Design of a PID tuner based on generalized-output errors. In: SICE annual conference, Tokyo, pp 2614–2618 5. Kolaj W, Mozaryn J, Syfert M (2016) PLC-PIDTuner: application for PID tuning with SIMATIC S7 PLC controllers. In: 21st international conference on methods and models in automation and robotics (MMAR), Miedzyzdroje, pp 306–311 6. Astrom KJ, Hagglund T (1995) PID controllers theory design and tuning. Instrument Society of America, Research Triangle Park, NC 7. Seborge DE, Edgar TF, Mellichamp DA (1989) Process dynamics and control. Wiley, New York 8. Vanavil B, Krishna Chaitanya K, Sheshagiri Rao A (2013) Improved PID controller design for unstable time delay processes based on direct synthesis method and maximum sensitivity. Int J Syst Sci 1–18 9. Yang X, Xu B, Chiu MS (2011) PID controller design directly from plant data. Ind Eng Chem Res 50:1352–1359 10. Shamsuzzoha M (2013) Closed loop PI and PID controller tuning for stable and integrating process with time delay. Ind Eng Chem Res 52:12973–12992 11. Meshram PM, Kanojiya RG (2012) Tuning of PID controllers using Ziegler-Nichols method for speed control of DC motor. In: IEEE-international conference on advances in engineering, science and management (ICAESM), Nagapattinam, India 12. Li S, Jiang Q (2011) Study on PID parameters tuning based on Matlab/Simulink. In: IEEE 3rd international conference on communication software and networks, Xi’an, China

194

M. Bala Abhirami and I. Thirunavukkarasu

13. Raza A, Pathak N, Anwar MN (2017) A PID controller tuning rule for FOPDT process to achieve better load disturbance rejection based on maximum sensitivity. In: ICSPACE: international conference on smart grids, power and advanced control engineering, pp 149–154 14. Pruna E, Sasig ER, Mullo S (2017) PI and PID controller tuning tool based on the lambda method. IEEE

Performance Prediction of Solar Cell Using Virtual Production Simulation B. Ashok Kumar, T. S. Bagavat Perumaal, S. Senthilrani, and Parthasarathy Seshadri

Abstract Recent advancements in renewable energy harvesting system have triggered the photovoltaic research in different dimensions. It has enabled rapid development of systems for renewable energy production. To yield an effective production of solar energy, it has become mandatory to predict the solar cell performance virtually. Such prediction will lead to performance analysis before the commencement of solar cell production. Virtual performance analysis of solar cell will support better design, analysis and control of photovoltaic systems. By an appropriate choice of software tool, performance prediction of solar cells will minimize time and cost by virtual production analysis. The work presented in this paper provides successful production rate of solar cell. For analysis, 10,000 samples have been taken into virtual performance analysis, where 9337 successful samples have been yielded. Thus, electrical yield of 92% have been obtained using photo luminescent (PL) imaging. It is observe that analysis without photo luminescent (PL) imaging have electrical yield of 86.2%. The defect rate of solar cells has been minimized through virtual performance prediction. Around 5.8% of increase in electrical yield has been observed through PL imaging. Keywords Renewable energy harvesting system · Photovoltaic systems · Virtual performance analysis · Electrical yield · Photo luminescent imaging B. Ashok Kumar (B) · T. S. Bagavat Perumaal · P. Seshadri Department of Electrical and Electronics Engineering, Thiagarajar College of Engineering, Madurai, India e-mail: [email protected] T. S. Bagavat Perumaal e-mail: [email protected] P. Seshadri e-mail: [email protected] S. Senthilrani Department of Electrical and Electronics Engineering, Velammal College of Engineering and Technology, Madurai, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_13

195

196

B. Ashok Kumar et al.

1 Introduction India is one of the thickly populated countries in the globe. The technological advancement and usage of more electronics and electrical gadgets by the people in such densely populated region has created power scarcity [1]. One such problem is satisfying consumer’s energy demand. Over the last few decades, the energy generation and energy demand couldn’t be met out effectively. The increase in demand can be satisfied with fossil fuel-based energy generation alone [2]. The depleting fossil fuel cannot support the increasing energy demand of the country in the next two decades [3]. Hence, there arises the need for alternative sources of energy. The shift toward alternative sources of energy has recently dominated by wind and solar energy. Thus, energy demand can be met easily with zero emission system (ZES) [4, 5]. Among the two types renewable energy-based systems, more preference is given to solar based because in country like India sunshine is abundant throughout the year [6, 7]. When compared to wind energy-based generation system, the installation and maintenance expenditure of solar energy-based generation system is cost effective. To promote clean energy system, Government offers subsidy to residential and industrial consumers. Indian government has set a target to achieve 100 GW by 2022 and this can be termed as solar revolution [8, 9]. Initially, the Indian government set its target as 20 GW capacity for 2022, which was revised as 100 GW of solar capacity by 2022 during the year 2015 [10]. To promote solar plant installation, India has established nearly 42 solar parks. Around 2.1 GW rooftop solar power generations satisfies 70% of commercial demand. In India, apart from large-scale grid-connected solar photovoltaic initiative, to satisfy local energy demand off-grid solar power systems are developed [8]. The rural energy needs are met out through solar lanterns and thus the utilization of kerosene has been drastically reduced. Similarly, under National program, 118,700 solar home lighting systems were installed and 46,655 solar street lighting installations were executed [11, 12].

2 Solar Cell Manufacturing Thin crystalline silicon is the raw material for solar cell fabrication, where the conductivity of cell is decided by intentional doping level of boron into silicon. Higher the level of boron increases the conductivity of solar cell. The doping process is a controlled process in the solar cell fabrication [13]. The most common method of silicon crystallization techniques are Czochralski (CZ) method and casting method [14]. Single-crystal ingot with high level of purity is obtained through CZ method, while in casting method multicrystalline silicon brick is obtained. Though casting method provides silicon with low cost, it comprises more defects and impurities.

Performance Prediction of Solar Cell Using Virtual Production Simulation

197

Hence, the electrical characterization will deviate in casting method. In both cases, silicon ingot is grown through controlled temperature chamber. The boron occupies crystal lattice when it is solidified. The resistivity of ingot is decided by the doped boron into silicon. During the fabrication process, a variation in the concentration of impurities can lead to auto doping process [15]. Such process occurs unintentionally in solar cell fabrication. The resultant of these processes is defects that increase recombination, thereby reducing the solar cell efficiency.

2.1 PV Factory In this work, the production of solar cells are analyzed using PV Factory, which a cloud-based simulation. The process of fabrication of PV cell in each step of processing can be analyzed using PV factory simulation tool. Since, the batch size of PV manufacturing is large in reality. Thus, analyzing the performance for such a large dataset will be time consuming. In order to perform simulation with smaller batch size, PV factory simulation tool can be deployed. In this tool, the step by step processing of PV cell is similar to actual manufacturing process of PV cell. PV Factory simulates the manufacturing of photovoltaic (PV) solar cells for smaller batch size of 10 or 20 cells per batch. The fabrication steps of PV cell from crystalline silicon wafers till electrical contacts are simulated by PV factory. The diffusing phosphorus into the silicon, deposition of silicon nitride film as antireflection coating and screen printing-based electrical contacts on either side of PV cells can be carried by this simulation tool. It is used to predict the factors relating PV cell efficiency with fabrication parameters. It provides a platform improve the efficiency of the solar cells by analyzing various parameters such as chemical concentrations, the temperatures and the gas flows.

2.2 PL Imaging The achievement of higher cell efficiencies, high yield at cost effective manner has resulted in adoption of improved quality control and process monitoring tools. PL imaging is well suited for such inline monitoring across the entire PV cell manufacturing process because of its minimum time of measurement and high repeatability. For the characterization of silicon samples across almost the entire photovoltaic (PV) value chain, Photoluminescence (PL) imaging is considered as a versatile technique. PL imagining pays a way for inline process monitoring in PV cell production. As fossil fuels are depleting in nature, there arises a need for search of alternate to fossil fuels.

198

B. Ashok Kumar et al.

Solar PV cells were considered as best alternative energy. To make this a cost effective technology, many researches are made toward cost reduction of PV cell manufacturing. Though there are various ways through which cost reduction can be attained, characterization of PV cell is considered to be the most essential method of cost reduction in PV cell manufacturing. An optical excitation is used by PL imaging, because it can be employed for wider range of samples. The throughput of PV cell production can be increased by reducing the risk of mechanical damage of samples. The defects in PV cell production can be minimized by deploying contactless imaging of samples. The silicon sample surface emits luminescence, when it is excited. This luminescent emission image is captured through a camera. Electroluminescence (EL) imaging requires electrical contacts. Thus, it is deployed to fully processed solar cells and modules alone. EL imaging is widely used for module inspection. Besides doping level, the thickness of silicon ingot that is sliced into thin wafers also decides the solar cell performance. During ingot slicing, the thickness of wafer determines the amount of light absorbed by solar cell. Higher the light absorption capability, higher the electron generations inside the solar cell [16]. To make highefficiency wafers from a wide range of wafers, manufacturers must produce more good wafers. Any defective wafer rejection will lead to unnecessary expense of solar cell production. For assisting this task, identification and removal of bad wafers at the start of a production line is highly mandatory. The condition or status of wafers produced can be examined with inline photo luminescent (PL) imaging [17]. For testing purpose, batch processing method has been adopted. Ingot is grown using casting method and boron doping with concentration of 4.78 × 1015 cm−3 the batch size of casting method is considered as 10,000 wafers with thickness of 190 µm. Since casting method-based ingot growth is cost effective, the proposed work deals with multicrystalline silicon ingot growth. At the end of PL imaging, 4.6% rejection has been obtained. Thus, 462 wafers are rejected from 10,000 wafers during PL imaging. Defect fraction and impurity fraction are set as rejection criteria during PL imaging [18]. The relation between number of wafers versus impurity fraction and defect fraction is shown in Figs. 1 and 2, respectively. The inference from it states that the rejection rate is minimum for each percentage of increase in impurity fraction. The rejection criteria applied to PL imaging is tabulated in Table 1 and based on which wafers are passed and rejected.

2.3 Chemical Etching After wafer slicing process, the solar cell needs isotexturing, which is process that chemically, etches the surfaces of multicrystalline silicon wafers [19]. Isotexturing removes saw-damaged surface layer and induce a textured surface. Such textures silicon cell can enhance the transmission of light into the solar cell [20].

Performance Prediction of Solar Cell Using Virtual Production Simulation Fig. 1 Statistics of passed and rejected number of wafers based on impurity fraction

Fig. 2 Statistics of passed and rejected number of wafers based on defect fraction

199

200 Table 1 Rejection criteria applied to PL imaging

Table 2 Impact of isotexturing on wafers

B. Ashok Kumar et al. Rejection criteria

Level (%)

Defect fraction

4.5

Impurity fraction

30

Solar cell parameters

Wafer Before isotexturing After isotexturing

Number of wafers Thickness (µm) Rate of removal through etchant (µm)

9532

9538

190

169.7

0

9.8

Since etching is a chemical assisted process and fabricators need to have an etch rate that is independent of the crystallographic orientation of the surface [21]. It is also important to ensure that sufficient etching of silicon at saw-damaged layer is performed. During the etching process, maintaining high optical transmission is also demanded for solar cell production. The isotexturing process is performed at a temperature of 25 °C with 25% of HNO3 and 20% of HF concentration. The chemical reaction takes places for 25 min time duration. Table 2 shows the impact of isotexturing on wafer, where the number of wafers placed for isotexturing are 9538 and wafers obtained after chemical reaction are 9532. The etchant removes 9.8 µm silicon from each side of wafer and thus yielding an average wafer thickness around 169.7 µm. The optical transmission angle of texturing is 58.9°. The textured solar cells are then loaded into a furnace subjected to high temperature in the presence of phosphorus gas. The diffusion of phosphorus atoms into silicon surface leads to n-type emitter region formation. The electron transportation to negative contacts is supported by emitter, which inturn minimizes the recombination and resistive losses in the solar cell [22].

3 Emitter Region Formation The performance of solar cell is dependent on emitter properties. When the diffusion of phosphorus concentration is high then resistive losses can be minimized [23]. The impact of diffusion on resistivity decides the number of wafer passed and rejected and is shown in Fig. 3. The diffusion of POCl3 at a temperature of 800 °C for a duration of 20 min leads to a sheet resistance of 126.8 Ω/sq. The average effective lifetime of electrons is 21.9 µs. The electron transportation provides an average emitter current density J0 as 841.8 fA/cm3 . The mechanical yield of POCl3 diffusion process provides 9529 wafers out of 9532 wafer supplied. Thus mechanical yield is 99.9% during POCl3

Performance Prediction of Solar Cell Using Virtual Production Simulation

201

Fig. 3 Statistics of passed and rejected number of wafers based on resistivity after diffusion of POCl3

diffusion process. The impact on number of wafers rejected and accepted after POCl3 diffusion based on impurity fraction and defect fraction are shown in Figs. 4 and 5. A very thin layer of silicon nitride (SiNx ) is coated in the front surface of solar cells is termed as antireflective coating (ARC) [24]. It reduces recombination by ‘passivating’ the surface and by providing a source of hydrogen that deactivates some defects in the bulk of the wafer. The rate of defects after diffusion in terms of wafers passed and rejected is shown in Fig. 6. A combination of plasma of silane (SiH4 ) and ammonia (NH4 ) gases react to form SiNx and hydrogen is supplied into a vacuum chamber were wafers are placed. The ratio of these gases and temperature impacts on the deposition rate and on properties of the film. A film thickness of 77.5 nm of SiNx is grown at a temperature of 375 °C. Thus mechanical yield is 99.8% during ARC process.

4 Extraction of Electricity A metal layer is formed to extract electricity from solar cell, where metal on front surface contacts n-type emitter and forms negative electrode. While metal on rear surface contacts p-type bulk of solar cell and forms positive electrode [25]. For

202 Fig. 4 Statistics of passed and rejected number of wafers based on impurity fraction after POCl3 diffusion

Fig. 5 Statistics of passed and rejected number of wafers based on defect fraction after POCl3 diffusion

B. Ashok Kumar et al.

Performance Prediction of Solar Cell Using Virtual Production Simulation

203

Fig. 6 Statistics of passed and rejected number of wafers based on DAD factor after POCl3 diffusion

preventing electrons from traveling to the positive contact and avoiding recombination at rear metal interface, a heavily doped p-type layer called a back-surface field (BSF) is formed. To ensure good contact between metal and silicon, the firing temperature and time must be properly selected. In case of more heat applied to paste, then metal layer may fail. Apart from this, firing process also releases hydrogen from the silicon nitride film and diffuses it into the bulk of the wafer where it can deactivate recombination defects, particularly crystallographic defects. With 45 µm finger width, 1500 µm spacing of metal grid, Ag paste thickness of 30 µm and Al paste thickness of 30 µm, a mechanical yield of 99.95% is obtained. This yield is obtained at a peak temperature of 810 °C.

5 Result and Discussion The virtual production analysis is carried out with a batch size of 10,000 wafers. The inference from analysis states that PL imaging-based production analysis provides electrical yield 92%, while without PL imaging the yield is 86.2%. The number of wafers failed during analysis is 745 for with PL imaging and 1376 for without PL imaging. Table 3 shows the comparison of virtual production analysis

204 Table 3 Comparison of virtual production analysis with and without PL imaging

B. Ashok Kumar et al. With PL imaging

Without PL imaging

Median V oc (mV)

640.6

640.3

Median J sc (mA/cm2 )

35.1

35.1

Median fill factor (%)

79.7

79.6

Median efficiency (%)

17.9

17.9

Cells failed

754

1376

Electrical yield (%) 92.0

86.2

Fig. 7 Comparison of with and without PL imaging based on J sc

with and without PL imaging. The comparison on number of cells based on J sc , Fill factor, V oc and efficiency are shown in Figs. 7, 8, 9 and 10.

6 Conclusion In recent days it has became mandatory to predict the solar cell performance in prior to installation. Such prediction will support better design, analysis and control of photovoltaic systems. In this work, performance prediction of solar cells using

Performance Prediction of Solar Cell Using Virtual Production Simulation Fig. 8 Comparison of with and without PL imaging based on fill factors

Fig. 9 Comparison of with and without PL imaging based on V oc

205

206

B. Ashok Kumar et al.

Fig. 10 Comparison of with and without PL imaging based on efficiency

PV factory emphasizes minimized time and cost by virtual production analysis. 10,000 samples under virtual performance analysis, yields 9337 successful samples. Thus, electrical yield of 92% is obtained using photo luminescent (PL) imaging. An electrical yield of 86.2% is obtained for analysis without photo luminescent (PL) imaging. Thus by using PL imaging, the defect rate of solar cells can be predicted before solar cell installation. It is inferred that using virtual performance prediction, defect rate is minimized.

References 1. Andrew RM (2020) Timely estimates of India’s annual and monthly fossil CO2 emissions. Earth Syst Sci Data 12:2411–2421. https://doi.org/10.5194/essd-12-2411-2020 2. Raghuwanshi SS, Arya R (2019) Renewable energy potential in India and future agenda of research. Int J Sustain Eng 12(5):291–302. https://doi.org/10.1080/19397038.2019.1602174 3. Owusu PA, Asumadu-Sarkodie, S, Dubey S (reviewing ed) (2016) A review of renewable energy sources, sustainability issues and climate change mitigation. Cogent Eng 3(1). https:// doi.org/10.1080/23311916.2016.1167990 4. Abbasi T, Abbasi SA (2012) Is the use of renewable energy sources an answer to the problems of global warming and pollution? Crit Rev Environ Sci Technol 42(2):99–154. https://doi.org/ 10.1080/10643389.2010.498754 5. Jurasz J, Canales FA, Kies A, Guezgouz M, Beluco A (2020) A review on the complementarity of renewable energy sources: concept, metrics, application and future research directions. Sol Energy 195:703–724. ISSN 0038-092X. https://doi.org/10.1016/j.solener.2019.11.087

Performance Prediction of Solar Cell Using Virtual Production Simulation

207

6. Bhattacharjee S, Bhattacharjee R (2021) Comprehensive solar energy resource characterisation for an intricate Indian province. Int J Ambient Energy 42(2):187–202. https://doi.org/10.1080/ 01430750.2018.1531257 7. Tyagi V, Rahim N, Abd Rahim N, Selvaraj J (2013) Progress in solar PV technology: research and achievement. Renew Sustain Energy Rev 20:443–461. https://doi.org/10.1016/j.rser.2012. 09.028 8. Goel M (2016) Solar rooftop in India: policies, challenges and outlook. Green Energy Environ 1(2):129–137. ISSN 2468-0257. https://doi.org/10.1016/j.gee.2016.08.003 9. India’s solar energy future: policy and institutions 10. India 2020—energy policy review 11. Sen R, Bhattacharyya S (2014) Off-grid electricity generation with renewable energy technologies in India: an application of HOMER. Renew Energy 62:388–398. https://doi.org/10.1016/ j.renene.2013.07.028 12. Feron S (2016) Sustainability of off-grid photovoltaic systems for rural electrification in developing countries: a review. Sustainability 8:1326. https://doi.org/10.3390/su8121326 13. Hudedmani M, Soppimath V, Jambotkar C (2017) A study of materials for solar PV technology and challenges. Eur J Appl Eng Sci Res 14. Kutsukake K (2019) Growth of crystalline silicon for solar cells: mono-like method. In: Yang D (ed) Handbook of photovoltaic silicon. Springer, Berlin, Heidelberg. https://doi.org/10.1007/ 978-3-662-56472-1_35 15. Koleske D, Wickenden A, Henry R, Twigg ME (2002) Influence of MOVPE growth conditions on carbon and silicon concentrations in GaN. J Cryst Growth 242:55–69. https://doi.org/10. 1016/S0022-0248(02)01348-9 16. Bhoopathy R, Kunz O, Juhl M, Trupke T, Hameiri Z (2020) Outdoor photoluminescence imaging of solar panels by contactless switching: technical considerations and applications. Prog Photovolt Res Appl 28:217–228. https://doi.org/10.1002/pip.3216 17. Schindler F, Giesecke J, Michl B, Schön J, Krenckel P, Riepe S, Warta W, Schubert MC (2017) Material limits of multicrystalline silicon from state of the art photoluminescence imaging techniques. Prog Photovolt Res Appl 25:499–508. https://doi.org/10.1002/pip.2836 18. Michl B, Padilla M, Geisemeyer I, Haag S, Schindler F, Schubert M, Warta W (2014) Imaging techniques for quantitative silicon material and solar cell analysis. IEEE J Photovolt 4:1502– 1510. https://doi.org/10.1109/jphotov.2014.2358795 19. Wu X, Li J (2020) Texturing technology on multicrystalline silicon wafer by metal-catalyzed chemical etching: a review. J Inorg Mater 361. https://doi.org/10.15541/jim20200361 20. Chen K, Liu Y, Wang X, Zhang L, Su X (2014) Novel texturing process for diamond-wire-sawn single-crystalline silicon solar cell. Sol Energy Mater Sol Cells 133. https://doi.org/10.1016/j. solmat.2014.11.016 21. Kolasinski KW, Unger BA, Ernst AT, Aindow M (2019) Crystallographically determined etching and its relevance to the metal-assisted catalytic etching (MACE) of silicon powders. Front Chem 6:651. https://doi.org/10.3389/fchem.2018.00651 22. Richards B, Richards S, Boreland M, Jamieson D (2004) High temperature processing of TiO2 thin films for application in silicon solar cells. J Vac Sci Technol A Vac Surf Films 339–348. https://doi.org/10.1116/1.1647594 23. Chen N, Ebong A (2015) Generalized analysis of the impact of emitter sheet resistance on silicon solar cell performance. Jpn J Appl Phys 54:08KD20. https://doi.org/10.7567/JJAP.54. 08KD20 24. Shanmugam N, Pugazhendhi R, Madurai Elavarasan R, Kasiviswanathan P, Das N (2020) Antireflective coating materials: a holistic review from PV perspective. Energies 13:2631. https:// doi.org/10.3390/en13102631 25. Lu Z, Lu PH, Cui J, Wang K, Lennon A (2013) Self-patterned localized metal contacts for silicon solar cells. J Mater Res 28:1984–1994. https://doi.org/10.1557/jmr.2013.204

Optimisation of FPGA-Based Designs for Convolutional Neural Networks P. L. Bonifus , Ann Mary Thomas, and Jobin K. Antony

Abstract Convolutional neural networks (CNN) is a widely known deep learning architecture and achieves higher accuracies in speech processing and computer vision applications. CNN has arisen as best option in the fields such as video surveillance, mobile robot vision, smart factories, and medical diagnostics. CNN has shown its tremendous potential for image understanding in Cardio-Vascular (CV) systems. But, the tremendous potential of CNNs came up with additional computational burden. To get an outstanding performance of CNN-based algorithms, it requires large amount of enormous applications and memory resources. Matrix multiplication (MM) is the most fundamental computational operation taking place in CNN. Researchers in this field found that multiplication operation is the most resource-intensive and power-hungry operation in CNN. This work aims to design an FPGA-based CNN using systolic array which can improve both the accuracy and hardware efficiency of convolutional neural network. Here, we made design optimisation in multiplier unit which brings a significant impact on overall performance of convolutional neural networks. Keywords Deep learning · CNN · FPGA · Bfloat-16

P. L. Bonifus (B) · A. M. Thomas · J. K. Antony Department of Electronics & Communication Engineering, Rajagiri School of Engineering & Technology, Kochi, Kerala, India e-mail: [email protected] A. M. Thomas e-mail: [email protected] J. K. Antony e-mail: [email protected] APJ Abdul Kalam Technological University, Thiruvananthapuram, Kerala, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_14

209

210

P. L. Bonifus et al.

1 Introduction 1.1 General Overview Hardware implementation of deep learning (DL) by means of convolutional neural networks (CNNs) has received wide acceptance recently. It has been widely used in various applications ranging from handwriting recognition to medical diagnosis and much more. Structure of optic nerves in living beings was the inspiration for CNNs. Approximately 86 billion (86 × 109 ) neurons are there in a human brain. Likewise, artificial neuron is the basic building block of neural network. Lot many artificial neurons are interconnected to form a Neural Network. Hence, CNN model demands a lot of computation. This makes the training and deployment phase to consume more power. In real-world applications, graphic processing units (GPUs) are commonly used as a platform for CNN. But, GPUs are inherently power-hungry devices; this will limit their usage in embedded applications like transportable gadgets and wearable devices. In low-power budget systems like portable devices and in other embedded systems, field programmable gate arrays (FPGAs) or application specific integrated circuits (ASICs) could be considered as an alternative for GPUs because of their lesser power consumption compared to GPUs. Among FPGA and ASIC-based designs, FPGA-based accelerators offer low non-recurring engineering (NRE) cost, less deployment time, and reconfiguration capability. Reconfiguration capability of FPGAs makes them a better choice for neural networks. FPGAs are configurable hardware devices that have adaptable components. When comparing with other processor-based CNN accelerators, accelerator implementations as soft logic on FPGA have the following advantages [1]. • Because of rich parallel computing resources and high energy efficiency, FPGAbased accelerators are considered as perfect for deep convolutional neural network. • Reprogramming property of FPGA reduces their electric requirements and offers them higher overall performance in terms of acceleration and throughput. • FPGA are more resistant to rugged settings and environmental factors and have longer life span than that of GPU. • FPGAs are already using in AI products by some companies. An example is Microsoft, which provides its FPGA-based powered machine learning technology for its Azure cloud services. • FPGA offers high energy efficiency compared to CPU or GPU. Therefore, FPGAs can be optimised for specific styles of architectures inclusive of CNN.

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

211

1.2 Motivation Although the basics of CNN were proposed way back in the 1980s, hardware implementation was not popular because of the lack of enough computational resources. Because of the advancements in VLSI technology, such as 3D stacking and Finfets, now transistor count and hence processing power is not a big concern. Due to this, hardware implementation of CNNs is again in limelight. Deeper and larger CNNs require more computational resources. Reducing the complexity required for these computational resources could reduce the overall design complexity and could reduce the computational delay and computational power required.

1.3 Objective This work aims to improve the hardware implementation efficiency of CNN by reducing its computational complexity in the multiply and accumulation (MAC) unit.

2 Literature Review The aim of this literature review is to study the existing FPGA-based CNN architecture and its design optimisations which improves the hardware implementation efficiency and prediction accuracy of Neural Networks. There are several researches were done to optimize or reduce the computational complexity of CNN by applying certain transformations to the matrix multiplication operation. Several methods proposed by different researchers are explained here. A most commonly observed issue in FPGA-based accelerator design is that the computations required may not match with the memory bandwidth available in an FPGA platform; this may cause either underutilisation of logic resources or memory bandwidth. In order to address this issue, a roofline model-based scheme using analytical design was proposed [1]. This implementation was having a throughput of 61.62 GFOP/s under 100 MHz working frequency. Limitation of this technique is that it only considered small CNN models for simple task. In [2], describes a dynamic-precision method which quantises the data to have a visible reduction in bandwidth required and reduction in the number of resources utilised. This research says that implementation of computational units on FPGAs using fixed point logic is comparatively more efficient than floating point computational units. This system offers performance at 187.8 and 137 GOP/s under 150 MHz frequency. Main drawback of this method is time-consuming when we retrain the model.

212

P. L. Bonifus et al.

In [3], explains about a FPGA-based accelerator in which all convolutional layers were mapped onto a single chip and thereby concurrent processing can be done by different layers. This technique of computations on batches helps to increase bandwidth utilisation required for memory access. This system achieves a maximum performance of 565.4 GOP/s under 156 MHz clock frequency and 391 GFPO/s under 156 MHz clock frequency. Main drawback is the sufficiently high time required to retrain the model. Depth-wise separable convolution design technique [4] studied the possibilities of reducing computational complexities in a standard convolution. Reduction in computational complexities could make neural network applications more suitable for embedded system platforms such as portable gadgets and wearable systems. This system is able to do each classification job 3.75 ms using ImageNet database and have 20× times speed if compared to CPU. Limitation of this technique is time consuming when we retrain the model. The method proposed in [5] can solve the issue of time-consuming when retraining a CNN model. Here refers a design technique called block-floating point arithmetic which increased hardware implementation efficiency and energy efficiency by three times. In this method, a highly optimised floating point unit makes use of blockfloating point concept is first developed to reduce the hardware implementation cost and off-chip data traffic. To improve the throughput of the hardware, a memory access pattern, optimised off-chip buffer memory, and a 3-level parallel convolution processing units were adopted. Researchers like Guo et al. and Han et al. found that computations with lowprecision arithmetic are sufficient for doing inference part in deep learning. They were able to cut down the implementation complexity for modelling the neural network, by reducing the arithmetic precision and found that accuracy reduction was within tolerance limit. Compared to the reduction in hardware complexity accuracy reduction was very minimal. Fast Fourier Transforms (FFTs) can be used to reduce the computational complexity in CNNs [6]. Accuracy loss can be prevented but found useful only for networks having large kernels. Toom and Cook proposed an algorithm based on minimum filtering called as Winograd’s algorithm. Winograd’s algorithm is able to reduce the count of multiplication (floating point) operations up to 2.25 times inside a CNN model. Here, instead of dot product, calculating the resultant matrix using specific formula. Main drawback of this algorithm is that it can only work in neural networks with small kernels [7].

3 Impact of Matrix Multiplier on CNN Convolutional neural network is basically made up of three different types of layers: convolutional layer, pooling layer, Rectified Linear Units (ReLU) correction layer,

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

213

and fully connected layer. The convolutional layer can successfully record the spatiotemporal dependencies of an image by applying relevant filters. Matrix multiplication is the fundamental operation inside a convolutional layer. That means convolution is nothing but a matrix operation between ‘k’ of filter and portion ‘p’ of image over which the filter is hovering. Consider an example of recognizing of a 224 × 224 RGB image using CNN when we are applying in a VGG-16 model, the complete process contains 138 billion (138 × 109 ) parameters. To do the feed-forward operation for an RGB image of size 224 × 224, the whole network requires 30 billion (30 × 109 ) addition and 30 billion (30 × 109 ) multiplication operations. Hence, we can summarize from this example that, CNN is a computationally complex network. Researchers in this field found that multiplication operation is the most resource-intensive and power-hungry operations in CNN. Thus, any improvement made on multiplier unit brings a significant impact on overall performance of CNN [5].

4 Proposed System To address such challenges dealt in previous sections, here we proposed a suitable architecture called systolic array architecture using bfloat16. This system is able to accept inputs one by one, perform multiplication of two inputs, and accumulate the result with previous multiplication results as required by the CNN architecture. This section discusses on systolic array and its advantages to develop a scalable CNN implementation.

4.1 Systolic Array Architecture In systolic array architecture (SA), processing elements (PEs) are arranged as a 2dimensional array arrangement with pipelining capability in order to reduce delay for data movement. This architecture is capable of accelerating general matrix multiplication (GEMM) and being accepted widely in academia as well as in industry. The systolic array family can reuse intra-PE operand and have high data path efficiency results in reduction of circuit area and power dissipation. The basic idea of systolic array architecture is mapping high-level computation into hardware structures. Systolic architecture allows multiple computations for each memory access. Hence, it can execute compute-bound problems at high speed without increasing input–output requirements. Main attractive features are regularity, re-configurability, and scalability. Because of the high level of parallelism incorporated, systolic array architecture can offer high rate of throughput. Data flow through this architecture is having a regular layout pattern and simple to understand. Systolic array features low data transfer latency and enables higher clock frequencies because of its own regular

214

P. L. Bonifus et al.

layout pattern and support for local communication. Thus, this architecture is suitable for complex parallel designs on FPGAs. Because of these advantages, systolic array architecture is being widely accepted for using in hardware implementation of matrix multiplications, especially in reconfigurable devices like FPGA [8].

4.2 Systolic Array Design Figure 1 shows the two-dimensional 4 × 4 systolic array representation. The 4 × 4 refers the arrangement of processing element in both row and column wise. There are total of 16 processing elements are present inside a 4 × 4 systolic array. Multiply and accumulate operation (MAC) is taking place inside each processing element [9]. Let’s look on how matrix multiplication operation carried out inside the systolic array which is illustrated in Fig. 2. At first, here it is Matrix B, fed in row-wise which is then moved down in the array. Next, the other matrix, in this case, Matrix A will be fed in a column-wise order and move to right from left. Till each processor is packed with a complete row and a complete column, dummy values will be issued. Once, each processor is packed with a complete row and a complete column, the final result of multiplication can be saved in an array and could be taken as output either as single row in one time instant or single column in one time instant.

Fig. 1 4 × 4 systolic array representation

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

215

Fig. 2 Working of 4 × 4 systolic array

In this systolic array, each rectangular box represents the processing element (PE). Processing elements placed on the periphery of the design can only communicate with the exterior environment. Processing steps in a single processing element can be summarized as it receives an input, multiply and accumulate (MAC) unit inside the processing element compute with the inputs, and finally transmits the corresponding output to neighbouring cell [9]. Figure 3 shows a single MAC unit. Systolic array is interconnection of array of processing elements. Each processing elements composed of multiply-accumulate unit (MAC) unit. MAC unit performs the operation of both multiply and accumulates. Principle operation of MAC unit is evaluating partial products and then partial products result is added by previous stage result. Here, the single element from matrix A is multiplied with single element from matrix B. Then the multiplication result is added with previous cell’s output [4]. As the first phase of project, implemented 32 × 32 integer-based systolic array in Verilog RTL. Vivado 2020.1 is used for simulation.

4.3 IEEE 754-Based Floating Point Advantage in using any fixed point denotation is performance and disadvantage is that we can only represent a very few range of values. Thus, it is usually inadequate for numerical analysis it does not permit sufficient numbers and accuracy. Currently, IEEE 754 is the widely accepted single precision standard used by personal computers

216

P. L. Bonifus et al.

Fig. 3 Single MAC unit

using MACs, Windows, and most of the Unix platforms. Float-based number representation is used when more precision is needed. To design our CNN system with more precision and accuracy, this project work opting floating point-based arithmetic operation in our architecture [10].

4.4 Bfloat-16 From the very beginning itself, the workloads for machine learning tasks may take hours or even days to run. In order to help organisations, by reducing the workload running time, recently Google Inc., designed and developed custom processors for deep learning applications named tensor processing units (TPUs). And this cloud TPU achieves high performance because it is working with innovative floating point format called as Bfloat-16 or Brain Float-16. The high-performance capability of

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

217

cloud TPUs was achieved with the help of Bfloat-16. And also it helps to improve hardware efficiency with minimal effort for switching from FP-32. Thus, we can achieve even performance with Bfloat-16 [11].

4.5 Floating Point Versus Bfloat-16 There are many floating point formats you can hear about in the context of deep learning. In this work, we are only dealt with IEEE 754 Floating point-32 and Bfloat16. As you can see in Fig. 4, single precision floating point format defined in IEEE 754 standard (FP-32) is divided into 1 bit for sign (S) representation, 8 bits for exponent part (E) representation, and 23 bits for mantissa (M) or fractional part representation. FP-32 is a 4-byte 32-bit floating point number ranges from ~1e−38 to ~3e38 . When it comes to fp-16 format, Mantissa bits are reduced to 10 bits but precision of FP-16 is reduced by half when compared to FP-32. As well, FP-16 can only represent less range of numbers. Here comes the relevance of newly developed innovative floating point format called bfloat-16. Bfloat-16 is a custom-tailored single-precision floating-point representation using only 16 bits, proposed for machine learning applications. Bfloat-16 consists of 1 sign bit (S), 8 exponent bits (E), and 7 mantissa bits (M). Even with this 16-bit representation, Bfloat-16 can represent same range of values as that of fp-32 floating point format. And also, there is no loss of accuracy and precision. Thus, artificial intelligent groups at Google found that bfloat-16 format provides same results similar to the fp-32 format while reducing the hardware implementation complexity, reducing the memory usage, and thus increasing the overall performance [11]. In this project work, simulated and synthesised 4 × 4, 8 × 8, 16 × 16, and 32 × 32 systolic arrays using both IEEE 754 floating-point standard and bfloat-16 format.

Fig. 4 fp32 versus fp16 versus bfloat16

218

P. L. Bonifus et al.

Output waveform obtained after the simulation of 32 × 32 IEEE 754 floating pointbased systolic array is shown in Fig. 5. Similarly, designed 32 × 32 bfloat-16-based systolic array and its corresponding waveform is shown in Fig. 6. Finally, compares the results with respect to power consumption, area utilisation, and critical path delay.

Fig. 5 IEEE 754 standard float-based 32 × 32 systolic array

Fig. 6 Bfloat-16-based 32 × 32 systolic array

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

219

5 Results and Discussions This project work includes the simulation and synthesis of 4 × 4, 8 × 8, 16 × 16, and 32 × 32 systolic arrays using both IEEE 754 standard floating point and Bfloat-16. The systolic array is modelled using Verilog HDL. For simulation and synthesis, Vivado 2020.1 is used. From the synthesis report targeted for Zynq xc7z020clg4841 device, obtained the information on how much power each type of systolic array (whether it is IEEE 754 standard float standard or Bfloat-16) is consumed, area utilization, and critical path delay. The exact values of power consumption, hardware utilisation, and critical path delay obtained after synthesising Verilog models of different systolic array sizes based on IEEE 754 implementation against Bfloat-16 are shown in Table 1. Based on that information, reached in an inference that Bfloat16-based systolic array architecture requires less power, less number of hardware resource, and hence less value of critical path delay when compared with IEEE 754 standard floating point. Hence, we can say that implementing bfloat-16-based systolic array architecture on CNN network can improve its hardware efficiency and gives us high throughput.

5.1 Analysis Table The comparison of power consumption and area utilisation by different size systolic array with respect to both IEEE 754 standard float standard or Bfloat-16 is plotted on Figs. 7 and 8. Figure 7 shows graphical representation of power consumed by different size of systolic array designed with respect to IEEE 754 standard floating point and Bfloat16. Here, x-axis depicts different size of systolic array, which are 4 × 4, 8 × 8, Table 1 Analysis of systolic arrays implemented using IEEE 754 standard floating point and Bfloat16 IEEE 754 4×4 array

BFloat16 8×8 array

16 × 16 array

32 × 32 array

4×4 array

8×8 array

16 × 16 array

32 × 32 array

LUT as logic (%)

19.06

45.8

71.3

86.2

8.6

34.2

63.9

70.09

Registers as flip flops (%)

2.43

9.6

24.3

47.1

1.1

4.67

16.9

33.3

Power (W) 607.17

3160.5

13,588.1

24,279.6

247.6

43.4

740.2

999.6

Critical path delay (ns)

20.725

30.15

45.4

11.03

12.61

15.22

20.3

13.73

220

P. L. Bonifus et al.

Fig. 7 Power consumption—floating point versus Bfloat

Fig. 8 Area utilization—floating point versus Bfloat

16 × 16, 32 × 32, and y-axis depicts power consumption in Watts (W). Blue line represents systolic array designed with respect to IEEE 754 standard float. While red line represents systolic array designed with respect to Bfloat-16 format. By analysing this graph, we can infer that systolic array based on bfloat-16 consumes less power when compared to IEEE 754 standard float standard format. Figure 8 shows graphical representation of area utilised by different size of systolic array designed with respect to IEEE 754 standard floating point and Bfloat-16. Here, x-axis depicts different size of systolic array, which is 4 × 4, 8 × 8, 16 × 16, 32 × 32, and y-axis depicts the percentage of area utilization. Red line represents systolic array designed with respect to IEEE 754 standard float. While violet line represents systolic array designed with respect to Bfloat-16 format. By analysing this graph, we

Optimisation of FPGA-Based Designs for Convolutional Neural Networks

221

can infer that systolic array based on bfloat-16 requires less area when compared to IEEE 754 single precision floating point standard.

6 Conclusion This paper aims to improve the hardware efficiency of CNN by reducing its computational complexity in multiplication unit of convolutional neural network. For that, here proposed a systolic array architecture to carry out matrix multiplication operation inside the convolutional layer of CNN. This architecture offer us to gain the advantages of parallel computing. Then, designed the same systolic array architecture implementation in two different floating-point number representations. They are IEEE standard floating point and Bfloat-16-based format. Finally, we arrived at a conclusion that bfloat-16-based systolic array architecture can reduce the area utilisation, reduce the power consumption, reduce the critical path delay and hence improve the overall performance metrics of CNN implementations targeted on FPGA.

References 1. Zhang C, Li P, Sun G, Guan Y, Xiao B, Cong J (2015) Optimizing FPGA-based accelerator design for deep convolutional neural networks. In: Proceedings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, Feb 2015, pp 161–170 2. Lian X, Liu Z, Song Z, Dai J, Zhou W, Ji X (2019) High-performance FPGA-based CNN accelerator with block-floating-point arithmetic. IEEE Trans Very Large Scale Integr (VLSI) Syst 27(8) 3. Li H, Fan X, Jiao L, Cao W, Zhou X, Wang L (2016) A high performance FPGA-based accelerator for large-scale convolutional neural networks. In: 2016 26th international conference on field programmable logic and applications (FPL), Sept 2016 4. Bai L, Zhao Y, Huang X (2018) A CNN accelerator on FPGA using depthwise separable convolution. IEEE Trans Circuits Syst II 65(10) 5. Qiu J et al (2012) Going deeper with embedded FPGA platform for convolutional neural network. In: Proceedings of the ACM/SIGDA international symposium on field programmable gate arrays, pp 26–35 6. Zhao Y, Wang D, Wang L, Liu P (2018) A faster algorithm for reducing the computational complexity of convolutional neural networks. www.mdpi.com/journal/algorithms. Received: 10 Sept 2018; Accepted: 16 Oct 2018; Published: 18 Oct 2018 7. Winograd S (1980) Arithmetic complexity of computations. SIAM, Philadelphia, PA 8. Liu Z-G, Whatmough PN, Mattina M (2020) Systolic tensor array: an efficient structured-sparse GEMM accelerator for mobile CNN inference. IEEE Comput Archit Lett 19(1) 9. Yang Z, Wang L, Ding D, Zhang X, Deng Y, Li S, Dou Q. Systolic array based accelerator and algorithm mapping for deep learning algorithms 10. https://www.sciencedirect.com/topics/engineering/floating-point-number 11. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-per formance-on-cloud-tpu

Design and Implementation of an Automated Fuel Station M. Jyothirmayi, Vibha B. Raj, V. Lekhana, and P. Manjunath

Abstract Currently, fuel stations are operated manually which requires a lot of time and staff to operate all the system. Customers are defrauded while dispensing the fuel by not delivering the exact amount of fuel, and during payments, invalid notes may be exchanged. So, an automated fuel station can be a solution to all the above-stated problems. In this project, an automated fuel station is designed and implemented which can work 24/7 conveniently. An automated fuel station consists of an automated fuel dispensing system, automatic level indicator of the fuel in the storage tank, automatic vehicle washing system, smart lighting system, and automatic fire extinguisher. An automatic fuel dispensing system is realized by interfacing RFID, GSM module with Arduino microcontroller. Level indication of the fuel in the storage tank is made automatic using an ultrasonic sensor. Vehicle washing is done by spraying water controlled automatically through the movable frame. Light intensity is automatically controlled based on the surrounding ambiance light and presence/absence of a vehicle. The accidental fire in the station is handled by an automatic fire extinguisher. Keywords Automatic fuel dispensing · Fuel level monitoring · Vehicle washing system · Arduino microcontroller · Lighting system · Fire extinguisher

1 Introduction In recent years, fuel stations have gone through major changes in terms of its functioning. The primary requirement of the fuel station is not only to supply petrol, gas, air, and other gasoline products, it also caters to other facilities like cafeteria, child play area, and other essentialities. To maintain these facilities for 24/7, huge M. Jyothirmayi · V. B. Raj (B) · V. Lekhana · P. Manjunath Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bangalore, India e-mail: [email protected] M. Jyothirmayi e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_15

223

224

M. Jyothirmayi et al.

manpower is required and more power is utilized. An automated fuel station is a solution to reduce the man power and save energy. Currently, most of the fuel stations in India are operated manually, though there is the use of a microcontroller for fuel dispensing. The main aim of this work is to develop a fully automatic fuel station by reducing the manpower used for filling the fuel, collecting cash, vehicle washing and checking the level of fuel in the storage tank and also power saving. Along with minimal usage of manpower, malpractices in dispensing of fuel should be prevented and timely service needs to be provided. The primary part of any fuel station is dispensing. An automated fuel station needs to have an automatic fuel dispensing system to get the exact amount of fuel as per the user requirement. This is achieved by providing a smart card (RFID tag) [1, 2] to all the customers using which they can operate individually for the facilities provided in the station and then enter the quantity of fuel to be dispensed. After the dispensing of fuel as per the customer’s requirement, the amount is deducted from the preloaded money in the RFID tag and the transaction details are sent to the customer’s mobile phone through the GSM module. Continuous level monitoring of the fuel in the storage tank is necessary to ensure that there is no shortage of fuel and provide 24/7 service. Therefore, an automatic level indication of the storage tank is designed which displays the fuel level using an ultrasonic sensor [3, 4], which is presently done manually using a dipstick/dip rod. An ultrasonic sensor is interfaced with the Analog-to-Digital Controller (ADC) which converts analog output from the sensor to the digital form and feed it to the microcontroller. Then, the microcontroller calculates the level according to the digital value obtained and displays the results. Providing quick services are essential in a fuel station. Hence, an automatic vehicle washing system is realized, which can save a lot of time for the customers. The washing process takes place in six steps, i.e., check if the vehicle is present on the conveyor, soapy water spraying, brushing, cleaning with normal water, drying, exit from the conveyor. IR sensors [5] are used to detect the vehicle in each step. The vehicle is parked on the conveyor and washed along with the movement of the conveyor [6]. Energy-saving is a prime objective of any system, therefore implementing a smart lighting system in the fuel station can reduce almost half of the energy usage. The implementation of smart lighting system in street lights is carried out using a motion sensor and a brightness sensor [7, 8]. The street lamps turn on automatically only when it is dark and a movement is detected. When there is no movement, the lamps are automatically turned off or adjusts intensity based on the ambient light in the environment. The sensor outputs are sent to the microcontroller, which analyses the data received and sets the intensity of the lights accordingly. Wireless technology is implemented using the ZigBee module for communication between the sensors, LED, and microcontroller. Thus, making it an energy-efficient system. Safety being an important parameter in fuel stations, a check for fire is a major requirement as fuel is a highly flammable. Most disasters can happen if proper precautions are not taken. Hence, an automatic fire extinguishing system [9, 10] is realized using a temperature sensor, smoke sensor, and solenoid valve. The sensors are used to detect the presence of fire and the valve regulates the opening and closing of the nozzle of the fire extinguisher.

Design and Implementation of an Automated Fuel Station

225

Since the advancement of technology has led to an increase in automation of all industries, our idea is to automate not just the dispensing pump, but also automate other systems such as level indication of the fuel storage tank, vehicle washing system, lighting system, and safety of the fuel station. Hence, this work attempts to integrate different sub-systems and to automate its use so that the petrol dispensing system works independently with very less manpower and to understand the working of different sub-systems for use as an automated system for managing the petrol station.

2 Methodology The suggested system is depicted in Fig. 1. The design of automated fuel station includes an automatic fuel dispensing pump, automatic fuel level indicator, automatic vehicle washing system, smart lighting system, and automatic fire extinguisher. The next sections go through the details of each system block.

2.1 Automatic Fuel Dispensing Pump Fuel dispensing in fuel stations demands more staff and is time-consuming. In addition to that, malpractices and fuel thefts are happening. To overcome these issues, it Fig. 1 Sub-systems of an automated fuel station

226

M. Jyothirmayi et al.

Fig. 2 Block diagram of an automatic fuel dispensing pump

is necessary to automate the dispensing pump. RFID and GSM technology are integrated with the fuel dispensing system to make it completely automatic. A detailed block diagram of the automatic fuel dispensing pump is depicted in Fig. 2. The block diagram shows various components interfaced with the microcontroller. Arduino Mega 2560 is the microcontroller used, which controls all the other components. A smart card (RFID tag) is issued to all the customers, which needs to be preloaded with money and it is scanned by the RFID reader. If the card is valid, the LCD displays a message that the access is granted and the process continues. Next, the customer has to enter the password using the keypad and if the password is correct, the option for entering the amount of fuel required in terms of rupees (Rs.) or milliliters (mL) is displayed on the LCD. The user has to select one of the options and enter the amount accordingly. The amount of fuel to be dispensed and money to be paid is calculated by the microcontroller using the input given by the customer. The payment amount is checked with the balance available in the card, and if it is within the balance the amount is deducted. The LCD shows the remaining balance in the card and the amount of fuel to be dispensed. The microcontroller gives a command to the relay to run the pump. The pump dispenses the exact amount of fuel as per the inputs given by the user. Every process taking place is displayed on the LCD for the user’s convenience. Also, a message is sent to the user’s mobile phone regarding the money deducted and the remaining balance using the GSM module. The data about the money collected and the amount of fuel dispensed to every customer is stored in the workstation.

Design and Implementation of an Automated Fuel Station

227

Fig. 3 Block diagram of an automatic fuel level indicator

2.2 Automatic Fuel Level Indicator The level measurement has been a major issue especially when it comes to flammable substances. So, a new method has to be adopted to measure and monitor the level. Continuous monitoring of fuel level in the storage tank can be done by implementing an automatic fuel level indicator (Fig. 3). In this system, an ultrasonic sensor is used to detect the object and to measure the distance between the object and itself. This sensor has two transducers one is the transmitter and the other is the receiver. The ultrasonic sensor will be placed on the top of the storage tank. The transmitter will generate ultrasonic waves for every microsecond and the receiver will absorb the micro sound waves which bounce back from the surface of the fuel level. Arduino will calculate the level of the fuel based on the time taken by the sound waves to touch the surface of fuel and bounce back to reach the receiver. The calculated result of the fuel level will be displayed on the LCD. If the level of the fuel is less than the minimum level, then the dispensing of the fuel will stop.

2.3 Automatic Vehicle Washing System In this developing world, automation has become a demand. It allows us to save time, money, and manpower. It’s necessary to have a well-functioning and effective method for keeping the vehicles clean. A vehicle washing system is a simple maintenance technique for keeping the vehicle’s exterior clean (Fig. 4). The exterior of a vehicle must be kept clean to prevent rust, oxidation, and scratches. With the help of a microcontroller, this technology cleans the vehicles automatically. This process is divided into two parts: washing and drying. To begin with, the user will scan an RFID tag on the reader, and the associated service charge will be displayed on the LCD. To begin the vehicle washing procedure, the user must press the start button, after pressing the button the appropriate amount will be deducted and the Arduino is initialized. Car washing consists of three major steps: first, clean water is sprayed over the vehicle, then detergent water is sprayed, and

228

M. Jyothirmayi et al.

Fig. 4 Block diagram of an automatic vehicle washing system

last, regular water is sprayed. To wash the car, the spraying frame will move forward and backward at a set speed and time for each step. After that, the vehicle is dried by switching the DC fans, fixed around the washing area. The traditional way of car washing takes a long time and uses a lot of water. This may be prevented by adopting this automatic vehicle washing system, which allows the vehicle to be washed in a shorter amount of time while using less water. Further the used water is also filtered and recycled in a water recycling thereby saving the water used.

2.4 Smart Lighting System Automatic light operation with power saving. Using the Light dependent resistor (LDR) sensor the ambient light is measured and the sensed value or the information is sent to the Arduino. Arduino accordingly calculates and sets the brightness of the LEDs (these LEDs are for the whole fuel station) as per the requirement of the station. Another set of LEDs are placed near the functional areas. Infrared (IR) sensor is used to detect object. As a vehicle approaches toward the station the IR sensor detects the vehicle and sends the information to the Arduino. Arduino checks for ambient light in the area where the vehicle has arrived, if the ambient light in that area is dark, then it makes the light to glow only in that part of the station. This makes the station use a minimum amount of power. Hence making it an energy-efficient system (Fig. 5).

Design and Implementation of an Automated Fuel Station

229

Fig. 5 Block diagram of a smart lighting system

2.5 Automatic Fire Extinguisher Fire causes damage to both life and property, therefore a fuel station’s assurance of safety is very necessary. An automatic fire extinguisher can prevent any type of losses by timely detection and quick extinguishing of the fire. The block diagram of an automatic fire extinguisher is given in Fig. 6. Fire in the fuel station is mainly caused due to fuel leakage or short circuit. Water cannot be used to extinguish this type of fire, instead CO2 -based fire extinguishers should be utilized. To realize a prototype of a compressed CO2 type of fire extinguisher, a balloon is used as a dupe to the compressed CO2 cylinder as shown in Fig. 7. A flame sensor and a smoke sensor are fixed in areas where there is a higher risk of fire. For extinguishing the fire, we need to burst the balloon (similar to the opening the nozzle of the fire extinguisher cylinder), for which servo motor is used. Buzzers are used to alert and alarm the surroundings. The sensors detect the fire and send the information to the microcontroller, which then gives command to the servo motor to rotate and turn on the buzzers. A needle is fixed to the servo motor so that the needle rotates along with the servo motor and bursts the balloon. The air inside the balloon rushes out when burst and blows off the fire.

Fig. 6 Block diagram of an automatic fire extinguisher

230

M. Jyothirmayi et al.

Fig. 7 Balloon used as a dupe to the compressed cylinder

3 Results An automated fuel station is designed and implemented to a project prototype, which is tried, executed and accurate results are obtained. In this section, we evaluate the characteristics of the components utilized in realizing the sub-systems of the automated fuel station.

Design and Implementation of an Automated Fuel Station

231

3.1 Automatic Fuel Dispensing Pump Calibration of the pump is critical in order to receive the correct amount of fuel, delivered by the dispensing system. A non-submersible DC pump is used in implementing the automatic fuel dispensing system. The effect of the fuel level in the storage tank on the quantity of fuel being dispensed by the pump is plotted in Fig. 8. Table 1 contains the quantity of fuel delivered by the pump at different fuel level in the storage tank for various time intervals. The table and plot indicate that the amount of fuel delivered is unaffected by the level of fuel in the storage tank. It is also important to know if the pump dispenses the fuel accurately, Fig. 9 depicts the plot and Table 2 contains the reading taken for the quantity of fuel dispensed by the pump versus the time interval it runs. According to the results, the pump can dispense 30 mL of fuel in 1 s and the relationship between the quantity and the time is linear. Figure 10a shows the message delivered to the customer’s mobile phone through the GSM module, and Fig. 10b depicts the data logged in the workstation for the quantity of fuel dispensed and money received.

Fig. 8 Plot depicting the effect of fuel level in the storage tank on the fuel quantity dispensed

Table 1 Level versus quantity for different time interval

Level

Time (s) 2

4

5

Full (mL)

60

120

150

Half (mL)

60

120

150

Minimum (mL)

60

120

150

232

M. Jyothirmayi et al.

Fig. 9 Plot depicting the quantity of fuel dispensed versus time Table 2 Quantity of fuel dispensed

Time (s) 1

Quantity (mL) 30

2

60

3

90

4

120

5

150

10

300

Fig. 10 a Messages received by the customer, b data logging

Design and Implementation of an Automated Fuel Station

233

3.2 Automatic Fuel Level Indicator The experiment was to calibrate level measurement from the reading obtained from the ultrasonic sensor to relate liter and distance measured in cm. The height of the storage tank is 22 cm (prototype) and width is 10 cm based on the distance the level is calculated. The ultrasonic sensor is fixed on top of the tank below the lid. When the distance is read as 2 cm, it means that the fuel level is full and when it shows 20 cm it means that storage tank is almost empty (0.2 L). Hence, the distance to the level of the fuel is inversely related as shown in Fig. 11 (Table 3).

Fig. 11 Plot depicting the level-distance

Table 3 Fuel level measurement

Distance in cm

Level in liter

2

2.0

4

1.8

6

1.6

8

1.4

10

1.2

12

1.0

14

0.8

16

0.6

18

0.4

20

0.2

234

M. Jyothirmayi et al.

Fig. 12 Process of vehicle washing system

3.3 Automatic Vehicle Washing System The automatic vehicle washing system is demonstrated as a prototype, in which a vehicle enters a washing station and is cleaned automatically. The entire system is automated, so no human intervention is necessary, the system is built on Arduino automation, so all work is done digitally, and only mechanical assembly is required. Vehicle is parked inside the washing area as shown in picture 1 (Fig. 12). Then, the regular water will be sprayed on the vehicle in the forward to backward movement of the frame attached with the sprayers as in picture 2 (Fig. 12). In the next forward to backward movement of the frame, soap water is sprayed. Again regular water is sprayed to wash off the soapy texture on the vehicle. At the end, the vehicle is dried by the dryers which are placed around the frame. By adopting this automation system, benefits such as decreased operating time and manpower can be achieved.

3.4 Smart Lighting System The IR sensor detects the objects and sends a signal in the form of voltage to the microcontroller through the analog pins. The data received by the microcontroller are simply voltage equivalent analog values, i.e., 5 V equals 1024 units. Sensor calibration of IR sensor is shown in Fig. 13. The analog value received by the sensor

Design and Implementation of an Automated Fuel Station

235

Fig. 13 IR sensor calibration

when there is no vehicle in fuel station is less than or equal to 140 units, and when vehicle is present the values obtained is greater than or equal to 865 units. The LDR sensor measures the ambient light intensity and sends the data in the form of voltage to the microcontroller through the analog pins. The data received by the microcontroller are simply voltage equivalent analog values, i.e., 5 V is equivalent to 1024 units. Sensor calibration of LDR sensor is shown in Fig. 14. The analog value received by the sensor when there is low ambient light intensity is less than or equal to 21 units, medium ambient light intensity the values are between 41 and 46 units, and under high ambient light intensity the values obtained is greater than or equal to 130 units.

3.5 Automatic Fire Extinguisher The analog pins on the microcontroller receive various readings from the sensor, which are voltage equivalent analog values, i.e., 5 V equals 1024 units or 0.0049 V (4.9 mV) per unit. The calibration graph of the flame sensor is shown in Fig. 15. The analog value received by the sensor when there is no flame detected is 1023 units, and when there is a flame detected, the analog value obtained is less than or equal to 33 units. As illustrated in Fig. 16, the smoke sensor gets rapidly shifting analog readings. The smoke is recognized when the resultant analog value exceeds the threshold value of 400 units.

236

M. Jyothirmayi et al.

Fig. 14 LDR sensor calibration

Fig. 15 Flame sensor calibration

4 Conclusion In this work, it is possible to operate all the systems conveniently 24/7 with reduced manpower and also the fuel is secured and malpractices can be avoided. The lighting system of whole fuel station is made completely automatic. Automatic Vehicle washing system saves the time of the customer. The level of the fuel in the storage

Design and Implementation of an Automated Fuel Station

237

Fig. 16 Smoke sensor calibration

tank is monitored and indicated automatically. In case of fire in the fuel station, automatic fire extinguisher reacts immediately. Fire extinguishing system is also a unique design that was used to extinguish the fire. This paper attempts to integrate different systems that work as a single system that is ideal for a complete automation system. The advancement of technology leads everything to get upgraded which led to this idea.

References 1. Janani G (2018) Petrol bunk automation with prepaid card using GSM identification. Int J Res Appl Sci Eng Technol (IJRASET) 2. Baqir ZM, Motlak HJ (2021) Smart automatic petrol pump system based on internet of things. Int J Electr Comput Eng (IJECE) 3. Rai N, Jagtap A, Rajguru PD (2020) Digital fuel level and battery life indicator. Int Eng Res J (IERJ) 3 4. Sunmonu RS, Sodunke MA, Abdulai OS, Agboola EA (2018) Development of an ultrasonic sensor based water level indicator with pump switching technique. Int J Res Electron Electr Eng 5. Singh RD, Nigam S, Aggrawal S, Neelgar MR, Kaura S, Sharma K (2018) Design and implementation of automatic car washing system using PLC. Int Res J Eng Technol (IRJET) 05(05) 6. Lin KK (2019) Arduino based automatic car washing system. Int J Trend Sci Res Dev 7. Fujii Y, Yoshiura N, Takita A, Ohta N (2013) Smart street light system with energy saving function based on the sensor network. Grant-in-aid for scientific research 8. Khachane MY (2018) Intelligent street lighting system. Int J Eng Res Comput Sci Eng (IJERCSE)

238

M. Jyothirmayi et al.

9. John AJ, Ashik K, Avinash Vishnu KS, Fahmi P, Henna P (2016) Automatic fire extinguishing robotic vehicle. Int J Sci Eng Res 10. Dey P, Das S, Das C, Ahmed T, Reza CMFS, Rahman M (2017) Design and implementation of an automatic fire extinguishing system based on fault secure multi-detectors. In: International conference on mechanical engineering and renewable energy

Heart Disease Prediction Using Machine Learning Algorithms Rea Mammen and Arti Pawar

Abstract Heart disease is synonymous with heart attacks and strokes. But, cardiovascular disease also includes maladies like coronary artery disease (CAD), heart arrhythmias, hypertension, congenital heart disease, etc. Heart disease plagues a majority of the population today and is the leading cause of death globally. Efficient prediction systems to diagnose heart diseases are a must in the health care industry. Such systems are already in use but there is scope for improvement and with technological advancement over the years, the accuracy of disease prediction has been improved. Machine learning is a branch of artificial intelligence that predicts several naturally occurring events by training a model with some data and then using unseen data to test it. This paper seeks to analyze a few machine learning algorithms and tests their accuracy in predicting heart diseases. Keywords Machine learning algorithms · Coronary artery disease · Logistic regression · SVM · Decision tree · Naïve Bayes

1 Introduction Heart disease patients experience different symptoms like shortness of breath, dizziness, chest pain, or discomfort, etc. Predicting the heart failure risk of patients suffering from these symptoms is vital in minimizing the risk of peracute heart ailments, as an early diagnosis can aid the afflicted in seeking immediate health care. The dearth of proper health care facilities/personnel in many areas and the increasing population may lead to delayed diagnosis of heart diseases. Hence, computational methods can help overcome these issues by providing a timely diagnosis. Machine learning algorithms play a significant role in predicting the presence of heart diseases R. Mammen · A. Pawar (B) ICAS, Manipal Academy of Higher Education, Manipal 576104, India e-mail: [email protected] R. Mammen e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_16

239

240

R. Mammen and A. Pawar

in a patient. This paper examines the accuracy of a few machine learning algorithms, namely, support vector machine (SVM), logistic regression, Naïve Bayes, and decision tree.

2 Background 2.1 Literature Survey Machine learning techniques are helpful to detect heart disease in patients based on their medical records. Different machine learning algorithms like SVM, Naïve Bayes, decision tree, and KNN, applied on a heart disease Framingham dataset, are analyzed. Using the ROC curve and computing the area under the curve, the accuracy of these algorithms is calculated and tabulated. KNN algorithm scored the highest accuracy (83.60%), followed by Naïve Bayes (80.66%), decision tree (75.58%), and lastly, SVM (65.56%) [1]. WEKA, an open-source software containing machine learning algorithms for data mining tasks, is used for comparing the accuracy of different decision tree classification algorithms. Logistic model tree, J48, and random forest algorithms are the classification tree algorithms under analysis. The Cleveland database in the UCI repository provides the required heart disease dataset. The J48 algorithm showed the best overall performance as it resulted in high accuracy (56.76%) and less build time. In contrast, the logistic model tree algorithm showed the lowest accuracy (55.77%) and a build time of 0.39 s [2]. SVM, ANN, KNN, logistic regression, classification tree, and Naïve Bayes are implemented and applied on the StatLog heart disease data set. The study involved partitioning the given data samples in tenfold with each fold being used exactly once as the validation data for testing and the remaining folds used as the training data. Logistic regression proved to have the highest classification accuracy of about 85%, followed by ANN with an accuracy of 84%. SVM emerged as the most suitable for detecting the absence of heart disease in patients. It has a high specificity of around 89%, i.e., it is good in identifying true negatives. Logistic regression and ANN have high sensitivities (89% and 87%, respectively), which means that they are good at identifying true positives, i.e., people with heart disease [3]. The expanding data analytics industry increases the necessity of more accurate disease prediction systems, especially cardiovascular disease prediction. Efficient techniques to forecast such fatal diseases need to be implemented to help detect and treat these diseases well in time. Hence, this study involves studying machine learning algorithms like Naïve Bayes, random forest, logistic regression, decision tree, SVM, and KNN, applied on the heart disease dataset taken from the UCI Repository. Analyzing the comparison of their performance accuracies helps to see which methodology performs best in accurately classifying the presence or absence of heart disease in a patient. After pre-processing, the data was split into the training set (70%)

Heart Disease Prediction Using Machine Learning Algorithms

241

and the testing set (30%). Random forest was the most accurate (91.80%), followed by Naïve Bayes (88.52%), decision tree (76.68%), SVM (88.52%), and lastly, logistic regression and KNN with an accuracy of 86.88% each. The study also shows more people afflicted with heart disease in the 50–60 age bracket [4]. SVM, a supervised machine learning algorithm, is used to predict cardiovascular disease in a patient. The model considers factors like age, sex, and pulse rate, etc., to predict the disease. The SVM algorithm results in greater prediction accuracy when compared to other machine learning algorithms. It has an average precision of around 91%. SVM accuracy proved to be 85.97%. It has high sensitivity (90.10%) and specificity (77.20%), proving it to be a good model overall compared to other models [5]. The proposition of a hybrid method for coronary artery disease prediction seeks to find an alternative to angiography. With the help of the genetic algorithm, the hybrid method improves the performance of neural networks by approximately 10%. The feature selection process involved popular ranking methods like Gini Index, weight by SVM, information gain, and principal component analysis. The proposed method tested on coronary artery disease (CAD) patients of the Z-Alizadeh Sani dataset achieved an accuracy of 93.85%, sensitivity of 97%, and specificity of 92% [6]. Supervised machine learning algorithms (Bayesian networks, logistic regression, decision trees, random forests, SVM, k-NN, and neural networks) are applied on the Titanic data set obtained from Kaggle. Factors such as accuracy, comprehensibility, and computation time form the basis of comparison among the supervised machine learning algorithms under analysis. The mean replaces missing values in the dataset if the data is continuous and the mode if the data is categorical. It is noteworthy that treebased algorithms like decision trees and random Forests, and discriminant analysis outperformed the other techniques. Combining two or more efficient algorithms to form a multimodal system can improve the accuracy of the predictive model [7]. In a PCA-SVM hybrid model, PCA is used to extract features and reduce the dimensionality of the dataset, and SVM is successively employed for classification. This hybrid model is effective in the early diagnosis and risk assessment of breast cancer. The model presents an accuracy of 97.62%, a sensitivity of 95.24%, and a specificity of 100% when tested on the dataset obtained from Lagos State University Teaching Hospital (LASUTH). These improved values confirm the viability of this hybrid model approach [8]. The C4.5 algorithm, k-NN, SVM, and Naïve Bayes applied on the Wisconsin Breast Cancer datasets show that SVM has the highest performance accuracy (97.13%) and lowest error rate in breast cancer risk prediction and diagnosis. The comparison drawn among the algorithms mentioned above is related to the accuracy, sensitivity, specificity, and precision in classifying the breast cancer data set. Both C4.5 and k-NN resulted in the highest error rate and consequently led to many incorrectly classified classes [9]. The heart disease dataset tests the performance of three Bayes classifiers, namely, Naïve Bayes, Bayes Net, and Naïve Bayes Multinomial. This performance estimation utilizes the cross-validation parameter. The experimental results indicate that the Naïve Bayes algorithm has the highest performance or classification accuracy and

242

R. Mammen and A. Pawar

the lowest error rate than Bayes Net and the Naïve Bayes Multinomial algorithms. The Naïve Bayes algorithm correctly classified 83.7037% of the instances, while the Bayes Net method correctly classified 83.3333%. On the other hand, the Naïve Bayes Multinomial algorithm correctly classified only 73.7037% of the cases and incorrectly classed nearly 26.30% of the cases [10]. Based on factors like sex, age, BMI, pulse rate, etc., an application to predict the presence of heart disease in a patient is developed using neural networks because of its accuracy and reliability. The dataset used in this study is the Cleveland heart disease dataset from the UCI repository. The proposed system uses the multi-layer perceptron (MLP) algorithm, a type of neural network, to predict the presence of CAD in an individual before its onset. The system obtained an overall precision of 91% and predicts if a person will have CAD in the future [11]. Machine learning algorithms k-NN, decision trees, linear regression, and SVM are applied to the heart disease dataset obtained from the UCI repository to compare their accuracies. About 73% of the collected data was used as the training set, and 37% as the testing set. The performance accuracy of k-NN turns out to be the best among the four algorithms under comparison, with an accuracy of 87%. SVM has an accuracy of 83%, followed by decision tree (79%) and linear regression (78%). Confusion matrices support the calculation of the accuracies mentioned above [12]. Dealing with large datasets involves a certain degree of complexity. The development of an Enhanced New Dynamic Data Processing (ENDDP) algorithm helps detect, more accurately, the existence of heart disease in its early stages. This study uses the processed UCI Cleveland dataset. The proposed system, which amalgamates several Bayesian functions that are independent of each other, has higher accuracy (97.98%), sensitivity (97.45%), and specificity (98.54%) as compared to the Naïve Bayes and Random Forest algorithms [13]. Analysis of different algorithms like logistic regression, k-NN, decision tree, random forest classifier, and SVM helps identify the appropriate technique to build a heart disease prediction system. This system could, in turn, aid the healthcare industry make predictions with a low error rate and high accuracy. The collected data was split into a training and testing set, with 90% of the data used for training and 10% for testing. SVM and the Random Forest classifier received a testing accuracy of 90.32% each, which proved to be the best among the algorithms under analysis. K-NN acquired the least testing accuracy of the lot (70.96%) [14]. Analysis of algorithms like k-NN, K-means clustering, neural networks, Naïve Bayes, logistic regression, and hybrid techniques like K-means clustering with Naïve Bayes classifier or Fuzzy k-NN help determine which of these machine learning algorithms would be the most helpful in predicting the presence of cardiovascular disease in a person. These algorithms are applied and tested on the Cleveland heart disease dataset. Experimental results show that hybrid approaches such as neural networks and Fuzzy k-NN performed better than the other algorithms, with neural networks resulting in an accuracy of 98% and Fuzzy k-NN with an accuracy of 94.19% [15]. Diabetes is a condition/disease in which the blood sugar levels rise above the ordinary level, and if not treated in time, it can prove fatal. Machine learning can

Heart Disease Prediction Using Machine Learning Algorithms

243

help with the early prediction of this disease. Naïve Bayes, k-NN, decision tree classifier, SVM, and random forest are the five supervised machine learning algorithms applied on the Pima Indians Diabetes dataset. The experiment involves five phases: Dataset collection, data preprocessing, setting classification metrics, applying ML classifiers, and K-fold cross validation. The dataset includes eight features. In the data preprocessing step, the missing values in the dataset are filled in with the calculated standard deviation of that particular feature. Precision, recall, F1 score, and accuracy are the classification metrics used. After performing K-fold cross-validation (K = 10 in this case), the k-NN classifier emerged as the most accurate algorithm, achieving an accuracy of 76%. SVM, Naïve Bayes, decision tree, and random forest also achieved accuracies above 70%, but k-NN received the highest [16]. Data mining and machine learning help analyze large data sets and can hence be used in the medical field to predict diseases like heart disease. To perform a comparative analysis to conclude which algorithm would be the best choice for a prediction system, a study of supervised machine learning models like Naïve Bayes, logistic regression, SVM, k-NN, random forest, decision tree, and ensemble machine learning technique XGBoost are implemented and tested on the Cleveland heart disease dataset. The dataset was divided into a training set (80%) and a testing set (20%). Each classifier was trained using the training set and their efficiency was tested using the testing set. Random forest obtained an accuracy of 86.89%, followed by XGBoost with 78.69%. These proved to be the most accurate, whereas k-NN achieved an accuracy of only 57.83% [17]. Improved classification and regression systems provide valuable data that help medical personnel in the early detection of diseases which, in turn, improves the chances of survival of patients. Classification algorithms are applied on three separate disease databases (heart, diabetes, and breast cancer). Any kind of missing data is replaced with the mean value (for continuous variables) or the mode value (for categorical variables). Backward elimination (p-value test) accomplishes the task of feature selection. For the breast cancer dataset, the ideal algorithm is AdaBoost, with an accuracy of 98.57%. SVM (linear kernel) achieved the highest accuracy of 85.71%, while SVM (radial basis function) attained the lowest (66.23%) when applied on the Pima Indians Diabetes dataset. Logistic regression scored the highest accuracy (87.10%) and SVM (RBF) the worst (54.84%) when applied to the heart disease dataset [18]. The Naïve Bayes classification model forms the basis of an application for predicting heart disease in a patient. Attributes such as the patient’s age, sex, BP, etc., fed to the Naïve Bayes classifier. About 80% of the collected data (obtained from the UCI repository) was used for training, and the remaining 20% for testing. An encryption algorithm (AES) is also used to securely transfer patient records to the database. The Naïve Bayesian technique of classification achieved a greater accuracy (89.77%) and lesser build time (0.01 s) in comparison to other methods like Bayes Net, MLP, and sequential minimal optimization (SMO) [19]. Diabetes is predictable with the help of supervised machine learning models like SVM and random forest. These models are tested on the dataset from Security Force Primary Health Care, Tabuk. The performance of the algorithms was deduced from

244

R. Mammen and A. Pawar

observing the ROC curve and the confusion matrix for each algorithm. SVM emerged with a prediction accuracy of 97%, while random forest performed better and scored an accuracy of 98.9% [20].

3 Dataset Description This project utilizes the Cleveland heart disease dataset, which has been taken from the UCI (UC Irvine) machine learning repository. The dataset comprises 303 rows and 14 columns or attributes (Figs. 1, 2 and 3).

4 Methodology This study analyses four machine learning algorithms: 1. 2. 3. 4.

Support vector machine (SVM) Naïve Bayes Logistic regression Decision tree.

(1) Support Vector Machine (SVM) It is a supervised machine learning technique utilized for both binary classification (predicting a categorical value) and regression (the method of predicting a continuous value) purposes. SVM works by mapping data points from a lower dimensionality feature space to a higher dimensionality feature space to categorize the data points easier. The main aim of an SVM is to find an optimal hyperplane that separates data points into two discrete classes. The plus-points of support vector machines are that they are accurate in high dimensional spaces and are memory-efficient because they use a subset of the training points in the decision function, called support vectors (these are data points closest to the hyperplane, which are responsible for building the SVM as they influence the orientation of the hyperplane). The drawbacks of this algorithm are that they are prone to overfitting when the number of features is more than the number of samples, and they are not very efficient on large data sets. The confusion matrix for the training data (Fig. 4) shows that of the 176 normal patients, 162 (92.05%) were correctly classified, and only 14 (7.95%) were incorrectly classified. Of the 66 heart disease patients, 54 (81.82%) correct classifications were made, and only 12 (18.18%) were incorrectly classified. The accuracy of the SVM for the training data set is 89.26%. In the confusion matrix for the testing data (Fig. 5), we see that of the 44 normal patients, 38 (86.36%) were correctly classified, and only 6 (13.64%) were incorrectly

Heart Disease Prediction Using Machine Learning Algorithms

Fig. 1 Brief description of the dataset attributes

245

246

R. Mammen and A. Pawar

Fig. 2 Distribution of the target variable

Fig. 3 Distribution of heart disease patients based on gender

classified. Of the 17 heart disease patients, 14 (82.35%) correct classifications were made, and only 3 (17.65%) were incorrectly classified. The accuracy of the SVM for the testing dataset is 85.26%. (2) Naïve Bayes It is a supervised machine learning technique and is “Naïve” because it assumes that each attribute/feature is independent of all others. It is a conditional probability model that assumes that for all the instances of a given class, the features bear little to no correlation with each other and that each attribute contributes equally to the outcome. The Naïve Bayes classifier is built on Bayes Theorem. Real-time applications include recommendation systems, spam filtering, and so on.

Heart Disease Prediction Using Machine Learning Algorithms

247

Fig. 4 Confusion matrix for the training data

Fig. 5 Confusion matrix for the testing data

The plus-points of Naïve Bayes classifiers are that they are suitable for highdimensional data like text since they are fast to train and use for prediction purposes. Since it is fast, it can help save computational time. It is also ideal for applications involving large datasets. Naïve Bayes assumes that each attribute/feature is independent of all others, which is not always correct in real-time problems. Another

248

R. Mammen and A. Pawar

Fig. 6 Performance accuracy of the Naïve Bayes algorithm on the training data

drawback of this Naïve Bayes is that it assigns a zero probability to a categorical variable present in the test dataset but absent in the training set. The confusion matrix for the training data (Fig. 6) shows that of the 176 normal patients, 158 (89.77%) were correctly classified, and only 18 (10.23%) were incorrectly classified. Of the 66 heart disease patients, 57 (86.36%) correct classifications were made, and only 9 (13.64%) were incorrectly classified. The accuracy of Naïve Bayes for the training data set is 88.84%. In the confusion matrix for the testing data, we see that of the 44 normal patients, 34 (77.27%) were correctly classified, and only 10 (22.73%) were incorrectly classified. Of the 17 heart disease patients, 14 (82.35%) correct classifications were made, and only 3 (17.65%) were incorrectly classified. The accuracy of Naïve Bayes for the testing data set is 78.69% (Fig. 7). (3) Logistic Regression A classification technique for categorical variables, logistic regression requires the explanatory variables to be continuous. In this machine learning method, one or more independent variables predict the target/dependent variable. The logistic regression model is useful when data is binary, linearly separable, and there is a requirement for probabilistic results. Logistic regression is identical to linear regression, but instead of predicting a discrete value, it predicts a categorical value. It cannot predict nonlinearly separable data. The confusion matrix for the training data (Fig. 8) shows that of the 176 normal patients, 164 (93.18%) were correctly classified, and only 12 (6.82%) were incorrectly classified. Of the 66 heart disease patients, 49 (74.24%) correct classifications were made, and only 17 (25.76%) were incorrectly classified. The accuracy of logistic regression for the training data set is 88.02%.

Heart Disease Prediction Using Machine Learning Algorithms

249

Fig. 7 Performance accuracy of the Naïve Bayes algorithm on the testing data

Fig. 8 Representation of the number of correct and incorrect predictions made by the logistic regression algorithm on the training dataset

In the confusion matrix for the testing data (Fig. 9), we see that of the 44 normal patients, 39 (88.64%) were correctly classified, and only 5 (11.36%) were incorrectly classified. Of the 17 heart disease patients, 14 (82.35%) correct classifications were made, and only 3 (17.65%) were incorrectly classified. The accuracy of logistic regression for the testing dataset is 86.89%. (4) Decision Trees As the name suggests, decision trees are “tree-shaped” algorithms/diagrams that lay down a course of action. Each branch of the tree indicates a decision based on a test. A

250

R. Mammen and A. Pawar

Fig. 9 Representation of the number of correct and incorrect predictions made by the logistic regression algorithm on the testing dataset

decision tree comprises a root node, internal nodes, and leaf nodes. They are used for classification and regression purposes. The training dataset is input to the decision tree, from where it splits into nodes. Hence, a feature is tested, and subsequent branching occurs depending on the test outcome. Decision trees are relatively easy to decipher and conceptualize. Moreover, they can be used with both numeric and categorical data. However, decision trees are susceptible to overfitting. The confusion matrix for the training data (Fig. 10) shows that of the 176 normal patients, 176 (100%) were correctly classified, and none were incorrectly classified. Of the 66 heart disease patients, 66 (100%) correct classifications were made, and none were incorrectly classified. The accuracy of the decision tree algorithm for the training data set is 100%. In the confusion matrix for the testing data (Fig. 11), we see that of the 44 normal patients, 33 (75%) were correctly classified, and only 11 (25%) were incorrectly classified. Of the 17 heart disease patients, 10 (58.82%) correct classifications were made, and only 7 (41.18%) were incorrectly classified. The accuracy of the decision tree algorithm for the testing data set is 70.49%.

5 Results We measured the testing and training accuracies for each of the algorithms implemented to examine if the problems of underfitting or overfitting occur. If the training accuracy is very high, the model is facing the problem of overfitting. On the other hand, a high testing accuracy indicates model underfitting. Therefore, an ideal classifier is one with comparable, testing, and training accuracies. For the Naïve Bayes and the decision classifiers, the training and testing accuracies differ by over 10%,

Heart Disease Prediction Using Machine Learning Algorithms

251

Fig. 10 Confusion matrix for the decision tree algorithm applied on the training dataset

Fig. 11 Confusion matrix for the decision tree algorithm applied on the testing data

with the training accuracy being much higher, indicating model overfitting. SVM performed reasonably well. Logistic regression performed the best, with a testing accuracy of 86.89% (Table 1).

252 Table 1 Training and testing accuracies

R. Mammen and A. Pawar Machine learning algorithm

Testing accuracy (%)

Training accuracy (%)

(1) Support vector machine

85.26

89.26

(2) Naïve Bayes

78.69

88.84

(3) Logistic regression

86.89

88.02

(4) Decision tree

70.49

100

6 Conclusion Machine learning can improve the accuracy of disease prediction exponentially. Early disease detection is vital to improve the chances of survival of a person who would otherwise succumb to a disease due to a delayed diagnosis. Algorithms like SVM, Logistic Regression, Decision Trees, and Naïve Bayes help build effective prediction systems like a heart disease prediction system. Cardiovascular diseases can affect all age groups and, if not detected in time, can be fatal. In this study, Logistic Regression has proven to be an ideal algorithm for a heart disease prediction system owing to its high accuracy of 86.89%. Future studies will address the efficiency and accuracy of these machine learning algorithms on other data sets, such as the Wisconsin Breast Cancer dataset. Applying the same machine learning techniques on different datasets will help formulate an interesting comparative analysis among the different datasets used. Principal Component Analysis (PCA), a dimensionality reduction algorithm, can also be applied to large datasets to reduce the complexity level. By reducing the feature space of large datasets, analysis of the features is more effective.

References 1. Marimuthu M, Deivarani S, Gayathri R (2019) Analysis of heart disease prediction using various machine learning techniques. In: Advances in computerized analysis in clinical and medical imaging, pp 157–168 2. Patel J, Upadhyay T, Patel S (2016) Heart disease prediction using machine learning and data mining technique. Int J Comput Sci Commun 7(1):129–137 3. Dwivedi A (2016) Performance evaluation of different machine learning techniques for prediction of heart disease. Neural Comput Appl 29(10):685–693 4. Malavika G (2020) Heart disease prediction using machine learning algorithms. Biosci Biotechnol Res Commun 13(11):24–27 5. Sandhya Y (2020) Prediction of heart diseases using support vector machine. Int J Res Appl Sci Eng Technol 8(2):126–135 6. Arabasadi Z, Alizadehsani R, Roshanzamir M, Moosaei H, Yarifard A (2017) Computer-aided decision-making for heart disease detection using hybrid neural network-genetic algorithm. Comput Methods Programs Biomed 141:19–26

Heart Disease Prediction Using Machine Learning Algorithms

253

7. Singh A, Thakur N, Sharma A (2016) A review of supervised machine learning algorithms. In: 2016 3rd international conference on computing for sustainable global development (INDIACom), pp 1310–1315 8. Akinnuwesi B, Macaulay B, Aribisala B (2020) Breast cancer risk assessment and early diagnosis using principal component analysis and support vector machine techniques. Inform Med Unlocked 21:100459 9. Asri H, Mousannif H, Moatassime H, Noel T (2016) Using machine learning algorithms for breast cancer risk prediction and diagnosis. Procedia Comput Sci 83:1064–1069 10. Nithya R, Ramyachitra D, Manikandan P (2016) An efficient Bayes classifiers algorithm on 10-fold cross-validation for heart disease dataset 11. Gavhane A, Kokkula G, Pandya I, Devadkar K (2018) Prediction of heart disease using machine learning. In: 2018 second international conference on electronics, communication, and aerospace technology (ICECA), pp 1275–1278. https://doi.org/10.1109/ICECA.2018.847 4922 12. Singh A, Kumar R (2020) Heart disease prediction using machine learning algorithms. In: 2020 international conference on electrical and electronics engineering (ICE3), pp 452–457. https://doi.org/10.1109/ICE348803.2020.9122958 13. Rao J, Prasad R (2021) An enhanced novel dynamic data processing (ENDDP) algorithm for predicting heart disease in machine learning. Int J Sci Res Comput Sci Eng Inf Technol 94–104 14. Bhunia PK, Debnath A, Mondal P, Monalisa DE, Ganguly K, Rakshit P (2021) Heart disease prediction using machine learning. Int J Eng Res Technol (IJERT) NCETER—2021 09(11) 15. Yadav SS, Jadhav SM, Nagrale S, Patil N (2020) Application of machine learning for the detection of heart disease. In: 2020 2nd international conference on innovative mechanisms for industry applications (ICIMIA), pp 165–172. https://doi.org/10.1109/ICIMIA48430.2020. 9074954 16. Lyngdoh AC, Choudhury NA, Moulik S (2021) Diabetes disease prediction using machine learning algorithms. In: 2020 IEEE-EMBS conference on biomedical engineering and sciences (IECBES), pp 517–521. https://doi.org/10.1109/IECBES48179.2021.9398759 17. Anbuselvan P (2020) Heart disease prediction using machine learning techniques. Int J Eng Res Technol (IJERT) 09(11) 18. Kohli PS, Arora S (2018) Application of machine learning in disease prediction. In: 2018 4th international conference on computing communication and automation (ICCCA), pp 1–4. https://doi.org/10.1109/CCAA.2018.8777449 19. Repaka AN, Ravikanti SD, Franklin RG (2019) Design and implementing heart disease prediction using Naïve Bayesian. In: 2019 3rd international conference on trends in electronics and informatics (ICOEI), pp 292–297. https://doi.org/10.1109/ICOEI.2019.8862604 20. Alanazi AS, Mezher MA (2020) Using machine learning algorithms for prediction of diabetes mellitus. In: 2020 international conference on computing and information technology (ICCIT1441), pp 1–3. https://doi.org/10.1109/ICCIT-144147971.2020.9213708

Design and Performance Evaluation of a Simple Resistance-to-Digital Converter for Tunneling Magneto-Resistance-Based Angular Position Sensor with 180° Range Kishor Bhaskarrao Nandapurkar Abstract Numerous applications require angular position measurement of a rotating shaft up to 180° range. Some examples include engine throttle-valve, joint angle of robotic arms, skid loaders, etc. Angle sensors based on tunneling magneto-resistance (TMR) effect can be used for many such applications due to their features such as compact and cost-effective design, low powered operation, and high sensitivity. The resistances of TMR angle sensor vary as sine and cosine function of the shaft angle. In this paper, a simple Resistance-to-Digital Converter for TMR angle sensor (RDCT) is presented. The proposed RDCT processes the nonlinear natured TMR resistances and render a linear digital output for 180° range. The digital indication of the shaft angular position is achieved without using any explicit analog-to-digital converter (ADC) and look-up table (LUT). The RDCT is based on standard dual-slope principle and hence possesses all the merits of dual-slope-based measurement method. It can also be employed to detect angular position of the target shaft rotating between 0 and 180° at slow speeds. The working principle of the RDCT is explained first in a detailed manner. Later, the effects of practical parameters of TMR angle sensor and the RDCT circuitry are analyzed. An effective offset-error reduction method is then discussed that reduces the output error due to the mismatch in TMR resistances. Extensive simulation studies are carried out to quantify the RDCT performance under ideal as well as non-ideal conditions. The effectiveness of the offset-reduction method is also proved in simulation studies. Keywords Angle sensor · Dual-slope-based technique · Error analysis · Resistance-to-digital converter · Tunneling magneto-resistance

K. B. Nandapurkar (B) Indian Institute of Technology (Indian School of Mines) Dhanbad, Dhanbad, Jharkhand, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_17

255

256

K. B. Nandapurkar

1 Introduction Angular position of a rotating shaft is a variable of interest for fields such as manufacturing and process industries, automobiles, medicine, military, aerospace, and several other. Some of the specific applications that require knowledge of angular position of the target shaft include detection of position of knobs, rotary valves, solar panels, vehicle engine throttle-valve, and motor shaft [1, 2]. In many of these applications, the target shaft rotates for angles up to 180° [3, 4]. Different types of angle sensors are commercially available that can be employed in such applications. These sensors can be broadly categorized as resistive potentiometers, optical, capacitive, inductive, hall-effect-based encoders [5]. Angle sensors based on tunneling magneto-resistance (TMR) effect are relatively new in the market, and they have many advantages over the aforementioned (conventional) angle sensors. The TMR angle sensors, as opposed to resistive potentiometers, offer non-contact measurement. In contrast to RVDTs, they are compact in design and consume less power. The output sensitivity of TMR angle sensors is high when compared with hall-effect-based encoders [5]. The availability of TMR angle sensors at low price, together with the advantages discussed above, makes them a favorable choice for shaft angle sensing. The TMR-based angle sensors yield two outputs which possess sine and cosine function dependency on the shaft angular position. It is to be interestingly noted that resolvers and hall-effect based encoders also possess such sine–cosine transfer characteristic [1, 5]. It is because of the two quadrature-shifted sine–cosine outputs, and shaft angles up to 360° range can be accurately estimated. However, such sine–cosine response of the sensor is nonlinear. So, despite having several merits, such nonlinear response sometimes makes an end user give a second thought on the selection of TMR angle sensor for the intended application. This issue can be easily addressed by designing a suitable interfacing circuit for the TMR angle sensor that linearizes the sensor response over the ‘desired measurement range’. It is to be noted that the phrase ‘desired measurement range’ is very wisely used in the previous statement. This is because the complexity of the interfacing circuit can be significantly reduced if the measurement range is reduced to 180° from 360°, which, as mentioned earlier, is sufficient for many applications. Some of these applications cover joint angle of robotic arms, skid loaders, solar panels, operating tables and dental chairs in hospitals, etc. [3, 4, 6, 7]. Researchers have designed several linearizing interface circuits for TMR-like sine–cosine angle sensors [8–13]. The features of these circuits are briefed next. Lopez-Martin and Carlosena have discussed three analog and digital interfacing circuits for GMR-based angle sensor in [8]. The maximum operating range achieved in this work is 10–170°. These circuits are complex as they either use programmable gain amplifier, encoder, multiplexers, and/or dedicated ADCs. A simple ADC-based digitizer for GMR angle sensor is presented in [9]. However, the angular resolution of the circuit is 1° which is quite poor. An interface circuit which requires flash ADC with nonlinear and non-standard resistors has been designed for MR angle sensor in [10]. Anoop and George have proposed an electronic scheme that estimates the shaft angle from the outputs of MR angle sensor using inverse-cosine operation.

Design and Performance Evaluation of a Simple Resistance-to-Digital …

257

This circuit requires a costly instrumentation amplifier and a sine wave oscillator for linearization and provides analog output [11]. A resolver-to-DC converter has been developed and discussed by Benammar et al. [12]. This circuit is significantly complex and employs costly analog multipliers for linearization. The directdigital interfacing circuits, described in [13, 14], also require a sine wave/quadrature oscillator and instrumentation amplifier. Thus, from the above brief discussion, it can be concluded that an efficient interfacing circuit is essential for TMR-like sine–cosine natured angle sensors. Considering the increasing demand for the digital measurement systems, the interfacing circuit should render a linear and digital output, preferably, without using any dedicated ADC(s) and LUTs. Such a digitizer should be simple-to-design, utilizing lowcost electronic components. The use of costly components such as instrumentation amplifier and analog multiplier should be discouraged. In this paper, a simple Resistance-to-Digital Converter for TMR angle sensor (RDCT) is presented that achieves all the aforementioned features. The proposed RDCT is a simplified version of the linearizing signal conditioner (LSC), presented in [15]. Similar to the LSC, the RDCT employs a DC excitation for TMR angle sensor, and hence, the parasitic capacitances of the sensor do not affect the measurement. The merits of the RDCT over [15] are highlighted below. 1. The LSC processes both sine and cosine half-bridges of the TMR sensor. On the other hand, the RDCT processes only cosine half-bridge and renders a linear output for half-circle range. As clarified earlier, 180° range of measurement is sufficient for many applications. 2. The LSC requires two integration phases and two de-integration phases for angle estimation, and hence, it is slower in operation. The proposed RDCT, however, requires one integration and one de-integration phase for the angle measurement. Since the RDCT measurement time is faster, intuitively, it can track dynamic inputs with a faster speed compared with the LSC. 3. The RDCT requires one switch and one control signal less compared to that of LSC. The next section covers the basic principle of the TMR angle sensor and the working of the proposed RDCT. In the subsequent sections, factors affecting the RDCT performance and the verification of RDCT working are covered.

2 Resistance-to-Digital Converter for TMR Angle Sensor The working principle of TMR angle sensor is explained first before heading toward the RDCT.

258

K. B. Nandapurkar

2.1 TMR Angle Sensor—Construction, Operating Principle, and Transfer Characteristic The TMR-based shaft angle sensing setup is shown in Fig. 1a. This sensing arrangement demands a diametrically magnetized permanent magnet (PM) to be attached to the rotating shaft. Such a PM produces magnetic field whose plane will be in parallel to the plane of the PM. The TMR angle sensor, which is available in a compact IC form [16], is placed below the PM. The TMR angle sensor requires saturating magnetic field of specific values for its proper operation. The range of this field depends on the manufacturer. For an instance, the AAT00X series of half-bridge TMR angle sensor, from NVE Corporation, requires magnetic field between 30 and 200 G [16]. This can be easily achieved by carefully adjusting the air gap between the TMR IC and the PM. Once the PM and the TMR IC are positioned properly, the sensing setup is ready to be used for angle measurement purpose. From the figure and the above explanation, it is clear that the TMR-based sensing arrangement is non-contact type and, hence, offer long life sensing solution. It is to be noted that this sensing arrangement need not be vertical, as shown in Fig. 1a. In applications such as motor shaft positioning, camera surveillance, the target shaft may be orientated in any direction. The TMR-based arrangement can be used in such scenarios as well. Commercially available TMR angle sensors are categorized as full-bridge type and half-bridge type. The full-bridge TMR sensors have eight ‘active’ TMR elements, embedded in them. On the other hand, as shown in Fig. 1b, a half-bridge-type TMR angle sensor internally consists of four active resistances. The word active in the previous two statements signifies that the resistance varies with the change in the direction of the biasing magnetic field of the PM. The proposed RDCT essentially requires a part of half-bridge TMR angle sensor for accurate shaft angle measurement.

Fig. 1 a TMR-based shaft angle sensing setup. b Internal circuit diagram of a half-bridge TMR angle sensor, comprising four TMR elements: RC1 , RC2 , RS1 , and RS2 . c TMR resistances vs shaft angle characteristic. Four TMR resistances have sine–cosine natured relationship with the shaft angle, θ

Design and Performance Evaluation of a Simple Resistance-to-Digital …

259

As shown in Fig. 1b, the resistances RS1 , RS2 , RC1 , and RC2 form half-bridge TMR angle sensor. To understand the working principle of TMR angle sensor, let us first discuss the structure of individual TMR resistances. A single TMR element (say, RS1 ) is a three-layered structure. The top layer is formed using a ferromagnetic material, and hence, its magnetization can align with the direction of magnetic field of the PM. The bottom layer is commonly called as pinned layer due to the fact that its magnetization is fixed or pinned in a particular direction [16]. These two layers are separated by an ultra-thin insulator layer (usually made of Al2 O3 or MgO). The nominal device resistance is mainly governed by the thickness of this insulator layer. For angle sensing applications, as shown in Fig. 1a, when the shaft rotates by some angle (say, θ ), the PM also rotates by the same angle. Thus, there is a rotation of the magnetic field ‘seen’ by the TMR sensor. The latter is powered using AC/DC voltage or current source of suitable magnitude. The resistance of the TMR element changes when the direction of the magnetic field changes. This is because of the change in the relative angle between the magnetization of free layer and pinned layer. For an instance, when the magnetizations of both the layers are parallel, then electrons are easily tunneled. This results in minimum resistance. On the contrary, when the two layers have their respective magnetization antiparallel, the device resistance is the largest. Since the resistance varies with the external magnetic field, this effect is known as ‘magneto-resistance’ effect. In particular to the TMR sensors, due to the presence of ultra-thin insulating layer, the electrons literally tunnel from the bottom layer to the top layer (or, vice versa). This is the reason of calling this overall phenomenon as ‘tunneling magneto-resistance’. More details on the fabrication and working of the TMR angle sensors can be found in [16–18]. The physical arrangement of four TMR elements RS1 , RS2 , RC1 , and RC2 , within the sensor, is different, and hence, their response to the change in the direction of magnetic field is also different. The mathematical expressions of these resistances are given in (1). RS1 = RN − ΔR sin θ ; RS2 = RN + ΔR sin θ ; RC1 = RN − ΔR cos θ ; RC2 = RN + ΔR cos θ

(1)

In the above expressions, RN stands for the nominal resistance of the TMR element and ΔR is the maximum change in the device resistance. Typically, RN is significantly greater than ΔR. For an ideal TMR angle sensor, the value of RN for all four resistances is equal. Also, ΔR is same for all four elements. Figure 1c shows the nature of variation of resistances of half-bridge TMR angle sensor with respect to the shaft angle, θ. From (1) and the graph, it is clear that the resistances possess unique sine and cosine relationship with θ. Of these four resistances, RS1 and RS2 form half-bridge HBS and RC1 and RC2 form HBC . This can be better visualized in Fig. 1b. From (1), we can also say that RS1 and RS2 (and RC1 and RC2 ) are arranged in push–pull configuration. Such arrangement keeps the total half-bridge resistance (RS1 + RS2 and RC1 + RC2 ) constant and independent of the sensing angle, θ.

260

K. B. Nandapurkar

From the sensor characteristic, shown in Fig. 1c, it is straightforward that the resistances RC1 and RC2 of HBC are unique for every value of θ ∈ (0◦ , 180◦ ). It is due to this property, the shaft angle, θ, can be accurately determined if just the RC1 and RC2 are processed using a suitable interfacing circuit. The proposed RDCT achieves this objective. Please note that RS1 and RS2 of HBS have distinct values for θ ∈ (90◦ , 270◦ ). The RDCT can process these resistances as well to determine θ. However, an offset of 90° is required in the sensing unit, i.e., between the sensor and the PM to render a linear indication from 0 to 180°. Since the measurement accuracy is independent of choice of resistances, HBC is chosen to avoid need of mechanical offset. Let us now slightly digress from the concept of processing the resistance variations to measure the shaft angle. The half-bridge-based TMR angle sensor, when powered using a suitable excitation voltage (say, V EX ), renders two sine–cosine natured outputs. These outputs (say, vSIN and vCOS ) can be given as vSIN = (0.5 + k sin θ ) × VEX and vCOS = (0.5 + k cos θ ) × VEX . Here, the term k = ΔR/(2RN ) refers to the transformation constant of the TMR sensor. The typical value of k is 80 mV/V. Thus, we can say that the half-bridge sensor outputs contain an offset term (= 0.5V EX ) which is significantly larger than the modulating term ‘k cos θ ’ or ‘k sin θ ’. In order to extract a linear output from such sine–cosine natured voltages, the offset term has to be eliminated. Later, the offset-free signals can be suitably processed by ADC and a processor for linearization and digitization. However, please note that the offset-free signals are small in magnitude and bipolar in nature. The strength of the signals can be enhanced using additional amplifier stage. However, the ADC must be bipolar to accept such signals or suitable demodulation stage is required. Considering factors of the ADC such as resolution, gain and offset error, and differential and integral nonlinearity, the ADC-based angle measurement system may not be the best and cost-effective choice. Therefore, an alternative to such ADC-based digital angle measurement system must be searched. The proposed RDCT targets this particular issue efficiently. Having discussed about the TMR angle sensor in detail and the limitations of the ADC-based measurement, let us now discuss the working principle of the proposed RDCT in an elaborated manner.

2.2 Working of RDCT The simplified schematic of the RDCT is shown in Fig. 2. As mentioned earlier, the RDCT, unlike many dual-slope-based digitizers, requires a single DC reference voltage (here, marked as V D ) for its operation. The RDCT is designed using three operational amplifiers: OP1 , OP2 , and OC, one single-pole double-throw switch SW, and a control and logic unit (CLU). Any pre-programmable microcontroller with sufficient number of digital input/output pins can be used as the CLU. The op-amp OP1 together with TMR resistances RC1 and RC2 forms an inverting amplifier. This inverting amplifier is supplied with V D as its input signal. The integrator circuit is designed using OP2 and passive components RI and C I . The selection of RI and C I

Design and Performance Evaluation of a Simple Resistance-to-Digital …

261

Fig. 2 Simplified schematic of the proposed RDCT. The half-bridge TMR sensor resistances RC1 and RC2 are processed to obtain linear digital indication for 180° range

is covered in the performance verification section. The op-amp OC is configured as a comparator to facilitate digital measurement. The components SW, integrator circuit, and comparator OC form the base for dual-slope-based measurement. As shown in Fig. 2, the inverting amplifier output, vS = −(RC2 /RC1 ) × VD is fed to position-0 of SW. The expressions of RC1 and RC2 are already given in (1). Since the value of RN is larger than ΔR, vS is negative irrespective of value of θ. The signal V D is applied at position-1 of SW. The SW position is controlled using the digital signal, DC1 , obtained from the digital port of the CLU. Note that the signal vS acts as a measuring input for the dual-slope-based circuit and the voltage V D acts as a reference signal. A typical measurement cycle of the RDCT consists of an integration phase which runs for a preset period of time (say, T I ) and a variable de-integration phase. The time duration (say, T V ) of this time-variable phase is dependent on the input signal, vS , which in turn depends on θ. The operation of the RDCT is explained assuming that the shaft is at a steady position for at least one measurement cycle. Before the start of the measurement cycle, it is important to ensure that the integrator output, vI , is zero. This is done to avoid saturation of OP2 which will result in huge measurement error. The process of making vI = 0 is commonly called as ‘auto-zeroing’ and can be executed easily using simple CLU-based programming. The auto-zeroing step does not require additional components. Also, it is to be noted that the auto-zeroing phase is required only once after powering on the RDCT. Once vI is ensured to be zero, RDCT is ready for the measurement. Initially, the CLU sets DC1 = 0 to feed the signal vS to the integrator circuit, as shown in Fig. 3a. This signal, being constant, causes the integrator output to increase linearly with time, t as can be seen in Fig. 3b. The value of signal vI at the end of the integration phase, i.e., at time t = T I , can be expressed using (2).

262

K. B. Nandapurkar

Fig. 3 Important waveforms of the RDCT. The integrator input and output are shown in a and b, respectively. The digital output, DP , of OC is shown in c. Here two measurement cycles are shown for better understanding. The input angle can be estimated using one integration phase (T I ) and one de-integration phase (T V ) alone

) {TI ( 1 RC2 vS · dt = − × VR · dt − RI C I RC1 0 0 ) ( RN + ΔR cos θ 1 × VR × TI vI (TI ) = × RI C I RN − ΔR cos θ

1 vI (TI ) = − RI C I

{TI

(2)

Once the integration phase is over, the CLU executes the de-integration phase. For this purpose, it places SW at position-1 by making DC1 = 1. The reference signal V D is fed to the integrator. The output of latter, as shown in Fig. 3b, starts falling linearly. The signal vI is supplied to the non-inverting terminal of OC. The latter’s inverting terminal is grounded. In the absence of feedback, for such a connection, OC acts as zero-crossing detector. Thus, the OC output, DP , is digital in nature. For vI ≥ 0, DP = 1, otherwise 0. In the de-integration phase, when the decreasing vI crosses zero (vide Fig. 3b), the signal DP changes its state from logic HIGH (or, 1) to logic LOW (or, 0), as can be seen in Fig. 3c. This signal is fed to the digital input pin of the CLU. The state change of DP is detected by the CLU, and this marks the end of the de-integration phase. Note that, as shown in Fig. 3, at the end of the de-integration phase, vI = 0. Thus, the charge-balance equation can be given as per (3). 1 × RI C I

(

RN + ΔR cos θ RN − ΔR cos θ

) × VR × TI −

VR × TV = 0 RI C I

(3)

The time period T V of the de-integration phase is measured by the CLU with the help of pulse DP . The mathematical expression of T V can be obtained from (3) as per (4).

Design and Performance Evaluation of a Simple Resistance-to-Digital …

( TV =

RN + ΔR cos θ RN − ΔR cos θ

) × TI which yields

263

TV RN + ΔR cos θ = TI RN − ΔR cos θ

(4)

The CLU then performs a simple ratiometric operation as per the following expression to obtain a variable x. x=

TV − TI RN + ΔR cos θ − RN + ΔR cos θ ΔR = × cos θ = TV + TI RN + ΔR cos θ + RN − ΔR cos θ RN

(5)

Thus, it can be said that processing the time periods, as per (5), is equivalent of processing the resistances RC1 and RC2 of the TMR angle sensor. The variable x, so obtained, varies as a cosine function of θ, which is again nonlinear in nature. The simplest method of linearizing any nonlinear function is to perform inverse of the nonlinear operation. Thus, here, inverse-cosine operation is suitably performed to obtain a linear indication (say, Dθ ) of θ for θ ∈ (0◦ , 180◦ ). This transformation is defined in (6). Dθ = cos−1

(

RN ×x ΔR

)

= θ for θ ∈ (0◦ , 180◦ )

(6)

Therefore, the proposed RDCT assists to obtain a linear estimate of shaft angular position for half-circle range. Please note that instead of inverse-cosine transformation, any sufficiently accurate approximation function can also be used to fetch linear output. This selection will depend on the computation time of the CLU and the accuracy of the approximation function. As mentioned in the previous section, the RDCT does not require costly instrumentation amplifier for the offset reduction and/or linearization. It does not use dedicated ADCs and LUTs for the output digitization. The DC excitation for the sensor ensures that the output is least influenced by the parasitic capacitances of the TMR sensor. The conversion time (say, T CONV ) of RDCT is given by T CONV = T I + T V . The maximum value of T CONV can be expressed as per (7). ( max(TCONV ) = TI + TV = TI ×

2RN RN − ΔR cos θ

) θ =0◦

( = TI ×

2RN RN − ΔR

) (7)

Thus, for a given TMR sensor, T CONV can be tuned by adjusting T I alone. Please note that a large value of T I will offer very high resolution. On the other hand, lower value of T I will facilitate dynamic input measurement. Therefore, the integration time should be chosen as per the application demand and the time resolution of the CLU employed in the RDCT. Next paragraph summarizes the working of the RDCT. The RDCT, during its integration phase, processes signal vS for T I s. This signal is dependent on the values of RC1 and RC2 and hence on θ. Later, the de-integration phase is executed, and the time period T V is measured by the CLU. The ratiometric operation, as given in (5), is then performed to obtain variable x. The linear indication

264

K. B. Nandapurkar

of θ is finally obtained using the transformation defined in (6). As explained earlier, the linear response of the RDCT, obtained for half-circle range, is sufficient for many applications such as engine throttle-valve, accelerator-pedal position, joint angle of robotic arms, and skid loaders [2–7]. Therefore, the RDCT can be well used in such applications.

3 Error Analysis of the RDCT The working of the RDCT and the measurement methodology assumes that both the TMR angle sensor and the electronic components employed in RDCT are ideal in nature. However, they may not behave ideally always. The effect of non-ideal parameters of sensor and circuit components is analyzed and presented in this section.

3.1 Error Sources of TMR Angle Sensor Mismatch in ΔR of RC1 and RC2 : Let δ be the mismatch between ΔR of RC1 and RC2 such that ΔRC2 = (1 + δ) × ΔRC1 . Here, ΔRC1 and ΔRC2 are the maximal change in resistance of RC1 and RC2 , respectively. For a nonzero value of δ, the variable x gets modified to x δ as per (8). xδ = [ΔRC1 (2 + δ) cos θ ]/[2RN + δ ΔRC1 cos θ ]

(8)

Since x /= x δ , the output angle, Dθ , gets affected for a nonzero δ. The quantity Dθ was mathematically computed using (6) and (8) for every value of θ ∈ (0◦ , 180◦ ). For δ = 0.05% (or, 0.0005), the maximum error in the measurement of shaft angle is close to 0.78%, and it occurs at θ = 180°. For the range 5–175°, this error always stays within 0.1%. It is to be noted that several industrial applications do not have tight requirements on the accuracy. In such cases, the proposed low-cost, less-complex RDCT is certainly a good choice. Mismatch in RN of RC1 and RC2 : Similar to the above non-ideality, the mismatch in RN value of two resistances can be modeled as follows: RNC2 = (1 + τ ) × RNC1 . Here, τ represents the mismatch extent. The new ratiometric quantity, x τ , in this case, can be described as per (9). Here, the term ‘τ × RNC1 ’ can be considered as an undesired offset. xτ =

2ΔR cos θ + τ RNC1 (2 + τ )RN

(9)

For τ = 0.05%, the worst-case error in the measurement of θ for half-circle range is 1.5%, whereas, for 5–175° range, this error is 0.56%. Thus, this non-ideality can

Design and Performance Evaluation of a Simple Resistance-to-Digital …

265

be considered as a major error source. The effect of nonzero τ can be reduced using a simple offset-error reduction technique, described next. Let us perform a new ratiometric operation as described using the following equation. TV − (1 + τ )TI (2 + τ )ΔR cos θ = TV + (1 + τ )TI 2 × (1 + τ )RNC1 − τ × ΔR cos θ (2 + τ )ΔR ∼ × cos θ = 2 × (1 + τ )RNC1

x, =

(10)

It can be seen that the newly obtained x , does not contain the offset term ‘τ × RNC1 ’ as in (9). The quantity x , can be easily realized in the CLU as the value of τ can be found from the sensor characteristic. The maximum error in Dθ was computed using x , for τ = 0.05%, and it was 0.72% for 180° range as opposed to 1.5% obtained using x τ . For the range of 5–175°, this error has also been reduced to 0.09% from 0.56%. Thus, it can be concluded that the offset-reduction technique is notably effective to minimize the output error due to nonzero τ. It is very important to note that the mismatch extent remains constant even when the surrounding temperature fluctuates. This is because both RC1 and RC2 increase or decrease with respect to the change in temperature equally. Thus, the proposed offset-error reduction technique needs no modification when the RDCT is employed in temperature fluctuating environments. Phase imbalance in RC1 and RC2 : The resistances RC1 and RC2 with a phase imbalance (say, β) can be given as RC1 = RN − ΔR cos θ and RC2 = RN + ΔR cos (θ + β). This factor also affects the ratio x. The modified expression of x β is given in (11). xβ =

ΔR × [cos θ + cos(θ + β)] 2RN + ΔR × [cos(θ + β) − cos θ ]

(11)

The typical value of β is 0.2°, which results in the worst-case error of 0.084% for 0–180° range and 0.064% for 5–175° range. Non-sinusoidal nature of resistance variations: Apart from the above-discussed non-idealities, the TMR resistances may not follow the expected sine–cosine natured behavior. The degree of non-sinusoidality is usually quantified in terms of percentage root mean square error (RMSE) [16]. The percentage RMSE can be computed from the sensor characteristic and curve-fitting technique. The effect of such nonsinusoidal nature of TMR sensor can then be found using a suitable simulation tool. The exact reason for the non-sinusoidal behavior of the sensor may not be known. Therefore, the error arising due to this non-ideality may not be fully nullified.

266

K. B. Nandapurkar

3.2 Error Sources of Circuit Components of RDCT Offset voltage and bias current of OP1 : The offset voltage (say, vOF1 ) and bias current (say, ib1 ) of inverting amplifier op-amp OP1 will alter its output, vS , as per (12). vS,

) ( RC2 RC2 − i b1 × RC2 =− × VD + vOF1 × 1 + RC1 RC1

(12)

The typical value of vOF1 and ib1 is around 60 µV and 2.5 nA, respectively. For these values, vS , was mathematically computed for every θ ∈ (0◦ , 180◦ ). Then, the intermediate ratio x and the final output, Dθ , were calculated. The maximum error observed for 180° range is 0.62%, whereas it is 0.07% for 5–175° range. ON resistance of SW: The ON resistance of SW (say, RON ) increases the input resistance of the integrator to (RI + RON ). This modified resistance remains same for both integration and de-integration phase. Therefore, this non-ideality does not affect the variable T V and hence the RDCT output, Dθ . This fact is further established by carrying out detailed simulation studies for different values of RON . The results will be discussed in the next section. Bias current of OP2 : In the proposed RDCT, OP1 and OP2 are realized using two ICs of same manufacturer. This suggests that the bias current of OP2 is also in nanoampere range. The modified charge-balance equation for this non-ideality is given using the following expression. [(

RN + ΔR cos θ RN − ΔR cos θ

) × VR −

[ [ [ TI i b2 VR i b2 × × TV = 0 − + CI RI C I RI C I CI

(13)

For ib2 = ib1 = 2.5 nA, the intermediate ratio x and the final output Dθ are computed using the suitable equations. The maximum error spotted for this non-ideality is 0.19% for half-circle range and 0.007% for 5–175° range. Offset voltage of OC: The nonzero offset voltage (say, vOF2 ) of OC affects its zerocrossing operation. The comparator OC will change its state from logic zero to logic one (or vice versa) when its input (vI ) crosses vOF2 (instead of 0 V). Thus, this nonideality will only result in a vertical shift of integrator output waveform (vide Fig. 3b) by vOF2 . The typical value of vOF2 is in millivolts. Thus, the vertical shift does not cause saturation of OP2 output and hence do not affect the angle measurement. Errors due to the time delay of SW, comparator response time, and the CLU accuracy: In the RDCT, SW can be realized with the help of a commercially available CMOS switch. These switches typically have a delay of 10 ns, and their response time is around 60 ns. Similarly, comparator also takes on an average 170 ns of time to respond to the step change in its input. On the other hand, the RDCT conversion time is in milliseconds. Therefore, the effects of these two parameters can be conveniently

Design and Performance Evaluation of a Simple Resistance-to-Digital …

267

ignored. However, the time resolution of the CLU plays a crucial role in deciding the integration time (T I ) which in turn governs the maximum speed limit for dynamic input tracking. A CLU with time resolution of 1 µs is a good choice in terms of its cost and dynamic input tracking capability. Other non-idealities: The passive components (RI and C I ) of the integrator may drift with time and/or temperature. However, such a drift does not affect the measurement as both the integration and de-integration phases are equally influenced. The assumption that drift in RI and C I is rapid and is in comparison with the conversion time is very impractical. Therefore, we can confidently claim that variations in RI and C I do not affect the measurement. Likewise, slow and minor variations in V D do not introduce any error in the RDCT output. Apart from all these error sources, the RDCT performance may get affected due to some random effects. The effect of random errors can be minimized by taking average of multiple readings of Dθ .

4 Extensive Performance Evaluation of the RDCT In the previous two sections, the working principle of TMR angle sensor and the proposed RDCT, and detailed error analysis have been discussed. In this section, simulation studies for ideal as well as various practical/non-ideal cases are described.

4.1 Simulation Studies for Ideal TMR Angle Sensor and Ideal RDCT The performance of the proposed RDCT was studied in simulation using the LTSPICE software from Linear Technology Corporation. The TMR resistances RC1 and RC2 were made to vary as per (1) so as to mimic the shaft rotation for half-circle range. This was achieved using SPICE programming. The values of RN and ΔR were chosen as 26 kΩ and 4.5 kΩ, respectively, which are true for the AAT003-10E half-bridge TMR IC [16]. Note that the RDCT will work equally well for any other TMR angle sensor with different value of RN and ΔR [19]. The components OP1 and OP2 were realized using the SPICE model for the OP07 IC from Texas Instruments. The switch SW was built in a similar way using equivalent model of MAX4053 IC. The value of RI and C I was chosen as 22 kΩ and 470 nF, respectively. These values ensure that the integrator output, vI , does not reach saturation when T I is set as 20 ms. As mentioned earlier, the time T I should be chosen sufficiently small in the case of dynamic input tracking. Therefore, the values of RI and C I must be selected such that the saturation of OP2 output is avoided as well as the de-integration time is measured accurately. The reference voltage V D is set as 2 V so as to restrict the net excitation voltage for the TMR sensor (V D + vS ) within the specified limit [16]. The programming part of the RDCT was achieved with the help of a precision monostable

268

K. B. Nandapurkar

Fig. 4 RDCT response obtained in LTSPICE-based simulation studies for ideal TMR angle sensor. As desired, the output angle varies linearly with the shaft angle, θ, for half-circle range

multivibrator that provides a pulse of 20 ms duration for the integration phase. The measurement of de-integration time T V was done using .meas command in LTSPICE. For the ideal simulation, different non-idealities, discussed in the previous section, were set to their respective ideal values. The resistances RC1 and RC2 were varied to mimic the shaft rotation over half-circle range, in steps of 3°, and the variable T V was recorded, for each step. Later, the intermediate variable, x and the output angle, Dθ were computed offline using (5) and (6), respectively. The obtained Dθ was plotted against θ and is shown in Fig. 4. As can be seen in this graph, Dθ varies linearly with θ for 180° range. The error in Dθ , for each step of θ, is shown in Fig. 4 using green-colored scatter plot. The maximum percentage error spotted in Dθ is 0.34%, which amounts to an angular error of 0.62°. For 5–175° range, the worst-case error in was found to be just 0.042% (= 0.08°). The high output error at θ = 0 and 180° is due to the combined effect of undesired mismatch in actual T I and expected T I and various parameters of RDCT which could not be set as ideal due to limitations.

4.2 Simulation Studies for a Practical TMR Angle Sensor In this part of the study, the TMR resistances RC1 and RC2 were modeled to have mismatch δ = τ = +0.05% and phase imbalance, β = +0.2°. Keeping all the other parameters same as that for the previous case, the simulation program was run and new T V , was recorded for every step of θ. The output angle was estimated from T V , . Results obtained in this part of the study are shown in Fig. 5a. It can be seen that the RDCT output, Dθ , fairly varies linearly with θ. However, the maximum percentage error for such a practical case is increased to 1.8% for 180° range. On the other hand, for 5–175° range, the maximum error is 0.49%. The offset-reduction technique, discussed in the error analysis section, was then applied and the output angle, Dθ , was estimated from x , . As expected, the output error is reduced greatly

Design and Performance Evaluation of a Simple Resistance-to-Digital …

269

Fig. 5 RDCT response for a practical TMR angle sensor with three non-idealities is shown in a. The output angle has significant error, mainly due to mismatch in RN . After implementing the offset-reduction strategy, this error, as shown in b, gets reduced significantly

to 0.41% for half-circle range. This proves the efficacy of the offset-compensation technique in simulation environment. Later, the effect of different RON of SW was also studied. The RON was varied from 50 to 300 Ω, in steps of 50 Ω, and the worstcase error in Dθ was found for each case. The maximum error was observed quite close for all the cases considered. Thus, as expected, RON was found to affect the output negligibly. Based on these simulation studies, it can be commented that the proposed RDCT indeed works as a resistance-to-digital converter for TMR angle sensor. The RDCT, as highlighted earlier, does not require costly instrumentation amplifiers, sine wave oscillator, and two reference voltage supplies for its operation. In real-time environment, when realized using suitable components, the RDCT will yield linear directdigital output for 180° range, without utilizing any separate ADC(s) and LUTs. The approximate component cost of RDCT (TMR angle sensor + RDCT circuit components) is just 15 USD. Thus, a low-cost angle sensing solution with sufficient accuracy is provided in this work for wide range of applications.

5 Conclusions and Future Work Plan In this work, a simple RDCT, suitable for measuring shaft angle over half-circle range, has been presented. The RDCT, in contrast to several existing interface circuits [8–15], is much simpler in design. The RDCT was designed with an objective of realizing an ADC-free, low-cost angle measurement system. This objective has been successfully achieved. Factors affecting the RDCT performance have been identified and analyzed. A simple and efficient offset-error reduction method was presented that drastically reduces the output error due to the mismatch in the nominal resistances of the TMR angle sensor. The performance of the RDCT was evaluated for ideal as well as non-ideal TMR angle sensor using detailed simulation studies. The results

270

K. B. Nandapurkar

obtained in these studies showcase the capability of RDCT to linearize and digitize the response of an inherently sine–cosine and hence, nonlinear natured response of the TMR angle sensor. The RDCT can be used in numerous applications such as tracking of solar panels, positioning of engine throttle-valve, accelerator-pedal, operating tables, and dental chairs in the hospitals. The future work is planned in the following manner. The experimental evaluation of the RDCT will be done to verify the methodology in real-time scenario. For this purpose, the TMR-based shaft angle sensing unit needs to be built. With the sensing setup, we plan to do several experiments for static as well as dynamic measurement. Since the RDCT is basically a digital instrumentation system, parameters such as signal-to-noise ratio (SNR), angular resolution, standard deviation, effective number of bits (ENoB), and repeatability error can be estimated. Note that these parameters can be suitably adjusted by tuning the integration time (T I ) of the RDCT carefully. It will also be interesting to verify the efficiency of the offset-error reduction technique in actual experimentation with the TMR-based sensing setup. The findings of these studies will be reported later.

References 1. Doebelin EO (2013) Measurement systems: application and design, 6th edn. McGraw Hill Publishing Co., New Delhi 2. The sixth sense for automotive applications. http://www.nxp.com/documents/brochure/750 15728.pdf. Accessed 05 Aug 2021 3. Zhang J, Hua S, Qi Z, Feng W (2011) The design and implementation of the shaft angle acquisition system used in the solar panel. In: Fourth international symposium on computational intelligence and design. IEEE, Hangzhou, pp 287–290 4. Robot sensor and actuators. http://robotics.sjtu.edu.cn/upload/course/5/files/Robot%20Sens ors%20and%20Actuators-new.pdf. Accessed 25 July 2021 5. Pallas-Arney R, Webster JG (2012) Sensors and signal conditioning, 2nd edn. Wiley, New Delhi 6. Arc sensor applications. https://cambridgeic.com/applications/arc-sensor-applications. Accessed 12 June 2021 7. Application note—transportation. https://sensing.honeywell.com/honeywell-value-addedsmart-rotary-boom-sprayer-arm-angle-position-sensing-000757-1en.pdf. Accessed 02 Aug 2021 8. Lopez-Martin AJ, Carlosena A (2009) Performance tradeoffs of three novel GMR contactless angle detectors. IEEE Sens J 9(3):191–198 9. Farcas C, Cosma A, Palaghitia N, Grama A, Gabor MS, Tiusan C (2014) Development of a noncontact angular transducer. In: International symposium on design technology and electronic packaging. IEEE, pp 187–191 10. Leitis K, Bonath W (2002) Magnetoresistive sensors and a new hardware-based interpolation method for length and angle measurements. In: Sensors conference. IEEE, Orlando, pp 1432– 1435 11. Anoop CS, George B (2012) Electronic scheme for computing inverse-cosine and its application to a GMR based angle sensor. IEEE Trans Instrum Meas 61(7):1991–1999 12. Benammar M, Ben-Brahim L, Alhamadi MA (2005) A high precision resolver-to-DC converter. IEEE Trans Instrum Meas 54(6):2289–2296

Design and Performance Evaluation of a Simple Resistance-to-Digital …

271

13. Bhaskarrao NK, Anoop CS, Dutta PK (2017) A simple and efficient front-end circuit for magneto-resistive angle sensors. In: International instrumentation and measurement technology conference. IEEE, Turin, pp 1–6 14. Kishor Bhaskarrao N, Sreekantan AC, Dutta PK (2018) Analysis of a linearizing direct digitizer with phase-error compensation for TMR angular position sensor. IEEE Trans Instrum Meas 67(8):1795–1803 15. Bhaskarrao NK, Anoop CS, Dutta PK (2021) A novel linearizing signal conditioner for halfbridge-based TMR angle sensor. IEEE Sens J 21(3):3216–3224 16. AAT003 TMR angle sensor kit. https://www.nve.com/Downloads/ag931-07.pdf. Accessed 10 Aug 2021 17. Wang D, Brown J, Hazelton T, Daughton JM (2005) Angle sensor using spin-valve with SAF structure. IEEE Trans Magn 41(10):3700–3702 18. Wang WG et al (2013) Parallel fabrication of magnetic tunnel junction nanopillars by nanosphere lithography. Sci Rep 3(1):1–6 19. AAT00X TMR angle sensor. https://www.nve.com/Downloads/AAT00X.pdf. Accessed 10 Aug 2021

The Use of LBP Features in Transform Domain for Object Recognition R. Ahila Priyadharshini and S. Arivazhagan

Abstract Local binary pattern (LBP) is very useful in various applications of machine vision problems. It is an effective multi-resolution texture descriptor. Various types of local binary patterns have been proposed till now. Even though deep learning framework has gained more attention in generic object recognition nowadays, in this article, we studied the effectiveness of LBP features in the transform domain for the task of object recognition. The comparison was done based on the performance measure such as recall, precision, equal error rate and F-measure. Images have been pre-processed using Log Gabor filter, and then, different types of LBPs are applied on the transformed images, and the statistical features are extracted and classified with support vector machine (SVM). The evaluation is done on the Graz databases. It is observed that rotation invariant LBP and direction coded LBP perform better with lesser error rate and higher recall, precision and F-measure. Keywords Object recognition · Local binary pattern · Log Gabor filter · Equal error rate

1 Introduction Object recognition is extensively used in the machine vision industry for the commitments of inspection, registration quality control, etc. The images are dependent on various parameters such as illumination, camera parameters and location. As an object should be recognized from the images of a scene which contains multiple objects, the complexity level of object recognition model relies on many aspects such as scene constancy, number of total objects in an image, image-models spaces, number of objects in the model database and possibility of occlusion. In the object recognition task, an input image is assigned with one of the N class labels. The main steps involved in this process are (i) preprocessing, (ii) feature analysis and (iii) classification. In the preprocessing, basic transformation is applied on the image R. Ahila Priyadharshini (B) · S. Arivazhagan Mepco Schlenk Engineering College, Sivakasi, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_18

273

274

R. Ahila Priyadharshini and S. Arivazhagan

[1]. In feature analysis, set of feature vectors are derived that represent the image highlighting the significant aspects. Most of the machine vision systems use either global features or local features. Global features represent the complete image as a single vector [2]. Local features are derived at several locations in the image and are robust to clutter and occlusion. In this work, local features are computed by applying uniform grid on the Log Gabor transformed images. The Gabor function is a prevailing tool in the field of machine vision. Gabor function is optimally localized in spatial and frequency domain. So, it is predominantly used for texture analysis [3]. Log Gabor filter has Gaussian response on logarithmic frequency scale. It captures additional information in high frequency regions [4, 5]. Hu et al. [6] used Log Gabor filter for fingerprint segmentation. Lajevardi et al. recognized facial expression using Log Gabor filter and LBP [7]. The performance based on Log Gabor filter gives better classification accuracy than LBPs. Seif et al. used Log Gabor filter for iris recognition [8]. The traditional LBP was proposed by Ojala et al. [9]. Fathi et al. [10] introduced a new local binary operator called rotational invariant local binary operator (LBPriu2 P,R ) for automatic blood vessel segmentation. Trefny et al. [11] presented two different encoding schemes for local binary pattern to represent the intensity function of local neighborhood. Wu et al. [12] proposed a novel local binary operator named as improved center symmetric LBP (ICSLBPP,R ) to extract more texture information from an image to increase the recognition rate. Martins et al. [13] demonstrated the classification of different forest species using original LBP. Nguyen et al. [14] proposed a modification in the LBP named as nonredundant LBP (NRLBPP,R ) and used it for detecting the objects. Liu et al. [15] used a set of six LBP-like features derived from local intensities and differences for face recognition. Arivazhagan et al. [16, 17] used Log Gabor transform for automatic target recognition in SAR images and vehicle recognition. The main focus of this work is to compare the performance of various types of LBPs for the object recognition task and to find the best binary pattern for each category of objects like bike, bike and person, person and car based on the performance measures such as recall, precision, F-measure and equal error rate. Instead of applying local binary pattern to the image directly, here, LBP is taken on the transformed image. Local features are derived here to capture various regions of complex images and are categorized using SVM classifier. The proposed method is evaluated on Graz database.

2 Materials and Methods First, the input is transformed using Log Gabor filter with four scales and six orientations. Then, a uniform rectangular grid is applied on the transformed images. Afterward different types of LBPs such as original LBP (o_LBP), rotation invariant LBP (ri_LBP), improved center symmetric LBP (ICS_LBP), transition LBP (tLBP), direction coded LBP (dLBP) and non-redundant LBP (NRLBP) are computed over the

The Use of LBP Features in Transform Domain for Object Recognition

275

Fig. 1 Proposed method’s block diagram

non-overlapping patches. Finally, mean and standard deviation are computed and classified using SVM classifier. The proposed method’s block diagram is displayed in Fig. 1.

2.1 Log Gabor Transform Gabor filter, a linear filter which has impulse response as a sine modulated Gaussian function. Due to the limitations of Gabor filter [18], Log Gabor function, an alternate to Gabor is suggested by D Field [19]. Log Gabor filters can be built with arbitrary bandwidth. The radial component G( f ) and the angular component G(ϕ) of Log Gabor function are defined in Eq. 1. 2  f log f 0 −   2 σ 2 log f f 0 

G( f, ϕ) = G( f ) × G(ϕ) = e

ϕ−ϕ0 )2 2σϕ2

−(

×e

(1)

where f 0 denotes the center frequency, σf denotes the bandwidth scaling factor, ϕ0 denotes orientation angle and σϕ represents angular bandwidth. The vital characteristics of Log Gabor functions are (i) DC component is not present in Log Gabor functions so that the contrast ridges and edges of images are improved. (ii) It has a prolonged tail transfer function, providing wide spectral information with localized spatial extent. Here, Log Gabor functions with six orientations and four scales are used. Figure 2 depicts 24 sub-bands of the transformed images.

2.2 Local Binary Patterns LBP is a powerful descriptor for texture classification [20]. LBP is robust to illumination, and it is computationally simple. Original Local Binary Pattern (LBPP,R )

276

R. Ahila Priyadharshini and S. Arivazhagan

Fig. 2 Output of Log Gabor transform with four scales and six orientations for a sample image

The original LBP is computed by thresholding a N × N neighborhood with the center value and is depicted in Fig. 3. The LBP has a limitation while using small neighborhood as it is unable to capture the prevailing features in huge structures. This can be overcome by extending the operator to cope with larger neighborhoods [9]. With the help of circular neighborhoods and bilinear interpolation, the neighborhood with arbitrary radius can be allowed. The LBPP,R generates 2 p dissimilar binary patterns. The mathematical representation of the LBPP,R is given by Eq. 2

Fig. 3 LBPP,R computation

The Use of LBP Features in Transform Domain for Object Recognition

LBPP,R =

p−1    s g p − gc 2 p

277

(2)

p=0

Rotational Invariant Local Binary Pattern (LBPP,R riu2 ) In traditional LBP, if the size of neighborhood increases, the number of prevailing texture patterns also grow extensively. To diminish the count of texture patterns, an extension, rotation invariant method, named LBPriu2 , was proposed [10]. The LBPriu2 is represented in Eq. 3

LBPriu2 P,R =

⎧ p−1 ⎨

  BCP,R (i ) U BCP,R ≤ 2

⎩ i=0 p+1

(3)

Otherwise

The LBPriu2 successfully extracts texture primitives in flat surfaces and straight edges [9]. Improved Center Symmetric Local Binary Pattern (ICS_LBP) The improved version of CS-LBP [12] named ICS-LBP is illustrated in Fig. 4 for 8 neighbors. The LBP yields 256 (28 ) dissimilar binary patterns, however, ICS-LBP gives only 16 (24 ) dissimilar binary patterns. The mathematical representation of ICS-LBP is given in Eq. 4 ICS-LBPP,R (x, y) =

p ( 2 )−1

 sICS-LBP pi , pc , pi+( 2p ) × 2i

(4)

i=0

Transition Local Binary Pattern (tLBP) Transition coded LBP gives the neighborhood pixel comparison in rightward direction for all pixels except the center. Each sequence with same binary values indicates ordered sequence of pixel intensities. Thus, the tLBP is gray-scale invariant.

Neighborhood

Fig. 4 ICS-LBP features for 8 pixel neighborhood

Binary Pattern

278

R. Ahila Priyadharshini and S. Arivazhagan

The mathematical representation for transition LBP is given in Eq. 5 tLBPP,R

P−1    = s(go − g P−1 ) + s g P − g p−1 2 P

(5)

p=1

Direction Coded Local Binary Pattern (dLBP) Motivation of dLBP is to provide improved information of local pattern in terms of direction functions similarly to CS-LBP [11]. In dLBP, intensity variation toward four base directions are encoded into two bits. The expression for dLBP is given in Eq. 6 dLBPP,R =

 P −1  





     s g p − gc g p + p − gc 22 p + s g p − gc − g p + p − gc 22 p +1

p =0

(6) Both the LBP and the dLBP rules encode if center pixel is an extrema. Unlike the LBP rule, the dLBP does not encode it as maximum or minimum but encodes if sign of first and second differential is the same. Non-redundant LBP (NRLBP) The two notable disadvantages of original LBP to encode object’s appearance [14] are storage requirement and discriminative ability. LBP codes indicated by red boxes in both the images in Fig. 5 are different although they characterize same structure.

Fig. 5 LBP code of a sample image and its inverted image

The Use of LBP Features in Transform Domain for Object Recognition

279

Non-redundant LBP (NRLBP) overcome both of the above issues and is given in Eq. 7   NRLBPP,R (xc , yc ) = min LBPP,R (xc , yc ), 2 P − 1 − LBPP,R (xc , yc )

(7)

The NRLBP treats the LBP code and its complement as equal. The number of histogram bins in NRLBP is decreased by half.

2.3 Performance Measures Recall gives the proportion of the objects that are detected correctly, and precision indicates how often the detections the system makes are false. These two quantities are given in Eqs. 8 and 9, respectively. Recall =

TP TP + FN

Precision =

TP TP + FP

(8) (9)

F 1 measure is shown in Eq. 10. Accuracy is the ratio of right predictions to the total predictions and is given in Eq. 11. F-measure = Accuracy =

2 × Recall × Precision Recall + Precision

(10)

TP + TN TP + FN + TN + FP

(11)

3 Experimentation Results The experiments are done on Graz database [21, 22]. The images in Graz databases are resized to 300 × 300 for experimentations. All categories of Graz 01 database and one category (car) from Graz 02 database are considered for experimentation. Every image in Graz database is convolved with Log Gabor function with six orientations and four scales and resulting in 24 sub-bands. For every sub-bands, a uniform grid of size 3 × 3 is applied. Then, for every 3 × 3 non-overlapping patches, six different types of LBPs such as original LBP, rotation invariant LBP, ICS-LBP, transition LBP, dLBP and non-redundant LBP are applied. The statistical features are derived and classified using SVM classifier (RBF kernel). The different types of LBP patterns for a sample image are shown in Fig. 6.

280

R. Ahila Priyadharshini and S. Arivazhagan

i

ii

iii

iv

v

vi

vii

viii

Fig. 6 Different types of LBP pattern (i) sample image, (ii) filtered image, (iii) LBPP,R , (iv) LBPP,R riu2 , (v) ICS_LBP, (vi) tLBP, (vii) dLBP, (viii) NRLBP

The training and testing ratio are maintained as in [23]. The training and testing are repeated for 10 times with non-overlapped random split. The performance measures and accuracy obtained using six different types of LBPs separately for Graz 01 database and Graz 02 databases are shown in Tables 1 and 2, respectively. Table 1 depicts that rotation invariant LBP gives highest recall, precision and Fmeasure for two categories such as bike and bike and person. For categories such as person and car, direction coded LBP yields the highest measure. Figure 7 depicts the average ROC curves of Graz 01 database and Graz 02 database using six different LBPs. The obtained AUC values are given in Table 3. The largest AUC values are highlighted. Equal error rates (EER) for the Graz databases are displayed in Fig. 8. From Fig. 8, it is understood, riLBP accomplishes well for two categories. For person and car tasks, direction coded LBP yields better performance. Table 3 shows the computed AUC values for Graz database. Table 4 shows the average EER of 10 runs for the Graz database. Our proposed method is comparable to the state-of-the-art methods.

Precision (%)

F-measure (%)

95.6

Car

85.8

79.2

95

76

87.8

84.9

97.4

72.4

70.4

97.3

86.4

97.4

o_LBP—original LBP, ri_LBP—rotation invariant LBP Bold specifies the highest performance measure

79.7

99.2

82.8

92.4

Bike and 92 person

86.8

75.8

Bike

Person

88.4

97

76

87.6

61.3

62.9

52.2

86.6

88.96

85.96

62.4

91.8

61.9

75.2

57.1

84.4

66.8

73.5

55.3

87.1

94.7

73.3

91.4

73.9

78.3

78.8

59.7

83.9

69.3

74.7

61.8

86.7

92.2

92

71.2

92

69.5

83.9

65.2

85.1

74.8

83.8

62.7

87.4

85.7 87

66.9

95.98 83

79.3

94.3

72.1

O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP

Category Recall (%)

Table 1 Recall, precision and F-measure of Graz database

The Use of LBP Features in Transform Domain for Object Recognition 281

282

R. Ahila Priyadharshini and S. Arivazhagan

Table 2 Accuracy for Graz databases AUC

Database

OLBP

riLBP

ICS-LBP

tLBP

dLBP

NRLBP

Bike

86.70

92.10

85.00

87.40

72.80

85.40

Person

58.20

66.50

59.40

56.90

94.10

62.50

Bike and person

68.90

91.50

81.80

81.10

77.50

85.20

Car

64.70

91.90

65.20

56.90

95.90

80.90

Bold specifies the highest performance measure

ROC curve for Person in Graz01-LBP 1

0.9

0.9

0.8

0.8

True positive rate

True positive rate

ROC curve for Bike in Graz01-LBP 1

0.7 0.6 0.5 0.4

RotLBP OriLBP ICSLBP TransLBP DirLBP NRLBP

0.3 0.2 0.1

0.7 0.6 0.5 0.4

RotLBP OriLBP ICSLBP TransLBP DirLBP NRLBP

0.3 0.2 0.1

0

0 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0

1

0.1

0.6

0.9

0.9

0.8

0.8

0.7 0.6 0.5 0.4

RotLBP OriLBP ICSLBP TransLBP DirLBP NRLBP

0.3 0.2 0.1 0.4

0.5

0.6

False positive rate

(iii)

0.7

0.8

0.9

0.8

0.9

1

0.7 0.6 0.5 0.4

RotLBP OriLBP ICSLBP TransLBP DirLBP NRLBP

0.3 0.2 0.1

0 0.3

0.7

ROC curve for Car in Graz02-LBP

True positive rate

True positive rate

0.5

(ii) 1

0.2

0.4

(i) ROC curve for Bike and Person in Graz01-LBP

0.1

0.3

False positive rate

1

0

0.2

False positive rate

1

0 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False positive rate

(iv)

Fig. 7 ROC curves of Graz databases. (i) Bike, (ii) person, (iii) person and bike, (iv) car

4 Conclusion LBP extracts structural primitive textures excellently. In this paper, the task of object recognition is investigated by using six different types of LBPs in the transformed

The Use of LBP Features in Transform Domain for Object Recognition

283

Table 3 Computed AUC values for categories of Graz database Database

AUC riLBP

OLBP

ICS-LBP

tLBP

dLBP

NRLBP

Bike

0.9210

0.8670

0.8500

0.8740

0.7280

0.8540

Person

0.6650

0.5320

0.5940

0.5690

0.9410

0.5870

Bike and person

0.9150

0.6890

0.8180

0.8110

0.7750

0.8520

Car

0.9187

0.6467

0.6520

0.7133

0.9593

0.8093

Bold specifies the highest performance measure

Fig. 8 Comparison of ROC EER for categories of Graz databases

Table 4 Average EER on the Graz database Database

[23]

[24]

[25]

[26]

[27]

ri_LBP

d_LBP

Bike

16.5

13.7

10

–

19.8

8.2

24.8

Person

23.5

17.7

13

–

16

49.8

9.2

Bike and person

–

–

–

–

–

16.2

31.4

Car

29.8

–

–

31.61

24.4

11.87

5.47

Bold specifies the highest performance measure

domain on the complex Graz database. It can be concluded that rotation invariant LBP and direction coded LBP are effective for object recognition task. The improved performance is due to the intensity inversion invariance property of direction coded LBP.

284

R. Ahila Priyadharshini and S. Arivazhagan

References 1. Kolsch T (2003) Local features for image classification. Diploma thesis, Lehrstuhl fur Informatik VI, RWTH Aachen University, Aachen, Germany 2. Hegerath A (2006) Patch-based object recognition. Diploma thesis, Human Language Technology and Pattern Recognition Group, RWTH Aachen University, Germany 3. Kanagalakshmi K, Chandra E (2012) Frequency domain enhancement algorithm based on Log-Gabor filter in FFT domain. Glob J Comput Sci Technol 12(7), Version 1.0 4. Cook J, Chandran V, Sridharan S, Fookes C (2005) Gabor filter bank representation for 3D face recognition. In: Proceedings of the digital image computing techniques and applications (DICTA’05). IEEE. ISBN 0-7695-2467-2 5. Liu C, Wechsler H (2003) Independent component analysis of Gabor features for face recognition. IEEE Trans Neural Netw 14(4):919–928 6. Hu C, Yin J, Zhu E, Chen H, Li Y (2010) A composite fingerprint segmentation based on LogGabor filter and orientation reliability. In: IEEE international conference on image processing (ICIP), pp 3097–3100 7. Lajevardi SM, Hussain ZM (2009) Facial expression recognition using Log-Gabor filters and local binary pattern operators. In: Proceedings of international conference on communication, computer and power (ICCCP’09), Feb 2009, pp 349–353 8. Seif A, Zewail R, Saeb M, Hamdy N (2003) Iris identification based on log Gabor filtering. In: Proceedings of IEEE 46th Midwest symposium on circuits and systems, vol 1, pp 333–336 9. Ojala T, Pietikainen M, Maenpaa T (2002) Multi resolution grayscale and rotation invariant texture classification with local binary patterns. IEEE Trans Pattern Anal Mach Intell 24(7):971–987 10. Fathi A, Naghsh-Nilchi AR (2014) General rotation-invariant local binary patterns operator with application to blood vessel detection in retinal images. Pattern Anal Appl 17(1):69–81 11. Trefny J, Matas J (2010) Extended set of local binary patterns for rapid object detection. In: Computer vision winter workshop, vol 1, pp 1–7 12. Wu X, Sun J (2009) An effective texture spectrum descriptor. In: Fifth international conference on information assurance and security (IAS), vol 2, pp 361–364 13. Martins J, Oliveira LS, Nisgoski S, Sabourin R (2013) A database for automatic classification of forest species. Mach Vis Appl 24(3):567–578 14. Nguyen D, Zong Z, Ogunbona P, Li W (2010) Object detection using non-redundant local binary patterns. In: IEEE 17th international conference on image processing, pp 4609–4612 15. Liu L, Fieguth P, Zhao G, Pietikäinen M, Hu D (2016) Extended local binary patterns for face recognition. Inf Sci 358–359:56–72 16. Arivazhagan S, Priyadharshini RA, Sangeetha L (2017) Automatic target recognition in SAR images using quaternion wavelet transform and principal component analysis. Int J Comput Vis Robot 7(3):314–334 17. Priyadharshini RA, Arivazhagan S, Sangeetha L (2014) Vehicle recognition based on Gabor and Log-Gabor transforms. In: IEEE international conference on advanced communications, control and computing technologies, pp 1268–1272. https://doi.org/10.1109/ICACCCT.2014. 7019303 18. Yu Q, Tan X, Wang H (2015) A face recognition method based on total variation minimization and Log-Gabor filter. In: International conference on electromechanical control technology and transportation 19. Field D (1987) Relations between the statistics of natural images and the response properties of cortical cells. J Opt Soc Am A 4(12):2379–2394 20. Oulu (2003) The local binary pattern approach to texture analysis—extensions and applications. Diploma thesis, Infotech Oulu, University of Oulu 21. http://www.emt.tugraz.at/~pinz/data/Graz-01/ 22. http://www.emt.tugraz.at/~pinz/data/Graz-02/ 23. Opelt A, Pinz A, Fussenegger M, Auer P (2006) Generic object recognition with boosting. IEEE Trans Pattern Anal Mach Intell (PAMI) 28(3):416–431

The Use of LBP Features in Transform Domain for Object Recognition

285

24. Lazebnik S, Schmid C, Ponce J (2006) Beyond bags of features: spatial pyramid matching for recognizing natural scene categories. In: CVPR, vol 2, pp 2169–2178 25. Boiman O, Shechtman E, Irani M (2008) In defense of nearest-neighbor based image classification. In: CVPR 2008, pp 1–8 26. Aldavert D, Ramisa A, Toledo R, de Mantaras RL (2009) Efficient object pixel-level categorization using bag of features. In: Advances in visual computing. Lecture notes in computer science, vol 5875, pp 44–54 27. Ghodrati M, Khaligh-Razavi S-M, Ebrahimpour R, Rajaei K, Pooyan M (2012) How can selection of biologically inspired features improve the performance of a robust object recognition model? PLoS ONE 7(2):e32357

Design and Simulation of Capacitive Pressure Sensor for Monitoring Lead-Acid Battery Charge Yashwant Adhav, Dayaram Sonawane, and Chetankumar Patil

Abstract In the present study, we have proposed a design of a novel, graphenebased micro-capacitive pressure sensor to measure minute variation in differential pressure developed in the air-purge system of lead-acid battery. Online state of charge (SOC) monitoring of lead-acid batteries using a sensor is a critical problem. The key principle is that specific gravity of the acid solution interlinks the amount of charge or discharge of the acid solution and the differential pressure of the air-purge system. We have obtained results from the analytical model, which are benchmarked with the numerical model of Comsol Multiphysics commercial software. The proposed graphene-based capacitive MEMS sensor has been designed for the range of 0– 1 kPa differential pressure variation. The gap between the two capacitive plates is maintained at 3 µm and the thickness of the circular diaphragm is 10 µm. It is found in this study that the minuscule range of small changes in the specific gravity (or pressure) is consistent with the sensitivity of the newly designed MEMS sensor. In a nutshell, it can measure a tiny pressure drop in the air-purge system. On the contrary, the micro-piezo resistance sensor is observed to be less accurate on the account of its temperature sensitivity. Through this study, it is observed that the graphenebased MEMS sensor shows greater accuracy compared to the silicon-based sensor. The change in capacitance of graphene is linear compared to silicon. In case of the graphene, capacitance changes from 4.875 × 10−12 to 4.955 × 10−12 pF whereas in silicon, capacitance changes from 2.3 × 10−12 to 3.5 10−12 pF. Keywords State of charge · MEMS capacitive sensor · Graphene · Air-purge tube · MEMS sensor design Y. Adhav (B) Cummins College of Engineering for Women, Pune 411052, India e-mail: [email protected] D. Sonawane · C. Patil Department of Instrumentation and Control Engineering, College of Engineering, Shivajinagar, Pune 411005, India e-mail: [email protected] C. Patil e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_19

287

288

Y. Adhav et al.

1 Introduction This study is about knowing the SOC of a lead-acid battery. Lead-acid batteries have stable outputs, low costs and capability to withstand at different temperature ranges. Such battery has a lot of scope in photo voltaic power systems, wind power plants, uninterpreted power systems, smart grids and submarine [1–6]. The hydrostatic principle is used to see the state of charge of the battery. When the battery gets charged or discharged, the concentration of acid changes. The specific gravity (SG) of acid changes due to a change in the concentration of the sulphuric acid. The state of the battery is monitored by measuring SG [3]. As per the hydrostatic pressure principle, if the hydrostatic head is constant, then pressure is directly proportional to the density of sulphuric acid. In a lead-acid battery, the bottom portion is covered with lead sheets. At the top side of lead sheets, a very little space is available for acid where these air-purge sensing tubes are immersed. Back pressure developed in this air tube is directly proportional to the specific gravity of acid in the battery. It is the most important point to design a sensor that can measure this back pressure developed in the bubbler tube. As the head is very low, it is challenging to measure a specific gravity of the solution having the head as almost 2–3 cm above lead sheets in the battery. A small air-purge system has already been developed [7] to know the SG of the acid solution. In this patent [7], MEMS-based piezoresistive pressure sensor is used to measure differential pressure. The main disadvantage of this pressure sensor is being more sensitivity to temperature. As altitude changes, the sensor can produce errors as pressure change is directly proportional to the altitude. To overcome this limitation in the current study, a graphene capacitive pressure sensor is designed and simulated using COMSOL Multiphysics simulation software. When load is applied to a thin diaphragm, it deflects. Pressure-deflection proportional relation can be simulated using a numerical method known as the finite element method. An analytical method based on large-small deflections is used to predict sensitivity and linearity [4]. The thickness of diaphragm can produce errors in linearity. Linearity error of ± full-scale span (FSS) is accomplished. Both circular as well as the rectangular diaphragm are examined for optimization. The analytical method in this study is used to compare the results obtained by the simulation finite element method. When pressure is applied to diaphragm, the gap between the plates changes non-uniformly. FEM is performed to detect that deformation and detect a change in capacitance of polysilicon plates [8]. For 0.1–1 MPa pressure displacement, the change in capacitance of 30–44 pF has been noted. The finite element method (FEM) used as the gap between plates varies nonuniformly when pressure is applied to one of the plates. PDMS (polydimethylsiloxane) deformable dielectric is used between the two plates. When the pressure is applied, dielectric layer separates the two electrodes [2], hence changes the sensor’s capacitance. Graphene is an emerging material in the MEMS field having high mechanical strength. This material consists of a thin layer of carbon atoms that can make thickness up to nanometers [9]. The sensitivity of this low-pressure sensor is 288 pF/Pa of 4 mm × 4 mm size.

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

289

Fibre optical sensor, multiple point fibre sensor, ultrasonic, quartz crystal micro balance technique, the piezoelectric sensor are used to estimate SOC which is discussed in [3, 5, 6, 10, 11] in this paper; micro technology is superior in performance over these macro techniques in terms of accuracy, size, strength and cost. Capacitive pressure sensor with novel liquid crystal polymer is designed to have circular thin plate [10]. Diaphragm having radius 0.0016 µm provides a total capacitance change of 0.277 pF for the pressure range of 0–100 kPa. In this study, circular diaphragm sensor is designed for the 0–6 kPa pressure range. The circular Silicon diaphragm and graphene diaphragm results are compared in terms of displacement of thin plate and capacitance change. The change in capacitance is non-linear in Silicon in comparison with the graphene. Capacitive sensors have lower noise and as well it requires lower power which is an added advantage. The fabrication process of the capacitive pressure sensor is comparatively more straightforward than the piezoresistive pressure sensor. For the measurement of low pressure, a more sensitive pressure sensor is required. With an increase in sensitivity of diaphragm, thickness of the sensor decreases, which in turn leads to the non-linear relationship between applied pressure and deflection of the diaphragm. In capacitive pressure sensor design, the diaphragm’s thickness, linearity and deflection are equally essential parameters [4]. As silicon thickness is reduced, it is difficult to fabricate the thin plate, and hence the sensor becomes bulky. CMOS (Complementary metal oxide silicon) material is used for the thin plate to reduce the noise and parasitic capacitance [12]. This pressure sensor has a sensitivity to 4.2 pF/bar. Flexible capacitive pressure sensor is developed for different applications like sitting posture and tongue pressure. The surface sensor is made with polyimide substrate stacked between ethylene vinyl acetate (EVA) and polyvinyl silicone (PVS) [2]. The circular diaphragm shows maximum deflection compared to rectangular and square diaphragm [13]. The circular shape is selected for the diaphragm to get more sensitivity and stiffness. In the past-related research, no researcher carried out work on designing a circular capacitive pressure sensor whose diaphragm is made from graphene material for an acid battery’s charge monitoring. Graphene material with high ultra toughness, high elasticity and electron mobility is also high, transparency and conductivity are also superior. As the hydrostatic head of sulphuric acid is minimal in the lead-acid battery, this sensor is designed for 0–1 kPa pressure range. In this study, we can observe and measure the diaphragm’s deflection when the battery is discharged and the displacement of the diaphragm when the battery is charged. Graphene diaphragm is one plate, and the other plate is placed with a gap of 3 µm. As for battery charge–discharge due to change in differential pressure, the graphene diaphragm deflects; hence, the capacitance of the sensor changes. By measuring capacitance, the state of the battery can be estimated. In this simulation, electromechanics physics is encompassing to design sensor available under structure mechanics. Structure mechanics integrate Solid Mechanics and Electrostatics and also allows to move mesh. The deformation can be modelled using electro statically actuated structures. This paper consists of three main parts. In Sect. 2, the design of MEMS capacitive pressure sensor using analytical method and design with Comsol Multiphysics

290

Y. Adhav et al.

are described. In Sect. 4, the results of analytical and finite element methods are compared. Section 5 is about the conclusion and future scope of research.

2 Design of MEMS Capacitive Pressure Sensor Using Analytical Method As shown in Fig. 1, the state of charge in terms of battery percentage and SG gravity is measured using a short air-purge system with integrated MEMS differential silicon pressure sensor. Minute pressure is developed from 215 to 246 Pa in the battery. We can use Pascal’s hydrostatic principle to measure SG of the acid solution and estimate battery SOC [7]. In the present study, graphene-based capacitive sensor is simulated to which pressure sensing tube can be connected and can get an added advantage over previous differential MEMS sensor.

2.1 Analytical Design of Graphene Capacitive Sensor A capacitive pressure sensor is designed analytically and this design is afterwards simulated in COMSOL. When pressure is applied to the diaphragm, it deflects at the centre. The change in the displacement of a thin plate causes a change in capacitance. This capacitive sensor consists of two plates made from graphene material. The thickness of the circular graphene diaphragm is 10 µm. The gap between the two

Fig. 1 Schematic diagram of SG measurement system

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

291

plates is maintained at 3 µm. The Poisson’s ratio of graphene is 1000 GPa. To get linearity, the diaphragm’s deflection should be 25–30% of the gap between the two plates. The sensor is designed for 0–1 kPa range. When a battery gets charged or discharged, the change in pressure is 215–246 Pa as per the hydrostatic pressure principle. The sensor is designed for the measurement of tiny variation in differential pressure. The radius of the sensor is calculated using the equation: / R=

4

Wmax × 4.13 × Eh 3 ( ) 1 − v2 × p

(1)

In the above Eq. (1), R denotes radius of the circular diaphragm (µm). Wmax is the maximum displacement of diaphragm at centre (µm), E represents Young’s modulus of elasticity (N/m2 ), h is the thickness of the diaphragm (µm), v denotes Poisson’s ratio and p represents applied pressure measured in the unit (bar). This sensor is designed for 0–1 kPa using graphene material. The linearity of the deflection of the diaphragm and the input applied pressure have prime importance. For linear displacement of the diaphragm, its maximum displacement at the centre should be 25–30% of the gap between the two plates (3 µm). In this design, 28% deflection of the diaphragm has been considered. Wmax indicates maximum displacement of diaphragm at the centre which is equal to 0.84 (µm), E stands for Young’s modulus of elasticity of 1000 GPa. h represents thickness of the diaphragm of 10 µm and v is Poisson’s ratio having value 0.17. p signifies pressure applied to the thin plate of 0.01 bar. The radius of the thin graphene plate is calculated using Eq. (1), the radius is 0.002432 µm for the pressure 0.01 bar. The displacement of the diaphragm is calculated using the equation: Wmax

) ( 1 − v2 × R4 × p = 4.13 × E × h 3

(2)

The maximum stress is calculated using the equation: σmax =

1.25 × p × R 2 h2

(3)

In the above Eq. (3), σmax represents maximum stress that can be sustained by thin diaphragm. R denotes radius of the circular diaphragm of 0.002432 µm. h means thickness of the diaphragm of 10 µm. P signifies pressure applied of 0.01 bar. As per the above Eq. (3), the maximum stress calculated is 7.54 MPa. The voltage applied between the two capacitive plates is 1 V. As shown in Fig. 2, voltage is distributed across the thin plate. For different applied loads, the voltage distribution in a thin plate is different. The change in the capacitance is calculated using an equation C=

∈o ∈r A d

(4)

292

Y. Adhav et al.

Fig. 2 Voltage distribution in the thin plate

In the above Eq. (4), ∈o denotes relative permittivity of air (8.854×10−14 F/cm). ∈r represents relative dielectric constant of medium. A means overlapping area between the plates and d stands for distance between the plates. The change in the capacitance is 2.945 × 10−11 pF for 100 Pa pressure. For 1000 Pa pressure, the change in the capacitance is 3.04 × 10−11 pF.

3 Design of MEMS Capacitive Pressure Sensor Using a Finite Element Method in Comsol Multiphysics Comsol Multiphysics uses the following main component to simulate the sensor environment. Parameters: In the Comsol simulation, initially, parameters have to be set where the operating condition of the sensor has to be mentioned. Geometry: To make the geometry, a 3D block is used with four layers of dimensions of 0.7 mm, 0.397 mm, 0.003 mm, 0.01 mm, respectively, to get an input port, sensor plate thickness, the gap between two capacitive plates and the base of sensor respectively. The cylinder is used to get a circular diaphragm of the capacitive sensor and circular port at one end. An average component coupling is used to compute the average of an expression over the selected geometric entities, which is the underside area of the circular graphene plate. Integration component coupling is used with a point as the source having maximum displacement at the centre. The collection of the boundaries done by XZ and YZ symmetries is used to specify the symmetry of boundary condition in physics interface. The first explicit domain is used for cavity selection, which is the

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

293

area between two main plates of the capacitive sensor. The second explicit is used to select all domains to expect the cavity to which linear elastic properties are assigned. Physics: Electromechanics physics is used to apply pressure and provide excitation to the plates and assign boundary conditions. Electromechanics physics equation form for stationary analysis is used as mentioned below. −∇ · σ = Fv

(5)

∇ · D = eq

(6)

The prescribed displacement is the constraint in Z-axis. The prescribed displacement node adds a condition prescribed in one direction and solid free to deform in the other directions. Boundary load is used to apply pressure to the top plate of the sensor. Terminal and ground are used to provide excitation to the sensor. Mesh: When any sensor is fabricated, it has an irregular shape; hence to solve the finite element equation, meshing is done by creating a mesh of different sizes and shapes. Meshing decides the area under tension and displacement that is the membrane of the sensor. The mapped mesh is used to lower the surface of the model. The maximum mesh element size is 4.88 mm. Auxiliary sweep is used to apply a pressure range. Study: This is used to observe stress and displacement and change in capacitance of the diaphragm due to the change in pressure.

3.1 Geometry Building of Sensor and Material Selection As shown in Fig. 3, cut section of the capacitive pressure sensor consists of a thin plate having a thickness of 10 µm. The radius of the diaphragm is 0.002432 µm. The second plate is made of graphene material having a thickness of 0.397 mm. Between the two plates, space is evacuated by vacuum. This cavity has a relative permittivity equal to 1. This pressure sensor is supported by borosilicate glass. The thickness of the borosilicate glass base is 0.7 mm. The circular port is etched out of 0.41 mm to enter air pressure towards the diaphragm. As shown in Table 1, the properties of the material are entered in the material contents.

3.2 Selection of Physics and Meshing This simulation has been undertaken under the electromechanics interface. It comes under structural mechanics. It combines solid mechanics and electrostatics. Electrostatically actuated structures deformation is modelled with moving mesh. As shown

294

Y. Adhav et al.

Fig. 3 Sensor geometry

Table 1 Material properties Sr. No.

Parameters

Silicon

Graphene

Borosilicate

1

Relative permittivity

11.7

2.14

4.8

2

Density

2330

2000

2230

kg/m3 GPa

3

Young’s modulus

170

1000

63

4

Poisson’s ratio

0.06

0.17

0.20

Unit

in Fig. 4, free tetrahydral meshing has been carried out for the capacitive pressure sensor. Fig. 4 Free tetrahydral meshing

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

295

Fig. 5 Pressure versus capacitance-silicon

4 Results and Discussion 4.1 Change in Capacitance for 0–6 kPa Pressure Range As shown in Fig. 5, when 0–6 kPa pressure is applied to silicon diaphragm, the change in capacitance changes from 2.3 × 10−12 to 3 × 10−12 pF. In Fig. 6, when input is applied to graphene diaphragm, the change in capacitance is from 4.875 × 10−12 to 4.95 × 10−12 pF. The change in graphene material finds more linear as compared to silicon.

4.2 Change in the Displacement of Diaphragm for Pressure Range 0–6 kPa As shown in Fig. 7, a thin plate deflects when the applied pressure is increased and vice versa. As shown in Fig. 8, red colour indicates more deflection of thin plate and blue colour indicates less deflection than the central deflection. The first sensor model is designed for pressure range 0–6 kPa. Change in the displacement of graphene and the silicon is linear. As shown in Fig. 9 in the case of graphene, the displacement changes from 0 to 0.32 µm. As shown in Fig. 10 in the case of silicon, the displacement changes from 0 to 1.8 µm. The difference in displacement in silicon is more than the graphene diaphragm displacement. The

296

Y. Adhav et al.

Fig. 6 Pressure versus capacitance-graphene

Fig. 7 Displacement of sensor

deflection of the graphene material is less than that of silicon because the graphene strength is more, so that this material can handle more substantial applied pressure. Also when the range changes, the radius of sensor differs and shows different displacement.

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

297

Fig. 8 Graphene thin plate

Fig. 9 Applied pressure versus graphene displacement

4.3 Capacitance and Diaphragm Displacement of Graphene for 0–1 kPa Finally, the sensor model is designed for pressure range 0–1 kPa. Change in the displacement of graphene diaphragm is linear. As shown in Fig. 12 in the case of graphene, the displacement changes from 0 to 0.6 µm. The displacement of the

298

Y. Adhav et al.

Fig. 10 Applied pressure versus silicon displacement

diaphragm is different for different ranges because the radius of the diaphragm changes as per the change in range. As shown in Fig. 11 in the case of graphene, the capacitance changes from 2.935 × 10−12 to 3.025 × 10−12 pF.

Fig. 11 Pressure versus capacitance

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

299

Fig. 12 Pressure versus displacement

4.4 Comparison of the Displacement of the Diaphragm by Analytical Method and Finite Element Method The load is applied to the thin material in the range of 0–1 kPa in the step of 100. The deflection of the diaphragm is calculated using Eq. (7) and calculated values and observed values of displacement and capacitance are entered in Table 2. Wmax

) ( 1 − v2 R4 p = 4.13Eh 3

(7)

As shown in Fig. 13, the analytical results are compared to the results obtained by Comsol Multiphysics software. The displacement of the diaphragm is linear with the applied pressure. The sensitivity of the designed sensor is 8.3 × 10−4 as per the finite element method and as per the analytic approach, it is 6.4 × 10−4 .

4.5 Diaphragm Displacement at the Different Thicknesses As shown in Fig. 14, pressure is applied to the diaphragm having thickness 0.010– 0.001 mm. To increase the sensitivity, thickness of the diaphragm is minimised and readings are entered in Table 3, but the linearity between the applied pressure and diaphragm decreases. In this design, the diaphragm’s thickness has been fixed and kept 0.010 mm to get good linearity and sensitivity.

300 Table 2 Displacement of diaphragm and capacitance

Fig. 13 Analytical method versus finite element method

Fig. 14 Diaphragm displacement at different thickness

Y. Adhav et al. Pressure (Pa)

Calculated displacement (µm)

Observed displacement (µm)

Capacitance (pF) × 10−11

100

0.08

0.06

2.945

200

0.16

0.13

2.955

300

0.25

0.19

2.965

400

0.33

0.25

2.975

500

0.41

0.31

2.985

600

0.50

0.37

2.995

700

0.58

0.44

3.005

800

0.66

0.5

3.02

900

0.75

0.56

3.03

1000

0.83

0.64

3.04

Design and Simulation of Capacitive Pressure Sensor for Monitoring …

301

Table 3 Displacement of diaphragm and applied pressure at different thickness Pressure (Pa)

Thickness (mm) 0.001 mm

0.003 mm

0.005 mm

0.007 mm

0.010 mm

100

2.97

2.96

2.95

2.95

2.94

200

3.02

2.99

2.97

2.96

2.95

300

3.07

3.02

2.99

2.98

2.96

400

3.13

3.06

3.02

2.99

2.97

500

3.2

3.1

3.04

3.01

2.98

600

3.28

3.15

3.07

3.03

2.99

700

3.38

3.2

3.09

3.04

3.00

800

3.52

3.25

3.12

3.06

3.01

900

3.7

3.32

3.16

3.08

3.02

1000

3.99

3.40

3.19

3.1

3.03

5 Conclusion The first sensor model is designed for the pressure range of 0–6 kPa. The change in the displacement of graphene and the silicon is linear. The difference in displacement in silicon is more than the graphene diaphragm displacement. Finally, the sensor model is designed for the pressure range of 0–1 kPa. The change in the displacement of graphene diaphragm is linear. In the case of graphene, the displacement changes from 0 to 0.6 µm whereas in graphene, the capacitance changes from 2.935 × 10−11 to 3.025 × 10−11 pF. The sensitivity of the designed sensor is 8.3 × 10−4 as per the finite element method and as per the analytic approach, it is 6.4 × 10−4 pF. In this design, the diaphragm’s thickness has been kept 0.010 mm to get good linearity and sensitivity. In this capacitive sensor design, only one port has been made to allow the applied pressure to a thin plate. For the lead-acid battery to measure differential pressure, a sensor having two ports is needed. We can use a single port sensor, but this will measure the gauge pressure. As long as the battery’s acid level is constant, this sensor shows output proportional to the electrolyte’s specific gravity. After a few months in a lead-acid battery, the level of acid changes due to acid evaporation, so it needs a differential pressure sensor to measure specific gravity. As per the hydrostatic principle, if the head is constant, the differential pressure is directly proportional to the specific gravity electrolyte. The output of this sensor is the change of capacitance. Most of the microcontrollers require the input signal in voltage form. So, if the capacitance to voltage converter is designed on the same platform, it could be easier to present the signal in the form of a state of charge battery. The first sensor model is designed for the pressure range of 0–6 kPa. Change in the capacitance of graphene is linear as compared to the silicon. As shown in Fig. 6 in the case of graphene, the capacitance changes from 4.875 × 10−12 to 4.955 ×

302

Y. Adhav et al.

10−12 pF. As shown in Fig. 5 in the case of silicon, the capacitance changes from 2.3 × 10−12 to 3.5 × 10−12 pF. Acknowledgements The authors thank for the financial support provided by the Instrumentation and Control Department, College of Engineering, Pune (C.O.E.P) for patent filing and further process on the research work.

References 1. Degla A, Chouder MCA, Bouchafaa F, Taallah A (2018) Update battery model for photovoltaic application based on comparative analysis and parameter identification of leadacid battery models behaviour. IET Renew Power Gener 12(4):484–493 2. Dinh THN, Joubert P, Martincic E, Dufour-Gergam E (2014) Flexible 3-axes capacitive pressure sensor array for medical applications. In: SENSORS, 2014. IEEE, pp 855–858 3. Hancke GP (1990) A fiber-optic density sensor for monitoring the state-of-charge of a lead acid battery. IEEE Trans Instrum Meas 39(1):247–250 4. Lin L, Chu H-C, Lu Y-W (1999) A simulation program for the sensitivity and linearity of piezoresistive pressure sensors. J Microelectromech Syst 8(4):514–522 5. Marcos-Acevedo J, Cao-Paz A, del Rio-Vazquez A, Quintans-Grana C, Martinez-Penalver C (2009) Density measurement into lead-acid batteries with multi-point optical fiber sensor. In: 2009 IEEE instrumentation and measurement technology conference, pp 715–718 6. Neto PBL, Saavedra OR, de Souza Ribeiro LA (2018) A dual-battery storage bank configuration for isolated microgrids based on renewable sources. IEEE Trans Sustain Energy 9(4):1618– 1626 7. Patil CY, Adhav YG (2020) A lead acid battery charge monitoring system. The patent office journal no. 15/2020 (application no. 201821037211 A), P–16896, 10 Apr 2020 8. Sathyanarayanan S, Juliet AV (2010) Design and simulation of touch mode MEMS capacitive pressure sensor. In: 2010 international conference on mechanical and electrical technology, pp 180–183 9. Zhang Y, Gui Y, Meng F, Gao C, Hao Y (2016) Design of a graphene capacitive pressure sensor for ultra-low pressure detection. In: 2016 IEEE 11th annual international conference on nano/micro engineered and molecular systems (NEMS), pp 192–195 10. Palasagaram JN, Ramadoss R (2006) MEMS-capacitive pressure sensor fabricated using printed-circuit-processing techniques. IEEE Sens J 6(6):1374–1375 11. y Paz AMC, Acevedo JM, Gandoy JD, del Rio Vazquez A (2006) Multisensor fibre optic for electrolyte density measurement in lead-acid batteries. In: IECON 2006—32nd annual conference on IEEE industrial electronics, pp 3012–3017 12. Unigarro E, Bohrquez JC, Achury A, Ramirez F, Sacristn J, Segura-Quijano F (2017) Differential capacitive pressure sensor design based on standard CMOS. Electron Lett 53(11):737–739 13. Sharma A, Singh J (2013) Design and analysis of high performance MEMS capacitive pressure sensor for TPMS. In: 2013 international conference on control, automation, robotics and embedded systems (CARE), pp 1–5

Development of Screw Press-Dewatering Unit for Biogas Slurry Madhuri More , Chitranjan Agrawal, and Deepak Sharma

Abstract The present research work aims to develop the screw press-dewatering unit of 50 kg per hour capacity for biogas slurry as waste and produces organic solid–liquid fertilizers. The performance effect on moisture content in biogas slurry and shaft speed of developed unit in terms of dewatering efficiency and dewatering rate was evaluated. The result showed the higher dewatering efficiency of 81.82% and dewatering rate of 49.38 kg per hour. The properties of separated solid fertilizer from raw biogas slurry were determined such as total solids content of 10.7%, ammoniacal nitrogen of 2.6 g per kg, total nitrogen content of 4.9 g per kg, phosphate content of 5.0 g per kg, and potassium content of 3.0 g per kg, respectively. In liquid, fertilizers were obtained such as total solids content of 4.07%, ammoniacal nitrogen of 3.1 g per kg, total nitrogen content of 5.8 g per kg, phosphate content of 2.2 g per kg, and potassium content of 5.8 g per kg, respectively. The organic solid fertilizer can be used as phosphorous-rich fertilizer for improving soil characteristics, and liquid fertilizers can be used as pesticide and fungicide to control pests on crop. The economic analysis shows that the manufacturing cost of whole process was 2325.35 US$ and the payback period was 0.88 years. Keywords Screw press-dewatering · Biogas slurry · Solid fertilizer · Liquid fertilizer

1 Introduction Biogas is efficient renewable source to reduce greenhouse gas (GHG) emissions and waste disposal [1]. In recent years, biogas plant (132,000 BP) is installed in the world, having a hug potential for biogas production. These biogas plants are M. More (B) · D. Sharma Department of Renewable Energy Engineering, CTAE, MPUAT, Udaipur, Rajasthan 313001, India e-mail: [email protected] C. Agrawal Department of Mechanical Engineering, CTAE, MPUAT, Udaipur, Rajasthan 313001, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_20

303

304

M. More et al.

used to produce biogas as well as biogas slurry. Biogas slurry production increased toward the years due to increasing demand in developing countries. From all available manure, 10 billion tons of nutrient-rich biogas slurry is produced under the process of anaerobic digestion and used as organic manure for production of crops [2]. In India, Ministry of New and Renewable Energy (MNRE) installed 12,019 number of biogas plants in the year 2019–20 and up to 78 MT of biogas slurry was produced per year [3]. Currently, increasing use of chemical fertilizers in farming systems creates environmental issues due to improper use of fertilizers, which includes an unacceptable percentage of nitrogen and phosphorous being released into soil and ammonia losses to the atmosphere [4]. Biogas slurry produced from the biogas plant has good amount of nutrients, such as macro (NPK) and micronutrients; therefore, it is very useful to the crop production as a valuable organic fertilizer [5, 6]. Biogas slurry is an inexpensive way to reduce the use of inorganic fertilizers in the soil and hence reduces the nature pollution [7]. Nowadays, the production of biogas slurry is increasing; however, there is need to store biogas slurry. Most of the farmers does not have sufficient field for proper storage of biogas slurry. Therefore, they directly spread biogas slurry on their own field [8]. The higher production of biogas slurry would pollute water bodies and environmental issues, i.e., nitrogen losses and ammonia losses because of improper management of biogas slurry waste [9]. The processing of biogas slurry is very costly and energy consuming [10]. For the proper utilization of biogas slurry, dewatering is a common method, which helps to reduce transportation, handling, and storage costs. Long-distance transportation justifies economically investment on slurry upgradation for which dewatering is commonly adopted process [11]. This process helps to separate solid–liquid fertilizers from biogas slurry. Dewatering of biogas slurry is a process to dewatered liquid fraction from biogas slurry and to separate solid–liquid fraction. This process can reduce nutrient losses such as nitrogen (N), phosphorus (P), and potassium (K) to the environment and manage the use of organic fertilizer. The separated solid fraction can be managed differently on-farm for slurry utilization and crop fertilization [12, 13]. Application of separated fractions to the soil can reduce the risk of the polluting soil surface and groundwater along with greenhouse gas emissions [10]. There are several techniques for the separation of biogas slurry like screw press, centrifugation, sieving or screening, belt filters, and sedimentation or chemically enhanced sediment settling [8]. Generally, screw press technology can be used to produce high dry matter (solid fraction) and low dry matter (liquid fraction). Al Seadi et al. [14] found that the solid and liquid partitioning divides most of the nitrogen in liquid fraction and phosphorous in solid fraction, which helps the plant nutrients management for better crop production. The phosphorous-rich organic solid fertilizer can be applied as phosphorous-rich fertilizer for improving soil characteristics. The major fraction deriving from the dewatering process is organic liquid fertilizers, which contain the part of nitrogen and potassium. This organic liquid fertilizer can be used as pesticide and fungicide [15]. In the previous studies, the dewatering of biogas slurry was carried out by traditional methods with the help of wheat straw, paddy straw, and agricultural waste [16]. However, these traditional methods are labor intensive and

Development of Screw Press-Dewatering Unit for Biogas Slurry

305

time-consuming. Only few investigations have focused on screw press technology, which is used in medium (2 m3 ) to large size (140 m3 ) biogas plants [8, 14]. The screw press technology has gained popularity, especially those farmers who need to export their extra fertilizer to other farms. By considering the scope, biogas slurry has abundant availability and great potential, but slurry utilization is very difficult. The research on utilization of biogas is slurry to produce organic solid–liquid fertilizers as a waste management. Therefore, the present study investigated production of organic solid–liquid fertilizers from biogas slurry in a screw press-dewatering unit of capacity of 50 kg/h. The effect of moisture content of biogas slurry and shaft speed of unit in terms of dewatering efficiency and dewatering rate was evaluated. The detailed characterization of produced organic solid–liquid fertilizers was determined by using standard procedure. The economic analysis of the present work was determined in terms of net present worth, payback period, benefit–cost ratio, and internal rate of returns for the commercialization.

2 Experimental Methods The fresh biogas slurry was collected from available biogas plants such as KVIC and Deenbandhu from the DREE, CTAE, MPUAT, Udaipur, Rajasthan (India). The experimental method was carried out, which is shown in Fig. 1.

2.1 Characterization of Biogas Slurry The properties of fresh biogas slurry were determined using ASTM standards Method 8075, Method 8190, and Method 8049 [17, 18]. The composition of biogas slurry in terms of pH, density, moisture content (MC), volatile solids content (VS), ash content (Ash), total solid content (TS), electrical conductivity (EC), nitrogen (N), phosphorous (P), and potassium (K) content was determined by using Kjeldahl apparatus and spectrophotometer.

2.2 Screw Press-Dewatering Unit The screw press-dewatering unit of capacity of 50 kg per hour was developed for the present research work (Fig. 2). The developed unit consists of hopper, screw press, outlet system, pumping system, and separate solid–liquid collector. The screw pressdewatering unit was made by mild steel having barrel length and inside diameter of 0.64 m and 0.16 m, respectively. The unit was operated with automatic power connection of screw press and slurry feed pump. Helical gearbox and slurry pump

306 Fig. 1 Experimental methodology for dewatering process

M. More et al.

Biogas Slurry

Screw Press Unit

Solid Fertilize Fertilizer

Liquid Fertilize Fertilizer

are connected to the panel system of the developed unit. The helical gear mechanism was used to reduce the speed of rotation. Slurry pump was used to transfer biogas slurry from outlet of biogas plant to the hopper. A motor was used to supply the power to rotate the shaft at variable speed controlled by a speed controller. The specifications of the screw press unit are shown in Table 1.

2.3 Experimental Procedure The performance evaluation of the developed unit was analyzed by moisture content of biogas slurry and shaft speed. Total nine sets of experiment with three replications of each set were performed to evaluate the performance of screw press unit in terms

Development of Screw Press-Dewatering Unit for Biogas Slurry

307

Fig. 2 Screw press-dewatering unit

of dewatering efficiency and dewatering rate. The screw press-dewatering unit test was performed at different moisture content (MC) of biogas slurry, i.e., 90–95%, 85–90%, and 80–85%, and shaft speed of 30 rpm, 35 rpm, and 40 rpm, respectively. The experiment was conducted with the following set of treatments: T 1 —90– 95% MC in biogas slurry and 30-rpm speed; T 2 —85–90% MC in biogas slurry and 30-rpm speed; T 3 —80–85% MC in biogas slurry and 30-rpm speed; T 4 —90–95% MC in biogas slurry and 35-rpm speed; T 5 —85–90% MC in biogas slurry and 35rpm speed; T 6 —80–85% MC in biogas slurry and 35-rpm speed; T 7 —90–95% MC in biogas slurry and 40-rpm speed; T 8 —85–90% MC in biogas slurry and 40-rpm speed; and T 9 —80–85% MC in biogas slurry and 40-rpm speed, respectively. Dewatering efficiency was determined by the ratio of the total weight of liquid dewatered from the biogas slurry to the total weight of biogas slurry by using equation [4, 19]; ED =

W1 × 100 W2

(1)

Dewatering rate represents the quantity of liquid dewatered from the fresh biogas slurry per unit time expressed by using equation [4, 11]; DR =

W1 × 100 T

(2)

where E D is the dewatering efficiency in percentage, DR is the dewatering rate in kg per hour, W 1 is the weight of liquid dewatered from the slurry in kg, W 2 is the total weight of slurry (kg), and T is the time taken for dewatering (h).

308 Table 1 Specification of screw press-dewatering unit

M. More et al. Parameters

Values

Capacity

50 kg/h

Diameter of the first screw

0.15 m

Diameter of the last screw

0.08 m

Pitch of screw

0.09 m

Tapered angle

3°

Speed of rotation

30, 35, and 40 rpm

Motor

0.746 kW, 1440 rpm

Slurry pump

0.746 kW, 6 m head

Bulk-density raw material

1220 kg/m3

Theoretical screw volume

0.0015 m3

Mass flow rate

7.10 kg/min

Helix angle

10.81°

Screw conveyor length

0.635 m

Number of threads

7

Volume per pitch

0.042 m3 /h

Torque on the screw

402.78 Nm

Power on shaft

0.70 kW

Pressure created on a barrel

7.24 MPa

Volume of a barrel on a screw

0.012 m3

Diameter of the screw shaft

40 mm

Load lifted by the screw

42.82 kN

Pressure to be developed by the screw

2.59 MPa

Pressing area

0.1649 m2

Compression ratio

20:1

Capacity of screw press

531.9 g/h

Power to drive screw press

0.75 kW

Specific energy consumption

0.0022 kWh/kg

2.4 Statistical Analysis Analysis of variance (ANOVA) was applied to analyze the data using randomized block design (RBD) and compared by Design Expert-11 software at the 5% level of significance in the Department of Statistics, RCA, MPUAT, Udaipur.

Development of Screw Press-Dewatering Unit for Biogas Slurry

309

2.5 Characterization of Separated Solid–Liquid Fertilizer The organic solid–liquid fertilizers were selected for the property analysis. The total solids content (TS), ammoniacal nitrogen (NH4 -N), total nitrogen (TN-Method 8075), phosphate content (P2 O5 -Method 8190), and potash content (K-Method 8049) were determined by using Kjeldahl apparatus and spectrophotometer [14, 18].

2.6 Economic Assessment The economic feasibility of production of separated solid–liquid by dewatering process using 50 kg per hour capacity unit was analyzed by the following parameters [20–22] i. ii. iii. iv.

Net present worth (NPW) Benefit–cost ratio (BCR) Payback period (PP) Internal rate of return (IRR).

3 Result and Discussion 3.1 Characterization of Biogas Slurry The characterization of biogas slurry including physicochemical properties is presented in Table 2. The biogas slurry contains 85.22% of moisture and high percentage of volatile matter content of 73.53%, which is suitable for the dewatering process. The physicochemical analysis of biogas slurry has good NPK content, which is much suitable for increasing crop productivity. The overall amount of nutrient content available in biogas slurry is sufficient for good crop growth [17].

3.2 Effect of Moisture Content in Biogas Slurry and Shaft Speed on Dewatering Efficiency (ED ) The performance of developed unit was evaluated in terms of moisture content of biogas slurry and shaft speed (Table 3). The experiment was carried out with nine treatments of biogas slurry having moisture content (MC) in the range of 90–95, 85–90, and 80–85% and speed of rotation as 30, 35, and 40 rpm. All treatments were arranged randomly with three replications for each treatment. Observations were recorded for dewatering efficiency at various level of moisture content and

310 Table 2 Physicochemical analysis of biogas slurry

M. More et al. Sr. No.

Parameter

Value

1.

pH

8.5

2.

Electrical conductivity (dS m−1 )

3.0–3.2

3.

Moisture content (%)

85.22

4.

Total solid content (%)

14.77

5.

Volatile solid content (%)

73.53

6.

Ash content (%)

26.47

7.

Nitrogen content (%)

1.51

8.

Phosphorus content (%)

0.77

9.

Potassium content (%)

0.81

different screw shaft speed. Figure 3 shows that the dewatering efficiency of developed unit was increased with decreasing moisture content and increasing shaft speed. Overall results concluded that Treatment 9 obtained maximum dewatering efficiency of 81.82% at 80–85% MC in biogas slurry and 40-rpm speed. The results’ data compared to Oladipo et al. [19] tested screw press unit with cassava mash, and they have founded an average dewatering efficiency of 38.15% at a 66.75% MC (w.b.) for fermented mash and 37.68% at a 70% MC (w.b.) for unfermented mash. The data were analyzed to check effect on dewatering efficiency of performance of developed unit by using ANOVA and the data compared to Design Expert-11 software. The observation shows that the 80–85% moisture content of biogas slurry and 40-rpm shaft speed came out to be significant treatment at 5% level of significance for evaluation of dewatering efficiency. The generated response surface diagram between moisture content (A) and shaft speed (B) clearly indicates that the decrease in moisture content and increase in speed of shaft gave better dewatering efficiency as shown in Fig. 4. It was observed that the developed screw press-dewatering unit is best suited for the performance and dewatering efficiency varies in the range of 78.01– 81.82%. The range being higher than the CD value points toward the significant effect in dewatering efficiency due to the moisture content and speed of shaft (Table 4). Table 4 shows that the Model F-value of 14.24 suggest that the model is significant and P-values less than 0.0266 indicate model terms are significant. Figure 5 indicates that the model is fit to relate the predicted and actual values and that the number of experimental combinations formed in the design of experiment was enough to find out the effect of independent variables (i.e., moisture content and speed) for dewatering efficiency.

48.84

DR (kg h−1 )

2.

T2

49.38

81.72

T3

39.26

81.53

T4 33.25

78.01

T5 32.66

78.94

T6 33.94

80.42

T7 32.3

79.49

T8 32.7

81.19

T9 30.24

81.82 49.38

81.82

Max.

30.24

78.01

Min.

18.84

3.81

Range

36.99

80.46

Mean

7.27

1.35

S.D.

0.7

0.42

SEM ±

2.42

1.24

CD (5%)

19.66

0.9

CV (%)

DE: dewatering efficiency, DR: dewatering rate, Max: Maximum value, Min: Minimum, S.D.: standard deviation, SEM ±: standard error, CD: critical difference, CV: coefficient of variation

T1

81.01

DE (%)

1.

Treatments

Parameters

Sr. No.

Table 3 Performance evaluation of screw press-dewatering unit at different moisture content and speed of shaft

Development of Screw Press-Dewatering Unit for Biogas Slurry 311

312

M. More et al.

Dewatering Efficiency, %

82 81 80 79 78 77 76 T1

T2

T3

T4 T5 T6 Treatments

T7

T8

T9

Fig. 3 Performance evaluation of moisture content and shaft speed on dewatering efficiency of developed unit

Fig. 4 Statistical presentation of generated response surface diagram between moisture content (A) and shaft speed (B) on dewatering efficiency of developed screw press-dewatering unit

Development of Screw Press-Dewatering Unit for Biogas Slurry

313

Table 4 Test analysis of dewatering efficiency (E D ) of developed screw press-dewatering unit Source

Sum of squares

df

Mean square

F-value

p-value

Quadratic model

14.11

5

2.82

14.24

0.0266

A-MC

4.61

1

4.61

23.27

0.0170

B-Speed

0.5087

1

0.5163

2.61

0.2049

AB

0.8190

1

0.8190

4.13

0.1350

A2

0.1089

1

0.1089

0.5495

0.5053

B2

8.05

1

8.05

40.64

0.0078

Residual

0.5945

3

0.1982

Cor. total

14.70

8

Significant

Fig. 5 Plot diagram of predicted versus actual on dewatering efficiency of developed screw pressdewatering unit

3.3 Effect of Moisture Content in Biogas Slurry and Shaft Speed on Dewatering Rate (DR) Observations were recorded for dewatering rate at various levels of moisture content and shaft speed. The dewatering rate of the developed unit was determined by the ratio of weight of raw material to the processing time. It can also be observed that the variation in dewatering rate of developed unit occurs due to variation in processing time.

M. More et al.

DR, kg/h

314

50 45 40 35 30 25 20 15 10 5 0 T1

T2

T3

T4 T5 T6 Treatments

T7

T8

T9

Fig. 6 Performance evaluation of moisture content and shaft speed on dewatering rate of developed unit

The processing time was decreased with increasing dewatering rate. Figure 6 shows that the dewatering rate of developed unit was increased with increasing moisture content in biogas slurry and decreasing shaft speed. From Fig. 6, the mean dewatering rate ranged from 30.54 to 49.38 kg h−1 . It can be observed that Treatment 2 can achieve the maximum dewatering rate of 49.38 kg h−1 for 85–90% moisture content of biogas slurry and 30-rpm shaft speed. The data were analyzed by ANOVA, compared by Design Expert-11 software, and concluded that the moisture content of biogas slurry and speed of shaft has a significant effect on dewatering rate at 5% level of significance. Table 5 shows that the Model F-value of 8.43 suggest that the model is significant and P-values less than 0.0181 indicate model terms are significant. Figure 7 shows that the generated response surface diagram between moisture content (A) and shaft speed (B) clearly indicates that the increase in moisture content and decrease in speed of shaft gave better dewatering rate. Figure 8 indicates that the model is fit to relate the predicted and actual values and that the number of experimental combinations formed in the design of experiment was enough to find out the effect of independent variables (i.e., moisture content and speed) for dewatering rate.

3.4 Characteristics of Organic Solid–Liquid Fertilizers The properties of organic solid–liquid fractions were carried out by respective ASTM standard methodology of total solids content (TS), ammoniacal nitrogen

Development of Screw Press-Dewatering Unit for Biogas Slurry

315

Table 5 Test analysis of dewatering rate of developed screw press-dewatering unit Source

Sum of squares

df

Mean square

Linear model

311.36

2

155.68

A-MC

F-value

p-value

8.43

0.0181

18.76

1

18.76

1.02

0.3524

B-Speed

292.60

1

292.60

15.84

0.0073

Residual

110.82

6

18.47

Cor. total

422.18

8

Significant

Fig. 7 Statistical presentation of generated response surface diagram between moisture content (A) and shaft speed (B) on dewatering rate of developed screw press-dewatering unit

(NH4 -N), total nitrogen (TN), phosphate content (P2 O5 ), and potassium content (K), respectively. Ammoniacal nitrogen (NH4 -N) was determined for decreasing nitrogen losses in separated solid and liquid fractions through ammonia volatilization and nitrogen leaching [23]. Total phosphate content of separated fractions was higher in organic solid fertilizer as compared to liquid fertilizer, which is very helpful for crop productivity. The ammoniacal nitrogen content, total nitrogen content, and potassium content were higher in organic liquid fertilizers due to increased ammonia content in liquid followed by dewatering process. The dewatered organic liquid fertilizers were used as pesticide and fungicide for controlling pest and fungus on crop growth [15].

316

M. More et al.

Fig. 8 Plot diagram of predicted versus actual on dewatering rate of developed screw pressdewatering unit

It can be also recycled into biogas plant for increasing biogas production in short period. Table 6 represents the properties of organic solid–liquid fertilizers produced from the dewatering of biogas slurry by using screw press-dewatering unit. The properties of organic solid fertilizers were determined such as total solids content of 10.7%, ammoniacal nitrogen of 2.6 g/kg, total nitrogen content of 4.9 g/kg, phosphate content of 5.0 g/kg, and potassium content of 3.0 g/kg, respectively. In the separated solid fertilizer, agriculture waste, neem leaves, and earthworm are added for increasing nutritive values and good quality compost. In organic liquid, fertilizers were obtained such as total solids content of 4.07%, ammoniacal nitrogen of 3.1 g/kg, total nitrogen of 5.8 g/kg, phosphate content of 2.2 g/kg, and potassium content of 5.8 g/kg, respectively. The dewatered liquid is mostly used as pesticide after the up-gradation and value addition process. Moller et al. [21] found that the dewatering of slurry reduces transport cost by separating the manure into a nutrient-rich solid–liquid fraction when the solid nutrient-rich phase is over high distances and liquid fraction is transported over low distances. The result compared to Table 9 concluded that the production of organic solid–liquid fertilizers is suitable for increasing crop productivity.

Development of Screw Press-Dewatering Unit for Biogas Slurry

317

Table 6 Distribution of nutrients in organic solid–liquid fertilizers Separated fractions

TS (%)

TN (g/kg)

NH4 -N (g/kg)

P2 O5 (g/kg)

K (g/kg)

Solid fraction

10.7

4.9

2.6

5.0

3.0

Liquid fraction

4.07

5.8

3.1

2.2

5.8

3.5 Economic Analysis To evaluate the economic analysis of dewatering process, some assumptions were made. The working life of developed screw press-dewatering unit was assumed 10 years, and the operation periods were considered 300 days in 365 days. The capital cost of developed system is 1722.48 US$ (1 US$ = 69.67 INR as of June 30, 2019). The capital cost includes the price of helical gear and mud pump. The repair and maintenance cost of developed unit is considered as 3% of capital cost. The developed unit can process 50 kg per hour of biogas slurry and calculate production cost of separated solid–liquid fractions. By selling the separated solid and liquid, the capital investment of developed screw press unit can be recovered within 0.88 years. The benefit–cost ratio of 1.59 specifies that the developed unit is economical for dewatering of biogas slurry. However, the unit capacity, types of raw material, and the dewatering efficiency of developed unit have direct impact on the production cost of organic solid–liquid fertilizers. The result of the present study compares with Kumar et al. [22], and it was found that the unit screw press unit advantage is that it is economically cost-effective than the rotary screen separator. Sievers et al. [24] performed techno-economic assessment of solid–liquid separator, and it was performed for 2000-T slurry per day and production cost is very high as compared to the present study. Tables 7 and 8 show the economic analysis and economic measures of the developed screw press-dewatering unit for production of organic solid–liquid fertilizers by dewatering process.

3.6 Comparison of Present Work and Literatures This section presents the comparison of literature related to the solid–liquid separation of biogas slurry and presents research work. Many researchers reported that the biogas slurry has potential to produce organic manure, which is rich in nutritive values. Khan et al. [25] studied on characterization of biogas slurry and concluded that the use of biogas slurry is good for increasing crop yield. Yu et al. [5] also studied biogas slurry and its properties for soil fertility for tomato crop. Hjorth et al. [11] discussed various slurry separation techniques such as sedimentation, centrifugation, and pressurized filtration (i.e., screw press) for solid–liquid separation and production of organic fertilizers. It was observed that the screw press technology for solid–liquid separation is good for recycling of organic matter and produced plant nutrients could mitigate environmental pollutions. The output of separation is

318 Table 7 Economic feasibility for the developed screw press-dewatering unit

M. More et al. Operating parameter/costs

Output

Production capacity

50 kg h−1

Operating days per year

300 days

Operating hours per day

8 h day−1

Capacity of utilization

80%

Raw material

Biogas slurry

Output

96 T year−1

The installed cost of screw press unit (A)

1722.48 US$

Cost Power @ 0.087 US$/kWh/year

258.37 US$

Maintenance includes @ 3% of (A) per year

15.93 US$

Manpower (1 labor) @ 5.74 US$/year

1722.48 US$

Discount rate @ 10% for screw press unit per year

172.25 US$

Total cost (B) per year

2325.35 US$

Return The net sale price of solid manure per tons

71.77 US$

The net sale price of solid manure per year (C) 6889.93 US$ Net annual saving per year (D = C − B)

4564.58 US$

Payback period

0.88

US$ 1.00 is equivalent to approximately 69.67 Indian rupees (cited on, 03 July 2019)

Table 8 Economic measures of the dewatering system

Economic measures

Value

Net present worth (NPW) after 10 years

9390.62 US$

Benefit–cost ratio (BCR)

1.59

Payback period (PP)

0.88 years

Internal rate of return (IRR)

105%

varied with the type of separation technology, type of raw material, and speed of unit applied. So far, very limited literature (small scale) is available on separation or dewatering of biogas slurry and its by-product characterization. The research aimed to study the production of organic solid–liquid fertilizers from biogas slurry by using screw press-dewatering unit. Al Seadi et al. [14] studied characterization of separated organic solid–liquid, which is nearly same as our results, but they studied all dewatering technologies. Aguirre-Villegas et al. [26] concluded that the solid–liquid separation of biogas slurry after digestion reduces GHG emission and increases separation efficiency for manure management. The present research work investigates production of organic solid–liquid fertilizer by using screw press-dewatering technology and characterization of organic solid–liquid fertilizer. Table 9 shows

Development of Screw Press-Dewatering Unit for Biogas Slurry

319

the comparative study of characterization of biogas slurry and separated organic solid–liquid fertilizers.

4 Conclusions In this study, biogas slurry is produced under the process of anaerobic digestion. The biogas slurry is used as organic manure, but it has higher transportation cost and storage cost. Therefore, the screw press unit was developed to separate solid– liquid fraction from biogas slurry by dewatering process. Performance evaluation of developed unit was carried out in terms of dewatering efficiency and dewatering rate of developed unit. The treatment (T 9 ) obtained maximum dewatering efficiency of 81.82% at 80–85% MC in biogas slurry and 40-rpm shaft speed, and treatment (T 2 ) can achieve the maximum dewatering rate of 49.38 kg h−1 for 85–90% moisture content of biogas slurry and 30-rpm shaft speed. It was found that the dependent parameters such as moisture content and shaft speed had major impact on solid– liquid separation. The properties of produced organic solid–liquid fertilizers indicate that it has potential to be used as organic fertilizer and can be up-gradated by using new technologies. Hence, the study concludes that the biogas slurry is utilized by producing organic solid–liquid fertilizers by dewatering process and beneficial as an economic and environment point of view. These produced organic solid–liquid fertilizers can be utilized to provide healthy environment, and ultimately, it will help to reduce nutrient losses and transportation problems. The economy of dewatering process is an important scope to produce organic fertilizers from biogas slurry waste of good quality and quantity in minimum production cost.

Liquid fertilizer

Solid fertilizer

–

–

–

7.64

38.1

52 µs/cm

–

–

9.2

–

–

–

–

2.31

5.7

2.9

1620 µs/cm

–

7.8

–

4.07

–

7.8

27.66

24.3

10.7

–

8.0

4.85

6.5

14.77

5.1

3.1 ds/m

1468 µs/cm

7.9

Biogas slurry

TS (%)

7.7

EC

pH

Characteristics

Table 9 Comparison of present work with literature

1.49%

–

0.98%

–

21.20%

–

0.91%

–

3.39%

–

0.78%

73.53%

VS

0.082%

–

–

–

6.46%

–

–

–

1.46%

–

–

26.47%

Ash

–

1.1%

–

–

23.3%

–

–

–

1.6%

–

Total C

3.49 kg/t

4.9 g/kg

2.63 kg/t

3.0 g/kg

–

3.1 g/kg

−5.8 g/kg 0.24%

4.50 kg/t

2.7 g/kg

–

2.6 g/kg

3.1 kg/t

3.2 g/kg

–

–

NH4 -N

8.15 kg/t

5.8 g/kg

1.34%

4.9 g/kg

4.6 kg/t

5.1 g/kg

0.23%

1.51%

TN

0.31 kg/t

2.3 g/kg

–

2.2 g/kg

6.52 kg/t

5.0 g/kg

–

5.0 g/kg

0.9 kg/t

2.3 g/kg

–

0.77%

P2 O5

–

6.2 g/kg

–

5.8 g/kg

–

5.8 g/kg

–

3.0 g/kg

3.5 kg/t

5.5 g/kg

–

0.81%

K

[14]

[8]

[27]

Present study

[14]

[8]

[27]

Present study

[14]

[8]

[27]

Present study

References

320 M. More et al.

Development of Screw Press-Dewatering Unit for Biogas Slurry

321

References 1. Martinov M, Scarlat N, Djatkov D, Dallemand JF, Viskovic M, Zezelj B (2020) Assessing sustainable biogas potentials—case study for Serbia. Biomass Conv Bioref 10(2):367–381 2. World Biogas Association (WBA) (2019) Global potential of biogas. Available from: https:// www.worldbiogasassociation.org/wpcontent/uploads/2019/09/WBA-globalreport-56ppa4_ digital-Sept-2019.pdf 3. Ministry of New and Renewable Energy. Potential of renewable energy in India. http://mnre. gov.in/file-manager/annual-report/2019-2020. Accessed 17 Sept 2020 4. Cocolo G, Hjorth M, Zarebska A, Provolo G (2016) Effect of acidification on solid–liquid separation of pig slurry. Biosyst Eng 143:20–27 5. Yu FB, Luo XP, Song CF, Zhang MX, Shan SD (2010) Concentrated biogas slurry enhanced soil fertility and tomato quality. Acta Agric Scand Sect B Soil Plant Sci 60(3):262–268 6. Koszel M, Lorencowicz E (2015) Agricultural use of biogas digestate as a replacement fertilizers. Agric Agric Sci Procedia 7:119–124 7. Islam R, Rahman SME, Rahman MM, Deog Hwan OH, Chang RA (2010) The effects of biogas slurry on the production and quality of maize fodder. Turk J Agric 34:91–99 8. Drosg B, Fuchs W, Al Seadi T, Madsen M, Linke B (2015) Nutrient recovery by biogas digestate processing. IEA Bioenergy 711 9. Du H, Gao W, Li J, Shen S, Wang F, Fu L, Zhang K (2019) Effects of digested biogas slurry application mixed with irrigation water on nitrate leaching during wheat-maize rotation in the North China Plain. Agric Water Manag 213:882–893 10. Fangueiro D, Lopes C, Surgy S, Vasconcelos E (2012) Effect of the pig slurry separation techniques on the characteristics and potential availability of N to plants in the resulting liquid and solid fractions. Biosyst Eng 113(2):187–194 11. Hjorth M, Christensen KV, Christensen ML, Sommer SG (2011) Solid–liquid separation of animal slurry in theory and practice. Sustain Agric 2:953–986 12. Fangueiro D, Hjorth M, Gioelli F (2015) Acidification of animal slurry—a review. J Environ Manage 149:46–56 13. Hupfauf S, Bachmann S, Juárez MFD, Insam H, Eichler-Löbermann B (2016) Biogas digestates affect crop P uptake and soil microbial community composition. Sci Total Environ 542:1144– 1154 14. Al Seadi T, Drosg B, Fuchs W, Rutz D, Janssen R (2013) Biogas digestate quality and utilization. In: The biogas handbook, pp 267–301 15. Groot LD, Bogdanski A (2013) Bioslurry: brown gold. A review of scientific literature on the co-product of biogas production. Food and Agriculture Organization of the United Nations (FAO) 16. Kadam ER, Sharma D, Pawar EA (2017) Filtration of biogas spent slurry and its chemical analysis. IJCS 5(3):405–408 17. Dahiya AK, Vasudevan P (1986) Biogas plant slurry as an alternative to chemical fertilizers. Biomass 9(1):67–74 18. Yadav N, Kumar R, Rawat L, Gupta S (2014) Physico-chemical properties of before and after anaerobic digestion of Jatropha seed cake and mixed with pure cow dung. J Chem Eng Process Technol 5(2):1 19. Oladipo NO, Olotu FB, Adamade CA, Agaja MO (2015) Development of NCAM manually operated single pole dewatering press 20. Mukhtar S, Sweeten JM, Auvermann BW (1999) Solid-liquid separation of animal manure and wastewater 21. Moller HB, Lund I, Sommer SG (2000) Solid–liquid separation of livestock slurry: efficiency and cost. Bioresour Technol 74(3):223–229 22. Kumar AK, Sharma S, Dixit G, Shah E, Patel A (2020) Techno-economic analysis of microalgae production with simultaneous dairy effluent treatment using a pilot-scale high volume V-shape pond system. Renew Energy 145:1620–1632

322

M. More et al.

23. Risberg K, Cederlund H, Pell M, Arthurson V, Schnurer A (2017) Comparative characterization of digestate versus pig slurry and cow manure—chemical composition and effects on soil microbial activity. Waste Manage 61:529–538 24. Sievers DA, Tao L, Schell DJ (2014) Performance and techno-economic assessment of several solid–liquid separation technologies for processing dilute-acid pre-treated corn stover. Bioresour Technol 167:291–296 25. Khan SA, Malav LC, Kumar S, Malav MK, Gupta N (2014) Resource utilization of biogas slurry for better yield and nutritional quality of baby corn. Adv Environ Agric Sci 32:382–394 26. Aguirre-Villegas HA, Larson RA, Sharara MA (2019) Anaerobic digestion, solid-liquid separation, and drying of dairy manure: measuring constituents and modeling emission. Sci Total Environ 696:134059 27. Holly MA, Larson RA, Powell JM, Ruark MD, Aguirre-Villegas H (2017) Greenhouse gas and ammonia emissions from digested and separated dairy manure during storage and after land application. Agric Ecosyst Environ 239:410–419

The Use of Photoplethysmography for Blood Glucose Estimation by Noninvasive Method Vandana C. Bavkar and Arundhati Shinde

Abstract Noninvasive blood glucose estimation system based on photoplethysmography is proposed in this paper. Photoplethysmography (PPG) is an optical technique which uses infrared light to determine the fluctuations in blood stream in the human body with every pulse. However literature reveals that, a functional correlation exists among pulse signal waveform and the glucose contents in the blood. Many researchers have shown estimation of glucose using different spectroscopy techniques which involves an optical sensor. The key involvement of this paper is withdrawal of different features with respect to time domain as well as frequency domain. Exploration of single pulse is also carried out with different features. A PPG data of 182 random people with varied health conditions like diabetic, blood pressure condition is recorded for about 1 min duration. The glucose estimation is done in two ways: the first one is feature extraction from complete PPG signal of a patient and the second one is single pulse wave analysis. All the features from above-mentioned methods are transformed into feature vector which acts as input to system and estimation of blood glucose is performed. We used Clarke Error Grid Analysis for blood glucose estimation which is clinically accepted. With time and frequency domain features we got 43.75% samples in region A and 50% samples in region B, so total 93.75% samples are in regions A and B. Using single pulse wave analysis we got 59.66% samples in region A and 34.61% samples in region B, so total 94.27% samples are in clinically accepted regions (A and B). As single pulse wave analysis shows significant improvement in region A (20% of actual blood glucose value), this method is good selection for glucose measurement. Keywords Blood glucose · Noninvasive · Photoplethysmography · Feature extraction · Neural network V. C. Bavkar (B) · A. Shinde Department of Electronics, Bharati Vidyapeeth (Deemed to be University), College of Engineering, Pune 411043, India e-mail: [email protected] A. Shinde e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_21

323

324

V. C. Bavkar and A. Shinde

1 Introduction Diabetes mellitus which is also known as diabetes is a group of metabolic disorders caused due to high glucose level for a long period of time. Untreated diabetes may lead to serious damages to heart, eyes, blood vessels, and nerves. Diabetes occurs due to one of the two mechanisms. The first one is insufficient creation of insulin or insufficient sensitivity of cells to the action of insulin. Here, insulin is a type of a hormone released by pancreas. Inside the cells, glucose is converted to energy. Depending upon insulin use, Type1 and Type2 diabetes was found in people. Type1 diabetes is also called as insulin-dependent diabetes. In Type1 diabetes pancreas doesn’t make insulin, because of this glucose is not utilized by the body’s cell to get energy. Treatment for Type1 diabetes involves injecting insulin into fatty tissues of your body. Type2 diabetes is non-insulin-dependent diabetes. In this case, pancreas make insufficient insulin treatment for Type2 diabetes which includes oral medications [1–4]. Currently existing techniques for blood glucose measurements are invasive in nature, and these methods are very painful and having discomfort in nature. This is the main cause of enormous research in noninvasive glucose measurement technique which is very comfortable and mainly no need of blood. Infrared spectroscopy is one of the popular techniques for this purpose. NIR spectroscopy is used for proposed system [5–11]. Photoplethysmography (PPG) is simple optical method to observe blood volume changes in circumferential circulation in cardiovascular system. It is low cost, noninvasive, and portable method because of which in last few decades this technique is commonly used in various modern health care systems. PPG signal also gives information about respiratory systems [12–16]. Researchers found that there is a correlation among PPG signal and glucose level [17]. Yadav describes a PPG signal-based system used to measure blood glucose. Galvanik skin response, temperature, and PPG data are used as input to machine learning system. The multisensory system considerably improves prediction error and correction coefficient in contrast to previous methods in literature [18]. Vahlsing describes noninvasive blood glucose measurement technique based on color-coded photoplethysmographic images of finger in NIR and visible range [19]. Cruz suggests application of PPG signal for analysis of glucose with the use of moving average filter. He found the correlation between glucose and peak-to-peak voltage. For non-diabetic patient, it is direct proportion, and for diabetic patient, it is indirect proportion [20]. Haroon developed a prototype for noninvasive measurement of blood glucose, pulse rate, and oxygen saturation level [21]. Philip developed a blood glucose monitoring system with acquired PPG signal along with bioimpedance. Prediction error gets reduced with bioimpedance method [22]. Venkataramanan proposed a blood glucose prediction system with NIR sensor output voltage. With this sensor, voltage in regression model, prediction of glucose is done with mean error of 0.8558 [23]. Chowdhury describes a blood glucose estimation using smart phone video which collects data from measurement site. These video frames are later transformed into PPG waveform. Principle component regression is useful for estimation of glucose from extracted features. The projected model

The Use of Photoplethysmography for Blood Glucose Estimation …

325

could estimate the glucose getting a standard error of 18.31 mg/dL for testing dataset [24].

2 Methodology 2.1 System Diagram The workflow for anticipated system is given in Fig. 1. It consists of patient dataset module which created using pulse signal acquisition system. This data is given to preprocessing module to filter the noisy signal. This signal is given to different feature extraction modules. Extracted features are converted to feature vector and given to system where the blood glucose values are predicted. Fig. 1 System diagram

326

V. C. Bavkar and A. Shinde

Fig. 2 NIR signal acquisition system

2.2 Database Data is collected from 182 people having 25–85 age groups chosen arbitrarily. These people are having various medical circumstances like high or low blood pressure, healthy, diabetic, and non-diabetic. The different attributes like age, gender, sex, skin tone, blood pressure level are considered for creation of in vivo database.

2.3 Data Acquisition System Near infrared (NIR) optical sensor is used for recording of PPG signal from finger as body site. When NIR light is passed through body site, a pulse signal is acquired by enlightening the skin and calculating variations in the absorption of light. The absorption depends upon the heart rate as the blood vessel through finger contract and expands in accordance with heart beat. At the output of optical detector, we get reflected signal in the form of pulse. Different parameters like arterial stiffness, pulse rate, blood pressure, and glucose level can be analyzed using PPG signal analysis [25]. Figure 2 shows NIR signal acquisition system.

2.4 Preprocessing of Signal The output of detector contains some noise due to motion artifacts, power line interference. The quality of detector output depends on skin thickness, i.e., length of an optical path, blood flow, and skin tone at the measurement. Noisy signal will affect the feature extraction and blood glucose measurement. Therefore, preprocessing of this noisy signal is necessary step. The input sequence x(n) for n samples is represented

The Use of Photoplethysmography for Blood Glucose Estimation …

327

as x(n) = {x1 , x2 , x3 . . . , xn } samples. The filter used for preprocessing is moving average filter which removes the baseline drifting from the signal. The output of moving average filter is given by the following equation: y(n) =

N −1 1  x(n − i ) N i=0

(1)

where y(n) output of the filter x(n − i) previous input N length of the filter (10 in our case).

2.5 Feature Extraction Time and frequency domain analysis is done for feature extraction. Analysis of physical signal with respect to time is time domain analysis and when signal is analyzed with frequency, it is frequency domain analysis. For these features, we have calculated statistical measures like mean (μ), quartiles (interquartile range) (iqr), variance (σ ), and skewness (skew). Statistical measures quantitatively describe or summarize features of collection of data or information. Statistical measure ‘mean’ measures the central tendency of the data. Interquartile range is used for measuring dispersion of data. It is the difference between 25 and 75th quartile. Variance shows how individual number relates to each other within dataset. It is calculated by arithmetic mean of the squares of deviations of all values in a set of numbers form their arithmetic mean. Skewness gives distortion or asymmetry in a normal distribution [17, 18]. Figure 3 shows PPG signal showing features like pulse interval, peak interval, etc. Different frequency domain features are extracted and calculated the statistical measures for each feature. Regression Coefficients Regression coefficient basically used to forecast upcoming values from past values. These coefficients describe the envelope of the spectrum. When we apply this on PPG signal, we can track the variations in shape of the signal because of blood flow. The coefficients are denoted as A0 , A1 , A2 , A3 , A4 , A5 . Power Spectral Density (PSD) This feature gives information about power present in the signal with respect to frequency. It also evaluates damping of signal, occurrence of noise, and the spectral

328

V. C. Bavkar and A. Shinde

Fig. 3 PPG signal

harmonic components. Statistical features mean PSDμn , interquartile range PSDiqr n , are calculated. variance PSDσn , and skewness PSDskew n Kaiser–Teager Energy (KTE) This is popular technique used for measuring energy of signal. For differentiating clean and noisy signal, this property is very useful. High energy signifies high amplitude signal with good quality and low energy signal signifies noisy signal with low amplitudes [17]. Statistical features like skewness, variance, interquartile range, and mean are calculated. Time domain features show signal variation over the time. Heart rate, pulse transit time, and peak interval these features are calculated from PPG signal. Heart Rate (HR) Heart rate is one of the factors affecting blood glucose level. Hence, this can be calculated from peaks of the PPG signal. Statistical features mean HRμn , interquartile σ skew are calculated. range HRiqr n , variance HRn , and skewness HRn Pulse Transit Time (PTT) The travel time of pulse between two arterial points is referred to as pulse transit time. For this measurement, we have to consider one frame. Statistical features mean PTTμn , σ skew are calculated. interquartile range PTTiqr n , variance PTTn , and skewness PTTn Peak Interval (PPInt) Peak interval is the width between consecutive peaks of PPG signal considering one σ frame. Statistical features mean PPIntμn , interquartile range PPIntiqr n , variance PPIntn , skew and skewness PPIntn are calculated [26–29].

The Use of Photoplethysmography for Blood Glucose Estimation …

329

Table 1 Time and frequency domain features Feature

Statistical measures (mean, interquartile range, variance, and skewness)

Auto-regression coefficients

A0 , A 1 , A 2 , A3 , A 4 , A 5

Kaiser-Teager energy

KTEn , KTEn , KTEσn , KTEskew n

Power spectral density Heart rate Pulse transit time Peak-to-peak interval Peak amplitude

μ

No. of features 6

iqr

4

iqr μ PSDn , PSDn , PSDσn , PSDskew n iqr μ HRn , HRn , HRσn , HRskew n iqr μ PTTn , PTTn , PTTσn , PTTskew n iqr μ σ PPIntn , PPIntn , PPIntn , PPIntskew n iqr μ PAmpn , PAmpn , PAmpσn , PAmpskew n

4 4 4 4 4

Total

30

Peak Amplitude (PAmp) Peak amplitude is the systolic pulse of PPG signal which shows rhythmic changes in blood because of arterial blood flow. Statistical features mean PAmpμn , interquartile σ skew are calculated. range PAmpiqr n , variance PAmpn , and skewness PAmpn Till now we have calculated 30 different features listed in Table 1.

3 Single Pulse Wave Analysis PPG signal is divided into single pulses, and each single pulse used for feature extraction is shown in Fig. 4. Features of single pulse: • Width period (T ) Fig. 4 Single pulse wave

A T0

T2 T1

C

B T

Width period

330

• • • • • • • • • • • • • • • • •

V. C. Bavkar and A. Shinde

Highest peak value (A) Time of highest peak value (T 0) Diastolic peak amplitude (B) Time of Diastolic peak (T 1) Notch amplitude (C) Time of notch (T 2) Time Difference Start-Peak (T 0) Time Difference Peak-Notch (T 2 − T 0) Time Difference Notch-Diastolic Peak (T 1 − T 2) Time Difference Diastolic Peak-End (T − T 1) Mean amplitude Standard deviation of single pulse Length of Single period Mean Start-Max Mean Max-Notch Mean Notch-Diastolic Peak Mean Diastolic Peak-End.

We have calculated 18 features from single pulse wave. Feature dimension method is used to reduce number of features as data samples are huge in number. Only seven features are inputted to system. The network configuration details are summarized in the following section.

3.1 Neural Network Configuration In last few decades, neural networks are recursively used in medical diagnosis and healthcare systems [30, 31]. All network structure is built in neural network toolbox of MATLAB software. At start, the testing is performed with different neurons present in hidden layer which affect the accuracy of the system. The hidden layer neurons are calculated by following equation. N=

I+O √ + S 2

(2)

where N I O S

Number of neurons Number of inputs Number of outputs Total samples.

The neurons considered in the experimentations are ±5 of neurons calculated as per above equation.

The Use of Photoplethysmography for Blood Glucose Estimation …

331

3.2 Clarke Error Grid Analysis Clarke Error Grid Analysis is a crucial tool for examining the scientific accuracy of blood glucose monitors. It shows the graphical relationship between reference glucose value versus predicted glucose value. Error grid divides graph into five different regions. In region A, samples are within 20% range of reference value. In region B, samples are outside of Region A but this does not lead to inappropriate medication, also called ‘Clinically uncritical decisions’. Region C samples are directing to redundant medication. Region D samples indicate hazardous malfunction to identify hypoglycemia or hyperglycemia, and lastly Region E samples would mystify medication of hypoglycemia for hyperglycemia. Region A and Region B are the clinically accepted regions [32].

3.3 Performance Measure Clarke Error Grid Analysis is used as performance measure for different neurons in hidden layer. The amount of samples in region A and region B varies according to deviation or change in number of neurons. MATLAB software is used to build neural networks and Clarke error grid graphs.

4 Results 4.1 Time and Frequency Domain Features Thirty features were taken out from pulse signal and given as input to network. At the network output, we get predicted blood glucose concentration level. The total data samples are separated to train and test dataset. 80% samples (150 samples) are applied to train the network and 20% (32 samples) samples are for testing purpose. According to Eq. (1), hidden layers were calculated in the range of 24–34. With different neuron configuration, (30, 1) as shown in Fig. 5, we get better accuracy than other neuron configurations. Clarke error grid graph is shown in Fig. 6 for this configuration. With time and frequency related features, 43.75% samples in region A and 50% samples in region B, so total 93.75% samples are in region A and region B.

332

V. C. Bavkar and A. Shinde

Fig. 5 Two layer neural network detailed architecture

Fig. 6 Clarke Error Grid Analysis using time and frequency domain features

4.2 Single Pulse Wave Features Four hidden layers are used as shown in Fig. 7. Total 12,196 single wave pulses of 182 patients are there. 80% samples (10,012 samples) are applied to train the neural network and 20% (2184 samples) samples are applied for testing purpose. Starting with two layers (15, 10) the network gives optimum results. Hence, we have

The Use of Photoplethysmography for Blood Glucose Estimation …

333

Fig. 7 Four layer neural network architecture

to increase one layer with (15, 10, 5) neurons. Further testing is done with different neuron combinations, and (20, 15, 10) configuration gives better accuracy with error grid graph samples as shown in Fig. 8. Using single pulse wave analysis, 59.66% samples are in A region and 34.61% samples in B region, so total 94.27% samples are in A and B regions.

Fig. 8 Clarke Error Grid Analysis using single wave features

334

V. C. Bavkar and A. Shinde

5 Conclusion Proposed system gives accuracy with 94.27% with single pulse wave analysis and that with time and frequency domain feature set is 93.75%. During this research work, different time domain and frequency domain features are taken out from PPG signal and from pulse wave. As single pulse wave analysis shows significant improvement in region A (20% of actual blood glucose value) of Clarke Error Grid Analysis, this method is good selection for noninvasive blood glucose measurement. Extraction of derivative features is the extended task for our research work for better results. Acknowledgements We would like to express thanks to Bharati Hospital and Research Center, Pune, India for allowing us to gather data from people in pathology laboratory.

References 1. International Diabetes Federation. About diabetes. http://www.idf.org/about-diabetes 2. Kovatchev BP (2012) Diabetes technology: markers, monitoring, assessment, and control of blood glucose fluctuations in diabetes. Scientifica 2012, 14 pages. Article ID 283821 3. Migdalis I, Leslie D, Papanas N, Valensi P, Vlassara H (2014) Diabetes mellitus. Int J Endocrinol 2014, 6 pages. Article ID 108419 4. Chen C, Zhao X-L, Li Z-H, Zhu Z-G, Qian S-H, Flewitt AJ (2017) Current and emerging technology for continuous glucose monitoring. Sensors 5. Senthi Kumar A, Kavitha S, Muthubharathi R (2017) Design and development of absorption spectrophotometric based non-invasive blood glucose measuring device. Res J Pharm Technol 10(1):91–97. https://doi.org/10.5958/0974-360X.2017.00022.1 6. Gamessa TW, Suman D (2019) Non-invasive blood glucose monitoring using visible laser light. Res J Pharm Technol 2(2):831–840 7. Premalatha J, Grace Kanmani Prince P (2019) Analysis of non-invasive methods to diagnose blood glucose level—a survey. Res J Pharm Technol 12(6):3105–3108 8. Thakur J (2019) Effect of maternal diabetes on fetus and newborn. Asian J Nurs Educ Res 9(3):463–465 9. Sirisha (2020) Recent advancements on colon targeted drug delivery systems. Asian J Pharm Res 10(4):268–274 10. Saudagar RB, Samuel S (2016) Ethosomes: novel noninvasive carrier for transdermal drug delivery. Asian J Pharm Technol 6(2):135–138 11. Ravindra HN, Christian K, Pooja G, Prem P, Priyal J, Rajat A, Riya R (2016) Knowledge and attitude on self monitoring of blood glucose (SMBG) among diabetic patients belongs to Waghodia taluka. Int J Adv Nurs Manage 4(4):398–403 12. Awasthi A, Sharma N (2014) Possibilities of using neural network for ECG classification. Res J Eng Technol 5(1):13–16 13. Walia MS (2017) Performance analysis of feature extraction techniques for iris pattern recognition system. Res J Eng Technol 8(4):431–435 14. Rajaguru H, Prabhakar SK (2018) Wavelet neural networks, Elman backpropagation and multilayer perceptrons for epilepsy classification from EEG signals. Res J Pharm Technol 11(4):1301–1306 15. Meenakshi K, Maragatham G (2019) Computational intelligence in diagnosis and prognosis of gestational diabetes using deep learning. Res J Pharm Technol 12(8):3891–3895

The Use of Photoplethysmography for Blood Glucose Estimation …

335

16. Moraes JL, Rocha MX (2018) Advances in photoplethysmography signal analysis for biomedical applications. Sensors 17. Habbu S, Dale M, Ghongade R (2019) Estimation of blood glucose by non-invasive method using photoplethysmography. Sadhana 18. Yadav J, Rani A, Singh V, Murari BM (2016) Investigations on multi-sensor based non-invasive blood glucose measurement system. J Med Devices 19. Vahlsing T, Delbeck S, Leonhardt S, Michael Heise H (2018) Noninvasive monitoring of blood glucose using color-coded photoplethysmographic images of the illuminated fingertip within the visible and near-infrared range: opportunities and questions. J Diabetes Sci Technol 12(6):1169–1177 20. Cruz FRG, Paglinawan CC, Catindig CNV (2019) Application of reflectance mode photoplethysmography for non-invasive monitoring of blood glucose level with moving average filter. In: International conference on biomedical engineering and technology, ICBET 2019 21. Haroon N, Tiwana MI (2017) Design and development of non-invasive prototype to measure pulse rate, blood glucose and oxygen saturation level in arterial blood. In: Future technology conference (FTC), Canada, Nov 2017, pp 226–233 22. Philip LA, Rajasekaran K, Jothi E (2017) Continuous monitoring of blood glucose using photophlythesmograph signal. In: Proceedings of IEEE international conference on innovations in electrical, electronics, instrumentation and media technology, ICIEEIMT 2017 23. Venkataramanan S, Kamble D, Bairolu A, Singh A, Rao R (2017) A novel heart rate and noninvasive glucose measuring device. In: International conference on communication and signal processing, ICCSP 2017 24. Chowdhury TT, Mishma T, Osman MS, Rahman T (2019) Estimation of blood glucose level of type-2 diabetes patients using smartphone video. In: Proceedings of the 6th international conference on networking, system and security, Dec 2019, pp 104–108 25. Habbu SK, Joshi S, Dale M, Ghongade R (2019) Noninvasive blood glucose estimation using pulse based cepstral coefficients. In: IEEE international conference on signal processing and information security (ICSPIS) 26. Zhang Y, Zhang Y, Siddiqui SA, Kos A (2019) Non-invasive blood-glucose estimation using smartphone PPG signals and subspace KNN classifier. Electrotech Rev 86(1):68–74 27. Jubadi W, Sahak S (2009) Heartbeat monitoring alert via SMS. In: Proceedings of IEEE symposium on industrial electronics & applications, vol 1, pp 1–5 28. Fu T, Liu S, Tang K (2008) Heart rate extraction from photoplethysmogram waveform using wavelet multi-resolution analysis. J Med Biol Eng 28(4):229–232 29. Linder S, Wendelken S, Wei E, McGrath S (2006) Using the morphology of photoplethysmogram peaks to detect changes in posture. J Clin Monit Comput 20(3):151–158 30. Bukhari MM, Alkhamees BF, Hussain S, Gumaei A, Assiri A, Ullah SS (2021) An improved artificial neural network model for effective diabetes prediction. Complexity 2021, 10 pages. Article ID 5525271 31. Bavkar VC, Shinde AA (2021) Machine learning algorithms for diabetes prediction and neural network method for blood glucose measurement. Indian J Sci Technol 14(10):869–880. https:// doi.org/10.17485/IJST/v14i10.2187 32. Clarke WL, Cox D, Gonder-Frederick LA, Carter W, Pohl SL (1987) Evaluating clinical accuracy of systems for self-monitoring of blood glucose. Diabetes Care 10(5):622–628

Single-Stage Stand-Alone Induction Motor Driven Solar Water Pumping System with Minimal Sensors Anup Shetty , K. Suryanarayana , and L. V. Prabhu

Abstract Water pumps are essential to extract ground water for agricultural purpose and majority of these pumps are driven by electric motors. However, in remote places due to power scarcity, people rely upon diesel-powered pumps. Solar-powered water pumping has gained attention in the recent years because of increased awareness about global warming, caused by the greenhouse gases released by the fossil fuel-based systems and encouraging government policies to reduce carbon footprint. Efforts are made to develop a cost effective and high performance solar water pump drives. A single-stage induction motor driven photovoltaic water pumping system with minimal sensors is presented in this paper. Proposed system consists of PV array, power inverter stage and scalar (V /f ) controlled induction motor (IM) to drive water pump. In the proposed system, PV array stage is connected to the DC bus of power inverter stage eliminating intermediate DC-DC converter stage to reduce overall cost and size. Optimal utilization of the PV system is assured with the help of variable step size-incremental conductance (VSS-INC) maximum power point tracking (MPPT) algorithm. Elimination of feedback signals in scalar control approach reduces the sensor count. The proposed system has been modeled and simulated using MATLAB/Simulink tool. System performance and effectiveness under atmospheric parameter fluctuations are simulated in detail. Keywords Maximum power point tracking · Photovoltaic system · Voltage source inverter · Variable step size-incremental conductance · Solar water pumping · V/f control A. Shetty (B) · K. Suryanarayana Department of Electrical and Electronics Engineering, NMAM Institute of Technology, Nitte, Karkala, Karnataka 574110, India e-mail: [email protected] K. Suryanarayana e-mail: [email protected] L. V. Prabhu HEXMOTO Controls Pvt. Ltd, Mysuru 570018, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_22

337

338

A. Shetty et al.

1 Introduction Renewable energy conversion and utilization is gaining interest due to the rise in adverse effects of environmental pollution caused by burning fossil fuels and escalating worldwide energy crisis [1]. Solar radiation energy is available in the cleanest form and abundant even in remote places making it favorite among all other renewable energy sources [2]. The recent advancements in photovoltaic cell technology and rapid growth in the semiconductor industry have empowered development of efficient solar energy harvesting solutions with minimal maintenance. Mass production of PV cells has led to steady drop in cost per watt-peak over the years. Possibility of increasing installed power capacity by adding panels to the existing system is one of the attractive features of solar energy harvesting. However, intermittence, high initial cost and low efficiencies are some of the drawbacks of current PV systems. Water pumping is a vital requirement for agriculture. Solar water pumping is most useful for remote areas with ample solar irradiance [3]. In solar water, pumping storage features such as batteries are not essential since water can be pumped and stored in tanks whenever the irradiance is available [4]. Based upon the water pumping requirement, a solar PV array with an appropriate rating could be selected. In the case where the water pump needs to be operated throughout the day with full power, a solar PV array with a battery system could be used. Energy stored in the battery even could be used to energize household appliances in remote areas. The natural connection present among the necessity of water pumping and the availability of solar irradiance, depending upon season requirement of water pumping changes. Solar water pumping (SWP) systems offer several advantages over the traditional fossil fuel-based generator powered pumps in terms of enhanced reliability, reduced operation and maintenance costs [5]. Solar energy is available only in the daytime and it is affected by environmental situations, weather changes and clouds. To maximize the energy extracted, the PV array should be operated at its point of maximum efficiency, called as maximum power point (MPP) [1, 6]. To locate maximum power point, tracking algorithms are being used based on the requirement. Several algorithms for MPP tracking have been presented by the researchers over the times. Perturb and Observe algorithm (P&O) is one of the simplest and effective tracking methods but has disadvantages such as oscillating around maximum power point, low response speed and low accuracy [7]. Incremental conductance (INC) MPPT algorithm overcomes the disadvantages of P&O algorithm and dynamic response could be improved with the help of variable step size technique [2]. Induction motors are commonly used in water pumps for their robustness and low cost despite low efficiency when compared to advanced motors [8]. Permanent magnet synchronous motors (PMSMs) could be an alternative for induction motors. PMSM has advantages of high efficiency, low noise, high air gap flux density, high power-to-weight ratio, swift acceleration and deceleration capability, high power factor and compact design [2] but suffers from high initial cost and control complexity. Benefits of advanced motors are surpassed by the cost effectiveness and

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

339

easy controllability of induction motor and is suitable for economical water pumping solution. Solar PV array generates DC voltage which needs to be converted to AC form with the help of inverter. Single-stage IM drives suffer from DC voltage fluctuation [9, 10]. Feed forward-based reference speed control with VSS-INC is used to minimize the DC voltage fluctuations and to improve dynamic performance. Speed of the motor is varied based upon the maximum power point tracker algorithm. Motor control in photovoltaic pumping must be focused on captivating maximum power. Power control of the induction motor from the PV is demanding because the motor mechanical time constant and electrical time constant of the system are different. Dynamic power flow control is required for effective operation of the system. Several IM control methods including scalar control, V/f control, vector control, field-oriented control and direct torque control are present in the literature [7]. Field-oriented control strategies are employed to adapt high performance induction motor control [5] with additional current sensors. Vector control methods involve complex transformations and additional current and position sensors. However, scalar control is good enough as accurate speed control is not required for water pumps. Scalar control is chosen for the system development because of its simplicity. This paper is organized as follows: Sect. 1 presents a brief introduction to the induction motor driven solar water pump. Section 2 discusses the system architecture of the proposed methodology. Control scheme of the solar-powered pump is discussed in Sect. 3. Section 4 presents the system modeling and simulation aspects. Section 5 gives the simulation results and observations and Sect. 6 presents the conclusion.

2 System Architecture The system configuration of the proposed scheme for stand-alone solar-powered water pumping is as in Fig. 1. The water pumping mechanism is driven by an induction motor controlled with conventional scalar control technique. The system consists of a PV array and three-phase voltage source inverter to control IM. Suitable number of PV modules are connected in series and parallel combination to obtain desired voltage and power ratings. The PV array is connected to the DC bus of VSI, eliminating intermediate DC-DC converter which is commonly used to implement solar MPPT algorithms. Power captivated from the PV array is being controlled with the help of variable step size-incremental conductance MPP tracking technique to draw maximum power with available radiation. Speed reference for the scalar control structure of IM drive is generated using MPPT algorithm.

340

A. Shetty et al.

Fig. 1 System configuration

2.1 Solar PV Array Design Photovoltaic cells are used to convert solar radiation energy into electrical energy. Power rating choice of the PV array depends upon the load requirement. PV array rating is designed by considering the losses in power converter stage and maximum irradiance available in the area of system installation. PV cells with power rating slightly greater than the peak load power could be selected to compensate for the losses. The proposed system makes use of a 4 kW induction motor; hence, PV array with peak capacity of 4.4 kW under standard test condition (1000 W/m2 , 25 °C) is preferred. PV array is formed by connecting solar cell modules in series and parallel manner to obtain desired voltage and power level. Standard solar module A10J-M60-220 by A10 Green Technology is considered for the design. PV array maximum power, PMP = 4.4 kW

(1)

Maximum power of single module, PM = 219.87 W

(2)

Total number of modules required is calculated as, n=

4.4 k ∼ PMP = = 20 PM 219.87

(3)

Single-Stage Stand-Alone Induction Motor Driven Solar Water … Table 1 Solar module and PV array specifications

Parameters

341 Value

Maximum power of single module

219.876 W

Open-circuit voltage of module (V OC )

36.06 V

Short-circuit current of module (I SC )

7.95 A

Voltage at maximum power point (V MPP )

30.12 V

Current at maximum power point (I MPP )

7.3 A

Number of modules connected in array

20

PV array open-circuit voltage (V OC )

721.2 V

PV array short-circuit current (I SC )

7.95 A

PV array maximum power (PMP )

4.4 kW

PV array voltage at maximum power (V MP )

602.4 V

PV array current at maximum power (I MP )

7.3 A

For a single-stage topology, V MP is selected near to V DC voltage level. Rated voltage of motor is 400 V; hence, DC link voltage could be chosen as, VDC =

√

2 ∗ VL−L ∼ = 566V

(4)

VDC is taken as 600 V. To obtain the desired DC link voltage, 20 solar modules are required to be connected in series. Specifications of the solar module and PV array are as in Table 1.

2.2 DC Link Capacitor Selection DC link capacitor is used to support the DC voltage level under transient conditions and the value of CDC is calculated as [2], CDC =

IO 4400/600 ∼ = = 390 µF 2 ∗ ω ∗ VDC 2 ∗ 314.16 ∗ 0.05 ∗ 600

(5)

where Io = PMP /VDC , ω = 2 ∗ π ∗ 50 = 314.16 rad/s and DC bus voltage ripple is taken as 5% of VDC . Standard capacitor of 470 µF is considered for the system design.

3 Control Methodology Control section of the proposed system primarily consists of variable step sizeincremental conductance MPPT algorithm and V /f control for the IM drive. Voltage

342

A. Shetty et al.

and current signals of PV array are sensed and fed to the MPP tracker to extract maximum power. Induction motor is driven with a scalar control technique which greatly reduces the control strategy complexity. VSS-INC algorithm generates the DC link voltage reference (V DCref ). Error between the actual DC link voltage (V DC ) and reference DC link voltage (V DCerror ) is fed to a PI controller to reduce the voltage error by adjusting the speed reference value. Assuming lossless system and steady state operation, output and input power could be related as, PPV = PCDC + Pm

(6)

where PPV , PCDC and Pm are the power of PV array, DC link capacitor and motor, respectively. Under steady state, PCDC will be zero thus, PPV = Pm

(7)

Pm = τm ∗ ωm

(8)

where

For a pump-type load using pump affinity law, speed and torque relationship for the motor is given as [1], τm ∝ ωm 2

(9)

Considering K m as the proportionality constant, Eq. (9) becomes, Km =

τm ωm 2

(10)

where τm is the pump load torque under steady state condition in N-m and ωm is the motor speed in rad/s. Using (10), Eq. (7) could be written as, PPV = K m ∗ ωm 3

(11)

From Eq. (11), it is evident that the power absorbed from the PV array (PPV ) is directly proportional to the cube of the motor speed (ωm ). Speed reference could be generated with the knowledge of power supplied by the PV array for effective MPP tracking. Accordingly, two speed reference components are generated as in Fig. 2. ωref1 from the DC link voltage regulator and ωref2 from the speed feedforward term employing the relation mentioned in Eq. (11). Speed feed forward term improves the system performance under dynamic conditions. DC link voltage regulating PI controller is used to fine-tune the ωref1 . Reference frequency for the V /f control is

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

343

Fig. 2 Control diagram

obtained from ωref1 and ωref2 . Sinusoidal pulse width modulation (SPWM) technique is used to generate the control pulses for the VSI.

3.1 Variable Step Size-Incremental Conductance MPPT Solar PV array PPV versus VPV curve is nonlinear in nature as in Fig. 3. For a fixed irradiance and temperature, power delivered by the PV cells greatly depend on the load impedance. To operate the solar array at the MPP, a tracking algorithm is a mandatory requirement. DC-DC converters in conjunction with the MPPT algorithms are universally used to match the load and PV array impedances for maximum power absorption. P&O technique suffers from the steady state oscillation about MPP and poor performance under irradiance changes. INC is another MPP tracking technique which is based on the slope of PV characteristics curve. INC basically monitors the slope of PV characteristics curve and makes use of the fact that at MPP, slope of the PV curve is zero to track MPP. Dynamic performance under irradiation changes is better in case of INC. Incremental conductance method is conventionally implemented with fixed step size for MPP tracking. Fixed step sized INC suffers from the possibility of oscillation about MPP under steady state, leading to loss of power. Even though slope of PPV versus VPV curve is tracked, steady state oscillation depends on the chosen step size. Step size could be made small to avoid power oscillation but dynamic performance of INC with small step size is poor and slower. These problems could be minimized by employing variable step size, depending upon transient conditions, step size is recalculated. Flow diagram of the VSS-INC is as in Fig. 4. MPPT algorithm of the proposed system controls the V DCref as, VDCref (k) = VDCref (k − 1) ± step Variable step size is calculated in each iteration and is given as,

(12)

344

A. Shetty et al.

Fig. 3 PV characteristics

   dPPV   step = K s ∗  dVPV 

(13)

where d PPV and dVPV are the change in PV power and PV voltage from present sampling instant (k) to previous instant (k − 1), respectively. Here, K s is the step size constant which is chosen appropriately by considering system stability. As system operating point reaches MPP, value of step tends toward zero, eliminating steady state oscillations. Maximum value of step size should be limited to stepmax so that under starting and transient conditions, speed of the tracking algorithm is limited. To ensure convergence, step size of VSS-INC should follow the condition as in    d PPV    K s < stepmax / dVPV step=stepmax

(14)

3.2 Induction Motor V/f Control Scalar or V /f is the common open-loop speed control technique used to drive induction motors. IMs are designed to operate at rated torque and speed for rated terminal voltage and frequency. When the variable speed is desired, voltage needs to be reduced and to get the rated torque performance of the IM, air gap flux needs to be maintained constant. This is achieved by varying both magnitude and frequency of terminal voltage so that IMs could perform best at all speed ranges. Air gap flux (ϕ) is related to voltage (V ) and frequency ( f ) as,

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

345

Fig. 4 VSS-INC algorithm flowchart

ϕ∝

V f

(15)

SPWM-controlled VSI is used to achieve scalar control of IM. Reference frequency is given based upon speed requirement neglecting the slip speed. Since V /f

346

A. Shetty et al.

constant for a motor is known, corresponding modulation index and hence voltage magnitude is calculated. Three sinusoidal reference signals with desired frequency, magnitude and 120 ◦ apart from each other are generated. Reference signals are compared with the carrier wave to generate desired SPWM control signals.

4 Modeling and Simulation Individual modules are modeled to form overall system model. PV array, induction motor and water pump load are modeled to match the practical system behavior. Static single-diode model is used for the system analysis [11, 12]. PV voltage and current relationship is given as,  VPV + IPV Rs Rsh     VPV + IPV Rs −1 Id = IS exp ηn c VT 

IPV = Iph − Id −

(16) (17)

where Iph is the PV generated current, Id is the PV diode’s forward bias current, dark saturation current is represented by IS , n c indicates number of series connected cells, η indicates diode’s quality factor, Rs and Rsh are the series and shunt resistances, respectively, VT is the thermal voltage. Induction motor model in d-q reference frame rotating with motor speed is considered for system simulation. Dynamic equations of IM [13–15] are given as, Vds = Rs i ds +

dϕds + ωm ϕqs dt

(18)

Vqs = Rs i qs +

dϕqs − ωm ϕds dt

(19)

Vdr = Rr i dr +

dϕdr dt

(20)

Vqr = Rr i qr +

dϕqr dt

(21)

where ωm is the rotor speed, Rs is the stator phase resistance, Rr is the rotor phase resistance, ϕds , ϕqs and ϕdr , ϕqr are the stator and rotor flux linkages in d-q frame, respectively. Vds , i ds , Vqs , i qs and Vdr , i dr , Vqr , i qr are the stator and rotor voltages and currents in d-q frame, respectively. A three-phase 4kW, 400V induction motor is used to drive the pump load. To model the load, pump affinity law is considered. Rated torque of the selected motor is 26.71 N-m and rated speed is 1430 rpm (149.75 rad/s). From (10), the motor

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

347

Fig. 5 Simulink model of the system

constant K m could be obtained as, Km =

26.71 = 1.191 × 10−3 N-m/(rad/s)2 (149.75)2

(22)

Proposed system model is developed and simulated to evaluate the performance of the control scheme. Simulation model developed is as in Fig. 5. PV array is formed by connecting 20 solar modules in series, and variable irradiance and temperature signals are given to the PV model to emulate real-world scenario. VSI stage is modeled with the help of power semiconductor device model. Load torque for the motor model is fed using the relationship given in Eq. (10). DC link voltage controller tuning is carried out based on the desired performance of the system under momentary change in operating conditions. Control strategy as explained in the control methodology section is modeled and VSI control pulses are generated. PV voltage and PV current signals are used to calculate reference speed for IM drive. Power delivered by the PV array is monitored to ensure faithful tracking of the MPP.

5 Results and Discussion Irradiance and temperature variation are simulated to analyze effectiveness of proposed methodology. Nominal steady state irradiance is taken as 900 W/m2 at temperature of 25 ◦ C. PV dynamics are emulated by changing irradiation to 100W/m2 and temperature to 40 ◦ C as in Fig. 6. PV voltage and current level vary during irradiation and temperature change. Voltage and current sense signals of the PV array for an irregular irradiation and temperature are as in Fig. 7. It is observed that the PV voltage is well regulated under

348

A. Shetty et al.

Fig. 6 Irradiance and temperature variation with time

steady state conditions and transient condition overshoots are limited well within the acceptable range. When the fall in irradiation is initiated at t = 2 s and temperature of the PV array is increased at t = 6 s, PV power starts to decrease as in Fig. 8. Fall in the power should be detected by the MPP tracker to find new MPP as operating characteristics of the PV changes with irradiance and temperature fluctuation. Accordingly, speed of the motor is adjusted to compensate for dynamic power variation. Figure 9 depicts the induction motor speed and electromagnetic torque variation for a given disturbance pattern. Induction motor current and voltages are monitored to assess the need of soft starting of the induction motor. Large starting current (34 A peak) is drawn when

Fig. 7 PV voltage and current

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

349

Fig. 8 PV power

Fig. 9 Motor speed and torque

the motor is operated without soft start as in Fig. 10 when MPP tracking is enabled all the time. To avoid high starting current, soft start technique is used in the system. Motor is driven by V /f control with slow ramp of speed reference signal without MPP tracking algorithm till the motor reference speed reaches a predefined limit. Once the reference speed attains set speed threshold, MPPT is engaged. Implementing soft start improves the life of the induction motor. Motor current and voltages with soft start are as in Fig. 11. It is observed that starting current drawn by the IM is reduced to 7.5 A peak.

350

A. Shetty et al.

Fig. 10 Motor startup voltage and current

Fig. 11 Motor voltage and current with soft start

6 Conclusion Model of the single-stage photovoltaic water pumping system with minimal sensors is analyzed in this paper. Simulation of the proposed model is carried out to assess the performance of the control strategy under various real-world conditions such as motor startup, irradiance and temperature variation. Steady state and transient performance of the VSS-INC algorithm are verified and it is observed that maximum power point tracking is carried out with greater efficacy. Overall system size and cost could be greatly reduced with the elimination of intermediate DC-DC converter stage. Simulation results of the modest control structure is found to be satisfactory.

Single-Stage Stand-Alone Induction Motor Driven Solar Water …

351

References 1. Singh B, Sharma U, Kumar S (2018) Standalone photovoltaic water pumping system using induction motor drive with reduced sensors. IEEE Trans Ind Appl 54(4):3645–3655. https:// doi.org/10.1109/TIA.2018.2825285 2. Murshid S, Singh B (2020) Single stage autonomous solar water pumping system using PMSM drive. IEEE Trans Ind Appl 56(4):3985–3994. https://doi.org/10.1109/TIA.2020.2988429 3. Montorfano M, Sbarbaro D, Morán L (2016) Economic and technical evaluation of solarassisted water pump stations for mining applications: a case of study. IEEE Trans Ind Appl 52(5):4454–4459. https://doi.org/10.1109/TIA.2016.2569415 4. Antonello R, Carraro M, Costabeber A, Tinazzi F, Zigliotto M (2017) Energy-efficient autonomous solar water-pumping system for permanent-magnet synchronous motors. IEEE Trans Ind Electron 64(1):43–51. https://doi.org/10.1109/TIE.2016.2595480 5. Maddalena ET, Moraes CGDS, Bragança G, Junior LG, Godoy RB, Pinto JOP (2019) A battery-less photovoltaic water-pumping system with low decoupling capacitance. IEEE Trans Ind Appl 55(3):2263–2271. https://doi.org/10.1109/TIA.2019.2900412 6. Caracas JVM, Farias GdC, Teixeira LFM, Ribeiro LAdS (2014) Implementation of a highefficiency, high-lifetime, and low-cost converter for an autonomous photovoltaic water pumping system. IEEE Trans Ind Appl 50(1):631–641. https://doi.org/10.1109/TIA.2013.2271214 7. Elgendy MA, Zahawi B, Atkinson DJ (2015) Operating characteristics of the P&O algorithm at high perturbation frequencies for standalone PV systems. IEEE Trans Energy Conver 30(1):189–198. https://doi.org/10.1109/TEC.2014.2331391 8. Jain S, Karampuri R, Somasekhar VT (2016) An integrated control algorithm for a single-stage PV pumping system using an open-end winding induction motor. IEEE Trans Ind Electron 63(2):956–965. https://doi.org/10.1109/TIE.2015.2480765 9. Short TD, Mueller MA (2002) Solar powered water pumps: problems, pitfalls and potential. In: 2002 international conference on power electronics, machines and drives (conf. publ. no. 487), pp 280–285. https://doi.org/10.1049/cp:20020129 10. Vitorino MA, de Rossiter Correa MB, Jacobina CB, Lima AMN (2011) An effective induction motor control for photovoltaic pumping. IEEE Trans Ind Electron 58(4):1162–1170. https:// doi.org/10.1109/TIE.2010.2054053 11. Venkatramanan D, John V (2019) Dynamic modeling and analysis of buck converter based solar PV charge controller for improved MPPT performance. IEEE Trans Ind Appl 55(6):6234–6246. https://doi.org/10.1109/TIA.2019.2937856 12. Costa de Souza A, Cardoso Melo F, Lima Oliveira T, Eduardo Tavares C (2016) Performance analysis of the computational implementation of a simplified PV model and MPPT algorithm. IEEE Latin Am Trans 14(2):792–798. https://doi.org/10.1109/TLA.2016.7437224 13. MacDonald ML, Sen PC (1979) Control loop study of induction motor drives using DQ model. IEEE Trans Ind Electron Control Instrum IECI-26(4):237–243. https://doi.org/10.1109/TIECI. 1979.351593 14. Lee RJ, Pillay P, Harley RG (1984) D, Q reference frames for the simulation of induction motors. Electr Power Syst Res 8(1):15–26. ISSN 0378-7796. https://doi.org/10.1016/03787796(84)90030-0 15. Krishnan R (2001) Electric motor drives: modeling analysis and control. Prentice-Hall, Englewood Cliffs, NJ, USA, pp 196–213. ISBN 0130910147

Automation of Weight-Based Sorting System Using Programmable Logic Controllers P. Chenchu Saibabu, R. Anjana, Manisha Kumari, and C. R. Srinivasan

Abstract Material sorting is an integral part of modern-day industries. In these industries, the production are made in batches ranging from several dozen to several thousand units, and thus, the raw material or any material used is also in abundance for example in case of coal industry, scrap industry, or in any industry that it’s quantity based or weight based factor that effects the production can be implemented. Using manual labour for sorting of material is expensive, time consuming and inefficient. This problem can be solved by automating the material sorting. In this paper, we have discussed how we have developed a system that effectively sorts the material according to weight using PLC as the controller. Sensor that is used for implementing the idea is a load cell which was kept under the specially designed conveyor to measure the weight of the object and then that in turn gives the analogous signal which is amplified and given to the PLC unit which gives signals to the respective solenoid valve and conveyor for sorting process. Keywords Sorting system · ANSYS · Programmable logic controller · S7-1200

1 Introduction Material sorting is a main concern in industries that deal with weight or quantity of the materials. For example, in mining industries the raw material mined is of different weights and sizes and requires different type of crusher according to their weight or size, without a proper sorting system before going to the crusher the raw materials of all sizes goes through all different types of crushes; thus, the material in smaller size are crushed to powder and is removed as waste [1]. This causes loss of a huge amount of natural resources in this case, thus using an automated sorting system to separate these raw material will make the process faster and reduce the amount of wastage. In other industries like the packaging industries or material distribution centres use P. Chenchu Saibabu · R. Anjana · M. Kumari · C. R. Srinivasan (B) Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_23

353

354

P. Chenchu Saibabu et al.

of automated sorting system can make the process more efficient, cost effective and less time consuming thus making it more economical, where manual labour is time consuming and can cause human error. In the postal industry, sorting of the mails require a lot of labour. The automation of the sorting of the mails reduces the labour charges and human interfere [2]. Thus, in many industries like coal industry where the problems like wastage could be identified, and this could be proposed as an effective solution and also in scrap industries where the weight of the scrap determines the cost could be monitored by using this simple yet noticeable factor that is determining the weight while offloading the material from the vehicle of transport via a conveyor where the materials could be sent batch by batch to the special conveyor one at a time and then to the recycling plant thus making the process more elementary. Using a PLC that is rugged can be used for various industrial purposes as it is safe and reliable and precise. Today, programmable controllers are a rudimentary segment of best in class robotic arrangements. Smart arrangements are not founded on a contention between these frameworks, but rather on solidarity. With vital preferences like, robustness and simple taking care of, the PLC keeps on ensuring its prosperity [3]. Its further improvement will be reliant on more and more undaunted mix of data and innovation. For the designing procedure, this implies an incorporated work process that covers all aspects. Automating the process reduces the human intervention, thus making it free of human errors and effective in time management. Here in the paper we discuss how we made the prototype that is responsible for sorting is made possible by the use of pneumatic machinery and PLC as a controller for automating the process where we have used a sensor that is load cell to sense the weight of the object [4]. Advance-based sorting system includes the vision-based sorting in order to sort the objects with respect to the colour and shape [5, 6]. In this paper work is carried out as three parts, first section deals about the structures which includes control elements, sensors and pneumatic components. Second section is the design of prototypes of stated system by using ANSYS tool [7]. Third section states the implementation of the designed prototype with the programmable logic controllers.

2 Methodology This work is divided into three parts namely structure, design and implementation. The structure is again divided into three blocks. The first is the bench, where we are mounted the sensors of the system. The second is the pneumatic valve desk, where are mounted the valves, which are handling the actuators. The third part is the controller, what is located under the bench and controlling the pneumatic valves, evaluating the sensory data. In the sorting system, the following types of sensors are used: (1) reflection light scanner (OBT200-18GM70-E5-V1), (2) inductive sensor (NBN8-18GM50-E2) and capacitive sensor (CJ8-18GM-E2). The reflection light sensor is acting in specimen reckoning and counting, while inductive and capacitive sensors are responsible for selections. The training system comprises components

Automation of Weight-Based Sorting System Using Programmable …

355

Fig. 1 NPN proximity sensor

that perform pneumatic movement and manipulation and control of them; these are controlled by a programmable logic controller.

2.1 Structure 2.1.1

Sensors

Proximity Sensor Capacitive-type NPN proximity sensors [8] with a range of 5–34 Vdc can be used to detect metallic and also non-metallic targets like paper, wood, plastic, glass, wood, powder, liquid, etc., without physical contact. The capacitive proximity sensor works on the capacitor principle. The main components of the capacitive proximity sensor are plate, oscillator, threshold detector and the output circuit (Fig. 1).

Strain Gauge It is an industrial strain gauge with input voltage range of 5 V–12 V. It can sense weight from 50 g to 5 kg [9]. When weight is applied on the free end of the instrument. The strain gauge consists of an insulating flexible backing which supports a metallic foil pattern. The gauge is attached to the object by a suitable adhesive, such as cyanoacrylate. As the object is deformed, the foil is deformed, causing its electrical resistance to change. This resistance change is usually measured using a Wheatstone bridge. This output is in millivolts. At a time, we can weight only one object weight using this strain gauge (Fig. 2).

356

P. Chenchu Saibabu et al.

Fig. 2 Strain gauge

Fig. 3 Pneumatic cylinder

Fig. 4 Solenoid valve

2.1.2

Pneumatic Valve Desk

Pneumatic Cylinder These are basic pneumatic cylinders [10] with stroke length of 5 cm. The air is filled from the inlet valve at the back. The air is filled inside the cylinder and it apply force on the piston, and hence, it moves forward. Then, the air exits the cylinder through the exit valve and the piston retreats. The speed of the stroke can be adjusted by adjusting the flow of air which is done by the solenoid valve (Fig. 3).

Solenoid Valve A solenoid valve is an electromechanically operated valve. The valve is controlled by an electric current through a solenoid. Here, we have solenoid valves with 24 V coil. This is an electromagnetic coil. When we supply current, it magnetizes the coil. There is a rod inside the valve that moves up and down due to the magnetization and thus create in and out flow pressure gauge used. The maximum air pressure this system could handle is 10 kg pressure lesser than that is adequate for the system operation (Fig. 4).

Automation of Weight-Based Sorting System Using Programmable …

357

Fig. 5 Conveyor belt

2.1.3

Control Elements

Programmable Logic Controller Programmable logic controller used in this work is Siemens S7 1200 series— CPU1214C, which is capable of handling analogue inputs/outputs and digital input and output signals. Step 7 software is used for the implementation of ladder logics in the programmable logic controllers.

Conveyor Belt We purchased a 75-cm-long conveyor belt, and it’s the structure. There is roller placed on both the ends with marks that prevents them from slipping one of the roller is aligned with a motor, and the other is the free end that can be adjusted according to the strain. By trial and error method, we have observed that it can hold weight up to 1 kg. Depending on the application weight range, we can choose the appropriate conveyor belt (Fig. 5).

2.2 Design 2.2.1

Electrical Circuit Design

Power Supply Circuits Different components that we are using in the model have different input ranges such as the proximity sensors require input of 5 Vdc, the solenoid valves requires input of 12 Vdc and the motor requires input of 6 Vdc. Due to this, we rigged up the conversion circuit which converts the 230 V ac supply to the regulated dc supplies of 5 V, 6 V and 12 V. For the supply conversions in this work, 7805, 7806 and 7812 ICs are used (Fig. 6).

358

P. Chenchu Saibabu et al.

Fig. 6 Circuit for 5, 6 and 12 V

Fig. 7 Amplifier circuit

Amplifier Circuit For measuring weights, we connected an industrial strain gauge. The output of this strain gauge is in terms of millivolts. The analogue input of the PLC cannot read such small changes; thus, we used an amplifier circuit to get the output in terms of volts (Fig. 7).

2.2.2

Mechanical Design

We first designed the model of the prototype deciding the type and the position of the components to be used to build the prototype and made a schematic diagram for the same as shown in Fig. 8. To build the structure of the prototype of the model, we first made a virtual model of the model of the structure using ANSYS tool. We made a 3D model of the structure to make it easier to make the structure. We designed the types of structure we require to support the conveyor belt, the proximity sensors, the strain gauge, the cylinder and the solenoid valves, and we made the model shown in Fig. 9. We considered a conveyor belt of length as 75 cm and breadth as 5 cm. There are total five number of capacitive-type proximity sensors, which have a supply of 24 V DC and give a signal of 24 V DC to PLC inputs when an object is detected.

Automation of Weight-Based Sorting System Using Programmable …

Fig. 8 Prototype model Fig. 9 Model developed by using the ANSYS

359

360

P. Chenchu Saibabu et al.

2.3 Installation 2.3.1

Process Installation

The first one, i.e. proximity sensor 1 detects whether the object is there on the belt. Once it detects the object, it sends a signal to the PLC which in turn sends the signal that starts the motor of the conveyor which needs 6 V to drive the motor; thus, signal turns the motor on and it starts moving. 24 V DC is the output of the PLC and it’s given to a voltage divider with heat sinks to reduce it to 6 V supply as per requirement of the motor. The next proximity sensor is placed just before the strain gauge that we are using. This sensor detects whether the object reached near the strain gauge. This sensor senses the object and sends a signal to the PLC. The object takes some time to reach from proximity sensor to the strain gauge which may be in milliseconds.

2.3.2

Weight Determination

When the proximity sensor 2 gives a signal to the PLC, the PLC is programmed to have a delay that is required by the object to reach form proximity sensor to the strain gauge and then sends a signal to the strain gauge to turn it on which is at 6 V and also turns the motor off of the conveyor. The strain gauge measures the weight of the object. We have decided some ranges of weight for this prototype. Here, we are sorting the material in two different ranges of weights. If the object is of 100 g it goes to rack 1, if it is 200 g it goes to rack 2, the last one is the reject rack. In our prototype, we are sorting the material for 2 different weights. We can increase the number of weights according to our requirement and desire. The strain gauge gave different output for different weights as AC input to the PLC unit. PLC determines the range of weight by the time the delay has got over the motor starts and conveyor starts working, and the signal from PLC is sent to the respective solenoid valve once the respective proximity sensor is turned on.

3 Results 3.1 Sorting For pushing the object into the rack, we are using pneumatic cylinder which are attached to the pneumatic solenoid valve. The pneumatic cylinders have two holes, one is inlet and the other is outlet. The electrical signal from the PLC is sent to the coil of the solenoid valve which is 24 V DC. The valve has a regulated air supply 4 kg which comes from a filter which is used to get a proper supply if air that is free of moisture and dust. The solenoid valve supply air to the cylinder inlet, thus causing the stroke to do the pushing action. Then, the air comes out of the cylinder through

Automation of Weight-Based Sorting System Using Programmable …

361

the outlet causing the retreating motion of the piston, and the object is pushed into the rack. Once the process is done, the value is reset for the next object (Figs. 10 and 11; Table 1). Fig. 10 Sorting of first weight

Fig. 11 Sorting of the second weight

Table 1 Result analysis

Case Weight (g) Pneumatic cylinder 1 Pneumatic cylinder 2 1

100

Activated

No action

2

200

No action

Activated

362

P. Chenchu Saibabu et al.

4 Discussion In the automatic sorting machine [10], they have sorted the materials based on length of the materials. In this technique, they have sorted only 5- and 10-cm-length objects. Expansion of this project based on the object lengths is going to increase complexity in the design, and the structure becomes very large. After the system design and commission, it is not possible to change the sorting parameters. In our project, we have used the sensors to measure the weight and object detection. Control actions are driven by the programmable logic controllers. In our project, we can easily expand the sorting parameters by incorporating the changes in programmable logic controller programming, addition of pneumatic cylinders and sensors.

5 Conclusion This concept makes the sorting process simpler and faster and thus more efficient. Using a programmable logic controller (PLC) as the device used in many industries to monitor and control various activities in production processes is a safe and reliable choice. Our objective was to propose the idea through the paper and make a prototype have been successful as we sorted the material according to the weight of the object. Also using a capacitive-type sensor gave us the chance to sort any type of material regardless metallic or non-metallic making the process feasible for any type of objects. Also the time required by the object to travel from one sensor to another may vary with changes in the weight of the object, thus changing the delay provided in the programming which in turn determines the position of the object as required. This gave us the idea that sorting of the material has been effectively done and has gave us an alternative effective method.

References 1. Bencsik AL, Lendvay M (2014) Sorting system for practical training of mechatronics. In: 9th IEEE international symposium on applied computational intelligence and informatics 2. Manjunatha (2015) Postal automation system for mail sorting. Int J Emerg Technol Adv Eng 5(3). ISSN 2250-2459 3. Bencsik AL, Lendvay M (2002) Industry-institute partnership for PLC education and training. In: IEEE international conference on information technology based higher education and training (ITHET 2002), Budapest, Hungary, 4–6 July 2002 4. Aruna YV, Beena S (2015) Automatic convey or system with in-process sorting mechanism using PLC and HMI system. Int J Eng Res Appl 5(11):37–42, ISSN: 2248-9622 5. Sheth S, Kher R, Shah R, Dudhat P, Jani P, Automatic sorting system using machine vision. In: Multi-disciplinary international symposium on control, automation & robotics, At DDIT, Nadiad, vol 1. https://doi.org/10.13140/2.1.1432.1448 6. Wahab DA, Hussain A, Scavino E, Mustafa MM, Basri H (2006) Development of a prototype automated sorting system for plastic recycling. Am J Appl Sci 3(7):1924–1928

Automation of Weight-Based Sorting System Using Programmable …

363

7. Sen S, Kumar A, Srinivasan CR, Saibabu C (2014) Efficient and low cost flow measurement using bend sensor flowmeter. Int J Electron Electr Eng 7(8):879–885. ISSN 0974-2174 8. Oladapoa B, Baloguna VA, Adeoyea AOM, Ijagbemib CO, Oluwolea AS, Daniyana IA, Esoso A, Asanta A, Simeona P (2016) Model design and simulation of automatic sorting machine using proximity sensor. Eng Sci Technol Int J 19(3):1452–1456 9. Sadani K, Prabhakar DA, Nag P (2018) A cardio pulmonary resuscitation device for stretchers. Int J Eng Technol 7(2.21):62–65 10. Prasad A, Gowtham M, Mohanraman S, Suresh M (2020) Automatic sorting machine. Int Res J Multidisc Technovation 7–12

Design, Development and Verification of a Fault Injection Capable Synchronous Generator Sona Meiyappan, P. Chaithanyasai, S. Swetha, M. Vishnu Deepika, and P. V. Sunil Nag

Abstract Virtual models capable of fault injection are crucial for Fault Diagnosis of Electrical machines like the Synchronous Generator (SG) and are researched widely for this purpose. The SG is used in many significant industries for power generation applications. Inter-turn stator faults account for the majority of SG faults. Multiple methods are being developed to detect these faults in the early stages to avoid industry shutdowns, reduce machine downtimes, and also as a cost-effective maintenance technique. A virtual fault injection capable model of SG is designed and developed to simulate and analyze the effects of inter-turn stator faults. The efficacy of the proposed model was checked by using a model of 10 kVA SG. The dimensions of this SG were calculated theoretically, and the resulting values were fed to ANSYS Maxwell software, where the 2D design of the SG was developed, and faults were introduced to replicate a hardware model closely. Faults were introduced in all three phases under two different load conditions. The system was verified to be fault injection capable with the help of fault signatures found in the stator currents and field current. The second and third harmonic components of the above-mentioned quantities are characteristic of turn-to-turn short circuit (TTSC) faults, and hence were used to validate the developed model. Keywords Fault injection · Synchronous generator · Fault diagnosis

1 Introduction Over the years, the consumption of electric power has grown exponentially. As new devices are added to the grid every year, a need for dependable electrical machines has risen [1]. One such important machine is the Synchronous Generator (SG). A SG S. Meiyappan · P. Chaithanyasai · S. Swetha · M. Vishnu Deepika · P. V. Sunil Nag (B) Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641112, India e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_24

365

366

S. Meiyappan et al.

is an extensively used generator for power generation [2]. Any faults in the SG can lead to a complete industry shutdown and compromise other systems and crew safety. Winding faults are the most common and significant fault in synchronous generators [3] and are mainly caused due to insulation failures [4]. The stator winding faults occur more frequently compared to rotor winding faults. The inter-turn faults occur in the initial stage of the stator winding failures [5] and can lead to complete damage to the machine. Thus, fault detection in the incipient stage has become exceedingly necessary. A crucial step in detecting a fault is to develop a test rig where the behavior of a faulty machine can be studied to develop methods capable of detecting the faults in early stages. Fault injection models are required for this purpose. Fault injection models are of two types: (1) Custom made Hardware setup [6], (2) Virtual machine. A SG has additional components compared with other electrical machines and hence is considered a complex machine. Researchers face difficulties in manufacturing and monitoring such a complex machine. A SG has additional components compared with other electrical machines and hence is considered a complex machine. A fault injection capable model of a SG is designed and implemented in ANSYS Maxwell to simulate both healthy and faulty generator data. Our work in this paper attempts to validate our virtual machine against a hardware setup. If verified, this model gives researchers the freedom to introduce different types of faults, those that may not be practically realizable with a hardware setup. Further, the design and implementation of this virtual setup discussed in this work is cost-effective and any aspiring researcher can easily design and develop a SG for research and exploration in the field of fault diagnosis. Analysis of the fault signature in the third harmonic component of stator current for a SG with R Load has been discussed in [7]. The main contribution of this paper is that it further builds on the mentioned work, by verifying the same for a SG with RL Load. Additionally, the second and third harmonic components of both stator and field currents were plotted and shown to be characteristics of a short circuit fault. The layout of the paper is as follows. The flowchart is displayed in Sect. 2. In Sect. 3, the design of SG is presented. Section 4 provides the Maxwell model obtained from the design data. Section 5 discusses about the injection of fault in ANSYS Maxwell. The plots of the induced currents and voltages obtained from the Maxwell model are given in Sect. 6. The results of the frequency analysis of various currents are discussed in Sect. 7.

2 Flowchart of the Working Model See Fig. 1.

Design, Development and Verification of a Fault Injection Capable …

367

Fig. 1 Flowchart

3 Design of a Synchronous Generator A Synchronous Generator (SG) was designed with the basic specifications of 10 kVA power, a Y connected load with a 36 slot stator, a 4 pole cylindrical rotor, a line to line voltage of 628 V rms , with a maximum load current of 4.5 A and a power factor of 0.8. The synchronous speed of the generator is 25 r.p.s. and has 220 V DC excitation and 50 Hz frequency. The design parameters required for the RMxprt are calculated referring chapter12 in [8]. A 10 kVA Synchronous Generator was designed theoretically and the dimensions were obtained with suitable assumptions for average mechanical loading (Bav = 0.4 Wb/m2 ) and average electrical loading (ac = 13,000 A/m).

4 Design of SG Using RMxprt and ANSYS Maxwell RMxprt is a template-based electrical machine design tool developed by ANSYS that is used to evaluate a design. After designing the analytical model of the generator, the performance of the design can be verified using ANSOFT RMxprt. This RMxprt

368

S. Meiyappan et al.

Fig. 2 RMxprt model of the SG

model can then be converted in 2D or 3D geometric creation in ANSYS Maxwell to perform finite element calculations [9]. Combined together RMxprt and Maxwell fully meet the electric machine design and analysis needs [10]. In the presented work, the analytical model of the generator is simulated in RMxprt (Fig. 2). After the design requirements are met and verified using RMxprt, this model is then used for 2D transient analysis using ANSYS Maxwell. A load of 1.5 A is used in all cases presented in this paper (Fig. 3). The SG’s symmetry is a significant component in obtaining a desirable output voltage [11]. Hence, the stator slot coils are split equally in the ANSYS Maxwell 2D model (Fig. 4). Fig. 3 ANSYS Maxwell model of the SG

Design, Development and Verification of a Fault Injection Capable …

369

Fig. 4 Circuit before splitting of coils

5 Fault Injection in ANSYS Maxwell Fault is introduced by splitting the stator coils of all three phases. Stator winding of each phase consists of 30 turns and a split in coil is introduced between 7 and 8 turns of the same phase [7] (Fig. 5).

Split coil

Unsplit coil

Fig. 5 Enlarged view of a spilt and unsplit coil

370

S. Meiyappan et al.

Fig. 6 Circuit after splitting of coils

An equivalent circuit of the Maxwell 2D model with the spilt in coils is developed in the circuit editor of ANSYS Maxwell software (Fig. 6). The fault is introduced by triggering a short circuit by adding a switching device to the external circuit, which is designed to close after a certain time [12]. Simulations are run for 1.5 s at a sampling frequency of 1000 Hz. It was observed that 0.6 s was necessary for the SG to attain the steady state and hence fault was introduced at 0.8 s, resulting in 0.2 s of healthy data and 0.7 s of faulty data. This external circuit is then manipulated to introduced fault into three phases each under two different load conditions namely, Resistive Load (R Load), and Resistive Inductive Load (RL Load), resulting in a total of 6 scenarios. AC motors like induction motors are the most common loads associated with SGs and these loads can be modelled as a combination of resistances and inductances. Hence, a R Load and RL Load cases were chosen for this study. A total of 6 simulation cases are run on a High-performance computing (HPC) machine. Induced and field current plots are obtained from each case for further analysis (Figs. 7, 8, 9 and 10).

6 ANSYS Maxwell Results Simulation values from Induced Current plot (stator currents) and Field Current plot after simulation are exported to MATLAB for FFT Analysis.

Design, Development and Verification of a Fault Injection Capable …

371

Switch R Load

Trigger

Fig. 7 External circuit of R load case: fault in phase A

Switch RL Load

Trigger

Fig. 8 External circuit of RL load case: fault in phase A

Fig. 9 Induced and field current plots (R load with fault in phase A)

372

S. Meiyappan et al.

Fig. 10 Induced and field current plots (RL load with fault in phase A)

7 Fast Fourier Transform 7.1 Results from FFT for Analysis of Stator Current The frequency analysis of stator current shows that the third harmonic component proves to be a suitable indicator for the detection of Turn-to-Turn short circuit (TTSC) fault. Figure 11 displays the stator currents of all the three phases for a Synchronous Generator under healthy operation for R Load condition. Figure 13 represents the plot of the stator currents for a machine under TTSC fault for R Load Condition. Comparing the two figures, it can be observed that the third harmonic component occurring at 150 Hz appears only when a fault is introduced. It can be inferred from the above analysis that the third harmonic component of stator current is a characteristic of TTSC fault, as stated in [13]. Figures 12 and 14 show the same frequency spectrum under RL Load condition for healthy and faulty conditions, respectively, and the same results obtained above were observed.

7.2 Results from FFT for Analysis of Field Current Fast Fourier Transform is performed on the field current to study the harmonic components. Figures 15 and 17 show the frequency analysis plots for a Synchronous Generator under healthy and faulty conditions, respectively, for R Load condition. Analysis of the plots shows that the second harmonic component occurring at 100 Hz arises only when a fault is injected in the machine. According to [13], this is a verified signature for the presence of fault. Hence, it is proved that the fault has been injected

Design, Development and Verification of a Fault Injection Capable …

373

Fig. 11 Frequency spectrum of stator current of a healthy machine under R load condition for phases A, B and C

Fig. 12 Frequency spectrum of stator current of a healthy machine under RL load condition for phases A, B and C

in the developed machine. As done previously, the machine was again simulated for RL Load condition to further verify that the fault signatures function for all linear load cases. Figures 16 and 18 show the frequency spectrum of field current of the machine under RL Load condition for healthy and faulty cases, respectively.

374

S. Meiyappan et al.

Fig. 13 Frequency spectrum of stator current under TTSC fault for R Load condition in Phases A, B and C

Fig. 14 Frequency spectrum of stator current under TTSC fault for RL load condition in phases A, B and C

Design, Development and Verification of a Fault Injection Capable …

375

Fig. 15 Frequency spectrum of field current of a healthy machine under R load condition for phases A, B and C

Fig. 16 Frequency spectrum of field current of a healthy machine under RL load condition for phases A, B and C

376

S. Meiyappan et al.

Fig. 17 Frequency spectrum of field current under TTSC fault under R load condition in phases A, B and C

Fig. 18 Frequency spectrum of field current under TTSC fault under RL load condition in phases A, B and C

Design, Development and Verification of a Fault Injection Capable …

377

Table 1 Comparison of results with prior art Results obtained in this paper

Results from existing work

Appearance of third harmonic component of stator Variation of 180 Hz (third harmonic) current under TTSC fault refer Figs. 13 and 14 component of terminal voltage under TTSC fault refer Figs. 4 and 5 of [14] Appearance of 2nd harmonic component of field current when a 7% stator winding fault is introduced refer Figs. 17 and 18

Increase in magnitude of 2nd harmonic component after introduction of 15%, as compared to the healthy machine. Refer Fig. 23a, b of [13]

7.3 Comparison of Results Obtained Table 1 compares the frequency spectrum obtained in this work with existing literature. From the comparison it can be concluded that the fault injection capable synchronous generator can simulate TTSC fault.

8 Conclusion In this paper, the development of a Fault Injection capable model of a Synchronous Generator is presented. A Synchronous Generator was designed for a power of 10 kVA and was simulated using the ANSYS Maxwell software. Simulation was done for various load conditions, with the introduction of fault into the machine. To verify the presence of fault after its introduction in the machine, frequency analysis was done on the stator current and on the field currents obtained from the simulation. The appearance of second and third harmonic components of the stator current and field current confirm the presence of fault in the developed machine. Conflicts of Interest The authors declare that they have no conflicts of interest to report regarding the present study.

References 1. DKG, Menon SR, Sreevathsava N, PM, Nag PVS, Kumar CS (2020) Stator inter-turn fault diagnosis and fault location in synchronous generator using dual extended Kalman filter and linear regression analysis. In: 5th international conference on communication and electronics systems (ICCES) 2020, Coimbatore, India, pp 52–57 2. Bruzzese C (2014) Diagnosis of eccentric rotor in synchronous machines by analysis of splitphase currents–part II: experimental analysis. IEEE Trans Ind Electron 61(8):4206–4216 3. IEEE Recommended Practice for the Design of Reliable Industrial and Commercial Power Systems—Redline. In: IEEE Std 493–2007 (Revision of IEEE Std 493–1997)—Redline, pp 1–426

378

S. Meiyappan et al.

4. Jee-Hoon J, Jong-Jae L, Bong-Hwan K (2006) Online diagnosis of induction motors using MCSA. IEEE Trans Ind Electron 53(6):1842–1852 5. Riera-Guasp M, Antonino-Daviu JA, Capolino GA (2015) Advances in electrical machine, power electronic, and drive condition monitoring and fault detection: State of the art. In: IEEE Trans Ind Electron 62(3):1746–1759 6. Gopinath R, Kumar C, Ramachandran K (2016) Scalable fault models for diagnosis of synchronous generators. Int J Intell Syst Technol Appl 15(1):35 7. Kamala KS, Induvadhani VV, Lakshmi VI, Mithra P, Sunil Nag PV, Kumar CS (2020) Electrical signature analysis (ESA) of a fault injection capable synchronous generator for inter-turn stator faults. In: 2020 5th international conference on communication and electronics systems (ICCES) Coimbatore, India, pp 171–175 8. Sawhney AK (1984) Electrical machine design, 1st edn. Dhanpat Rai, India, pp 702–833 9. MK, Warrier GS, Pathivil P, Kanagalakshmi S, Archana R (2019) Design and performance analysis of brushless DC motor using ANSYS Maxwell. In: 2019 2nd international conference on intelligent computing, instrumentation and control technologies (ICICICT), Kannur, Kerala, India, pp 1049–1053 10. Allirani S, Vidhya H, Aishwarya T, Kiruthika T, Kowsalya V (2018) Design and performance analysis of switched reluctance motor using ANSYS Maxwell. In: 2018 2nd international conference on trends in electronics and informatics (ICOEI), Tirunelveli, pp 1427–1432 11. Dallas SE, Safacas AN, Kappatou JC (2011) Interturn stator faults analysis of a 200-MVA hydrogenarator during transient operation using FEM. IEEE Trans Energy Convers 26(4):1151– 1160 12. Prasob K, Kumar NP, Isha TB (2017) Inter-turn short circuit fault analysis of PWM inverter fed three-phase induction motor using finite element method. In: 2017 international conference on circuit, power and computing technologies (ICCPCT), pp 1–6 13. Nadarajan S, Panda SK, Bhangu B, Gupta AK (2015) Hybrid model for wound-rotor synchronous generator to detect and diagnose turn-to-turn short-circuit fault in stator windings. IEEE Trans Ind Electron 62(3):1888–1900 14. Sottile J, Trutt F, Leedy A (2006) Condition monitoring of brushless three-phase synchronous generators with stator winding or rotor circuit deterioration. IEEE Trans Ind Appl 42(5):1209– 1215

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars K. Bipin and P. V. Sunil Nag

Abstract A self-driving car is one of the engineering marvels in recent time due to autonomous capabilities. An autonomous car can sense its surroundings through various integrated sensors and navigate the vehicle based on sensor data with Model Predictive Control System. Vehicle Navigation in self-driving car helps to determine the vehicle position by localization through high precision Global Navigation Satellite System and Inertial Navigation System (GNSS/INS) technologies with various sensors like Light Detection and Ranging (LIDAR), Radio Detection and Ranging (RADAR) and Global Positioning System (GPS) Sensor. Vehicle Navigation is completely depending up on the localization of the vehicle. The most popular method like particle filter and Kalman filter-based approach lacks stability with random particle distribution, and some of the GNSS-based methods are very expensive in terms of implementation. So here introducing an unscented Kalman filter and H-Infinity filter (UKF/H-Infinity)-based novel approach for the low-cost localization technique in self-driving cars. This robust control model helps to handle the position error from the GPS sensor with H-Infinity filter and UKF by the help of Constant Turn Rate Acceleration (CTRA) motion model. This work includes the detailed discussion of pose estimation of vehicle with better yaw stability and minimum computational overhead in estimation. Keywords CTRA · GPS · H∞ · LIDAR · PF · UKF

1 Introduction Vehicle positioning or localization is the primary information needed for any Autonomous Transport System. This information can be collected by using various integrated sensors and methods. Self-driving cars use sensors and maps for vehicle positioning and understanding its surrounding infrastructure. The existing methods K. Bipin · P. V. Sunil Nag (B) Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641112, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_25

379

380

K. Bipin and P. V. Sunil Nag

available for determining the vehicle position in self-driving cars have numerous problems and challenges. GNSS suffers from accuracy problem [1] in Urban Environment, and Simultaneous Localization and Mapping (SLAM) methods have error accumulation problem [2]. Another state-of-the-art localization technique 3DLIDAR also has same problem [3]; this technique observes the surrounding infrastructure and compares it with prior known 3D point cloud for estimating the position of the vehicle in the map [4]. Here, we need massive storage space on the vehicle, or simultaneous cloud access is necessary. Even though GPS has low accuracy problem [5], we can improve the performance by using various algorithms.

2 Background Understanding the localization process [6] along with various models is a timeconsuming process, because each model depends on different dynamics of the vehicle. GPS-based localization technique is commonly used, but we cannot completely rely on GPS as the only solution due to noise and its poor performance in indoor and urban environments. Accuracy in any localization technique is crucial for self-driving cars, because lack of accuracy leads to road accidents. We went through the study of various approaches used in vehicle localization. Some of the methods referred for understanding the problem statement of the work are as follows. Odometry is a technique in localization where wheel displacement calculation is used for the estimation of vehicle position, but this technique is inaccurate [7] since some vehicle parameters are not taken into account for prediction. Another popular method called SLAM uses the MAP [8]. The reference paper taken here for the implementation has Kalman and particle filter-based approach [9] that compares the observations of range sensors with the surrounding infrastructure map and creates particles around the ego vehicle by using an iterative approach to make similar observation as that of map.

3 UKF/H-Infinity Filter We are proposing a new approach called UKF/H-Infinity robust control model with CTRA motion model. The H-Infinity filter is commonly used in control theory due to its robust nature with better stability; it can also be used in GPS/INS [10] and here we are using UKF [11], as optimal control. The H-Infinity filter estimates better position from noisy GPS sensor data by considering the acceleration and velocity noise from the Inertial Measurement Unit (IMU) fusion filter. The UKF motion prediction can be achieved in process space by using the CTRA motion model and the state position input from the H-Infinity algorithm. When these two algorithms are combined, we can call this as mixed model or UKF/H-Infinity robust control model. H-Infinity algorithm helps to reduce the

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

381

estimation error and gives better performance with constrained environment, but sometimes outliers affect the performance of the algorithm. But choosing maximum bounded value that can tolerate the outlier will help to avoid this problem. H-Infinity filter’s robustness nature of handling [12] unknown noise resources help to reduce noise from the localization system. Figure 1 is the architecture of UKF/H-Infinity robust model controller, where we can see the IMU fusion filter with GPS, UKF/H-Infinity filter model and LIDAR sensor module in it. From the IMU/GPS fusion filter, we get the ego vehicle’s lateral and longitudinal parameters like linear velocity, angular velocity, orientation of the vehicle and noisy GPS position of the vehicle. While implementing model in MATLAB, we can add Gaussian and non-Gaussian noise in the IMU fusion filter for simulating the real-world scenario. The input of H-Infinity controller is the noisy GPS position from the fusion filter, and it consists of the three states (x, y, heading angle), angular velocity, linear velocity and acceleration noise. After the H-Infinity process estimation, a new state vector with size 3 is given as the input to UKF filter. The UKF filter again does the estimation by using the measurement space value from the LIDAR, and the final estimation result will be the same size as input with best accuracy and less computational overhead. The detailed description of H-Infinity filter is shown below.

Fig. 1 UKF/H-Infinity model architecture

382

K. Bipin and P. V. Sunil Nag

3.1 H-Infinity Filter The H-Infinity filter [13] controller shown in Fig. 2 and general equation can be written (1) as follows, where A, B are known matrices, and we are estimating the state considering the unknown noise. X k+1 = A ∗ X k + B ∗ U + K ∗ Mv

(1)

If we consider the velocity, acceleration and angular velocity in model, it can be written in the following form. Here, state vector “X” has the position of vehicle from GPS sensor with the external noise. “U” is the unknown vehicle acceleration (without the noise). M v is the measured velocity affected by the measurement “z” noise from the plant. K has two elements, and it is considered as H-Infinity gain and is constantly updated through the process. The system model can be formulated (2) with some uncertainty. x(k+1) = (A + Aunc ) ∗ xk + (B + Bunc ) ∗ u k + wk

(2)

xnoise,max = xk + xoffset

(3)

ynoise,max = yk + yoffset

(4)

  xk+1 = xk + xnoise,max . cos(θh )

(5)

  yk+1 = yk + ynoise,max . sin(θh )

(6)

Fig. 2 H-Infinity control block

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

X h f,k+1

   x = = y

vk [sin(θk + θk (t)) − sin(θk )] θk t vk [cos(θ k ) − cos(θk + θk (t))] θk t

383

 (7)

The motion model (7) predicts the position of the vehicle by considering the reduced positional error in the state space x and y with H-Infinity filter. This is considered as one of the simplest vehicle models generally used for the reduction of computational expense in vehicle motion model. The advantage of using this model is that it is only considering few parameters (8) for lateral and longitudinal motion, and hence, it does not affect the yaw rate stability of the vehicle. The motion model will be added to the state space matrix of the general equations with acceleration and velocity noises. The dynamics of Bicycle Vehicle Motion Model [14] is shown in Fig. 3. The outliers are the boundaries of the noise; x noise,max is the maximum noise in x-direction (3) and ynoise,max is the maximum noise in y-direction (4). The x offset and yoffset are the outlier offset parameters and Eqs. (5) and (6) are constraints with this offset.  Vpos = x y θ x Ego vehicle position in x-direction y Ego vehicle position in y-direction

Fig. 3 Bicycle vehicle motion model

(8)

384

K. Bipin and P. V. Sunil Nag

θ Heading angle v Velocity of ego vehicle. “A” is the transition matrix with dimension “m × m” having the time space between the previous and next position, “B” is the input matrix with dimension “m × n” having the position estimation through the motion model, “C” is the measurement matrix with dimension “n × m”, which keeps updated by the previous measurement state and X hf is the initial state vector with dimension m × n; where m = 2 and n = 1. The initial measurement state “Y ” is initialized below with state and measurement update. Y = C ∗ X hf

(9)

The measurement can be initialized by measurement matrix C and initial state vector X hf by using Eq. (9). The H-Infinity gain controls the algorithm by using the below equations, we can apply the steady state gain if the system performance is poor or else, and we can use time-invariant gain to get better control over the computation. This can be achieved by Eqs. (10) and (11). L = (I − g ∗ Q ∗ P + C T ∗ V −1 ∗ C ∗ P)

(10)

K = A.P. L −1 . C T . V −1

(11)

Here, “I” is identity matrix, M v is the measurement variable, M Noise measurement noise, PNoise process noise, Q weight matrix, P covariance matrix, K is H-Infinity gain and mNoise nominal measurement values. Measurement state (14) and velocity update (15) initialized with previous state. Equation (13) forces the matrix P (12) to be symmetric.  P=

10 01

 (12)

  P = P + P T /2

(13)

Mx = Y − C ∗ X k

(14)

Mv = Y − C ∗ Vk

(15)

Process noise and measurement noise can be calculated by Eqs. (16), (20), where we are considering the acceleration noise and “B” matrix.  PNoise = 2 ∗ aNoise ∗ B ∗

R1 R2

 (16)

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

385

X k+1 = A ∗ X k + B ∗ U + K ∗ Mx

(17)

Vk+1 = A ∗ Vk + B ∗ U + K ∗ Mv

(18)

X k+1 = A ∗ X k + B ∗ U + K ∗ PNoise

(19)

MNoise = m noise ∗ (rand − 0.5)

(20)

Here in Eq. (17), we can see two known and two unknown matrices, and we can consider the known matrices as the acceleration and velocity of the ego vehicle and unknown matrices as the noise of both. The noise is always coming with acceleration and velocity of ego vehicle from the fusion filter module due to the external or internal disturbance. We have taken acceleration and velocity as the system model parameter and angular velocity as the external parameter to update the state (19) and velocity (18) with measurement state. This can be applied in adaptive filter modelling, and it helps the time invariant system to adapt the relative noise with constrained boundaries. This method outperforms the particle filter due to its robust nature.

3.2 UKF The UKF [15] is suitable for nonlinear space motion model. UKF takes the points from the whole distribution called sigma points and Gaussian distribution given into a nonlinear function [16] called unscented transformation, and this will generate the corresponding sigma points in the prediction and measurement model with the state space vector [17]. Then, we calculate the mean of state vector and covariance matrix from these corresponding sigma points to find the Gaussian distribution of the predicted or measurement state space. By taking the correlation of both state and covariance matrix, we get the Kalman gain and using this and LIDAR measurement space, the state and covariance matrix will be updated. The overall workflow of the UKF algorithm is shown in Fig. 4 for better understanding. UKF Prediction Generating sigma points from the state vector (21) is the first step involved in the UKF algorithm, and it is considered as the symmetrical region around the mean value of state. We augment this state matrix with acceleration and velocity noise, then the resultant state vector is given to the UKF prediction. In the UKF prediction, CTRA motion model is used for the prediction of new state. The CTRA model is a nonlinear motion model used for vehicle state estimation, and this model has six state vectors, and it can be converted to eight by adding additional augmented state to prediction model.  T X CTRA = x, y, v, a, h, w

(21)

386

Fig. 4 UKF workflow

K. Bipin and P. V. Sunil Nag

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

387

Here, x is the ego vehicle position in x-direction, y is the ego vehicle position in y-direction, v and a are the velocity and acceleration of the vehicle and w are the heading angle and heading turn rate of the vehicle. ⎤ ⎡ ⎤ (vω + aωT ) · sin(h + ωT ) xhfukf,k ⎢ (−vω − aωT ) cos(h + ωT ) ⎥ ⎢ yhfukf,k ⎥ ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎢ v + aT ⎥ 0 ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎢ ⎥ 1 ⎢ 0 ⎥ ⎢ a ⎥ X hfukf,k+1,aug = ⎢ ⎥ ⎥+ 2⎢ ⎥ ⎢ h + ωT ⎥ ω ⎢ 0 ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎢ ω ⎥ 0 ⎥ ⎢ ⎢ ⎥ ⎦ ⎣ ⎣ 0 ⎦ 0 0 0 ⎤ ⎡ a ∗ cos(h + ω.T ) − vω sin(h) − a ∗ cos(h) ⎢ a ∗ sin(h + ωT ) + vω cos(h) − a ∗ sin(h) ⎥ ⎥ ⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ 1 ⎢ 0 ⎥ ⎥ ⎢ 2 ⎥ 0 ω ⎢ ⎥ ⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎦ ⎣ 0 0 ⎡

(22)

The motion model can be written as (22), where acceleration parameter is added in the model to achieve the best performance in cornering. Here, we have four-velocity models for the ego vehicle. This makes the system to work more efficient in handling noise in cornering. Usually, CTRA model gives better results than CTRV in urban environment [18, 19] due to handling the acceleration parameter in the motion model equations. UKF Updates Here, both process space and measurement space state mean and covariance will be updated by taking the correlation of both spaces. By taking the correlation of these two spaces we compute the Kalman gain. The state and covariance will be updated using LIDAR measurement values and the Kalman gain. Finally, we are getting the processed state estimation from UKF/H-Infinity filter with reduced positional error and better stability.

4 Implementation The implementation of UKF/H-Infinity filter is done in MATLAB by using various MATLAB toolboxes. Some of the MATLAB toolboxes are described below with respective area of development in this approach.

388

K. Bipin and P. V. Sunil Nag

4.1 IMU–GPS Fusion Filter The IMU fusion filter [20] is created by using MATLAB sensor fusion and tracking toolbox for generating the true and noisy GPS data and other dynamics parameters. For understanding these parameters with noise, we can add Gaussian and nonGaussian noise into the virtual sensor. Tables 1 and 2 show IMU and GPS noise parameter used in the work.

5 Simulation Result MATLAB automated driving toolbox and driving scenario toolbox are used for creating road models for testing different velocity profiles with ego vehicle. I-Road model has a vehicle profile “Fig. 5a” with virtual LIDAR/RADAR sensor embedded in it and have road length of 60 m. Here in S-Road model “Fig. 6b”, the vehicle profile is same as the I-Road model except the road length of 4 km. The accuracy of algorithm is given in Table 3 and can observe that this approach [20] works well in Gaussian noise. In real-world applications, we can’t expect Gaussian noise, so here we had concentrated on non-Gaussian noise. The UKF/HInfinity approach gives better result in non-Gaussian noise with CTRA motion model. Apart from UKF/PF uses the random distribution of particle in state prediction, hence the accuracy of estimation changes on each test cycle, while the ego vehicle taking cornering. UKF/H-Infinity filter circumvent this problem by considering the H-Infinity filter in this approach. Table 1 IMU noise parameter table S. No.

Noise parameter (m/s2 )

Accelerometer

Gyroscope

Magnetometer

19.62

4.363

1200

1

Measurement

2

Noise density

0.0012361

0.0000872

[ 0.6 0.6 0.9 ]

3

Resolution

0.0023965

0.00013323

0.1

Table 2 GPS noise parameter S. No.

GPS sensor parameter

Non-Gaussian noise

1

Horizontal position accuracy

Normrnd (9.65, 12.2)

2

Vertical position accuracy

Normrnd (8.34, 12.33)

3

Decay factor

0.80

5

Reference location

[ 40 − 50 30 ]

6

Random stream

mt19937ar with 100 seed

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

(a) I-Road Model

389

(b) S-Road Model

Fig. 5 Road models

(a) H-Infinity Filter Estimation

(b) UKF Estimation

Fig. 6 I-road model vehicle profile

Table 3 UKF/H-infinity filter position accuracy table Velocity

Noise

UKF/PF

UKF/PF

UKF/H-infinity

UKF/H-infinity

km/h

m

GA (m)

Non-GA (m)

GA (m)

Non-GA (m)

60

29

1.4510

1.6552

0.9822

1.2182

80

29

1.6519

1.4402

0.9713

1.2477

100

29

1.5012

1.6167

0.9408

1.2938

120

29

1.7721

1.4541

0.9900

1.2626

5.1 I-Road: Ego Vehicle at 60 km/h See Fig. 7.

5.2 S-Road: Ego Vehicle at 60 km/h See Figs. 8 and 9.

390

K. Bipin and P. V. Sunil Nag

Fig. 7 I-road model—UKF/H-infinity position estimation of ego vehicle

(a) H-Infinity Filter Estimation

(b) UKF Estimation

Fig. 8 S-road model vehicle profile

6 Conclusion The UKF/H-Infinity robust model is implemented in MATLAB with four different velocity profiles and two road models. From the simulation result, we can conclude that the accuracy of the algorithm increases marginally when compared with UKF without H-Infinity filter and from the figures we can view the stability in the position estimation in cornering. We can further improve the accuracy by doing rigorous tuning of the filter. This approach uses GPS and 2D/360/LIDAR for localization but the implementation of UKF/H-Infinity filter with 3D-LIDAR measurement can be a future enhancement of this approach.

UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars

391

Fig. 9 S-road model—UKF/H-infinity position estimation of ego vehicle

References 1. Lohan ES, Borre K (2016) Accuracy limits in multi-GNSS. IEEE Trans Aerosp Electron Syst 52(5):2477–2494 2. Jiang G, Yin L, Jin S, Tian C, Ma X, Ou Y (2019) A simultaneous localization and map-ping (SLAM) framework for 2.5D map building based on low-cost LiDAR and vision fusion. Appl Sci 9(10):2105 3. Chen CP, Xu L, Danping C, Qi Z, Zhu Y, Li Y, Tao (2019) Trajectory optimization of LiDAR SLAM based on local pose graph. China satellite navigation conference (CSNC) 2019 proceedings, p 360370 4. Yin H, Berger C (2017) Mastering data complexity for autonomous driving with adaptive point clouds for urban environments. In: 2017 IEEE intelligent vehicles symposium (IV) (2017), pp 1364–1371 5. Pawlowski E (2015) Experimental study of a positioning accuracy with GPS receiver. In:12th conference on selected problems of electrical engineering and electronics WZEE’2015, Kielce, Poland, vol 12 6. Woo A, Fidan B, Melek WW, Zekavat S, Buehrer R (2019) Localization for autonomous driving. In: Handbook of position location: theory, practice, and advances, IEEE, pp 1051–1087 7. Goh V, Fischbeck P, Gerard D (2007) Identifying and correcting errors with odometer readings from inspection and maintenance data: rollover problem for estimation of emissions and technical change. Transportation research record 8. Frintrop S, Jensfelt P, Christensen HI (2006) Attentional landmark selection for visual SLAM. In: IEEE/RSJ international conference on intelligent robots and systems, pp 2582–2587. https:// doi.org/10.1109/IROS.2006.281711 9. Lin M, Yoon J, Kim B (2020) Self-driving car location estimation based on a particleaided unscented Kalman filter. Sensors (Basel) 20(9):2544. https://doi.org/10.3390/s20092544. PMID: 32365721; PMCID: PMC7249166 10. Tsogas M, Polychronopoulos A, Amditis A (2005) Unscented Kalman filter design for curvilinear motion models suitable for automotive safety applications. In: 7th international conference on information fusion, p 8. https://doi.org/10.1109/ICIF.2005.1592006

392

K. Bipin and P. V. Sunil Nag

11. Jiang C, Zhang S-B, Zhang Q-Z (2016) A new adaptive H-infinity filtering algorithm for the GPS/INS integrated navigation. Sensors 16(12):2127 12. Nag S, Silla G, Gummadi V, Harishankar B, Ray V, Kumar C (2016) Model based fault diagnosis of low earth orbiting (LEO) satellite using spherical unscented Kalman filter, IFACPapersOnLine 49, pp 635–638. https://doi.org/10.1016/j.ifacol.2016.03.127 13. Warrier E, Nag S, Kumar C (2019) A novel H-infinity filter based indicator for health monitoring of components in a smart grid. Energy transfer and dissipation in plasma turbulence, pp 221– 230. https://doi.org/10.1007/978-981-13-59538 14. Polack P, Altch´e F, Novel B, de La Fortelle A (2017) The kinematic bicycle model: a consistent model for planning feasible trajectories for autonomous vehicles? In: 2017 IEEE intelligent vehicles symposium (IV), pp 812–818. https://doi.org/10.1109/IVS.2017.7995816 15. Sudheesh P, Jayakumar M (2018) Nonlinear tracking using unscented Kalman filter. In: Advances in intelligent systems and computing, vol 678, pp 38–46, Springer Internal, Advances in Signal Processing and Intelligent Recognition Systems 16. Sudheesh P, Jayakumar M (2018) Tracking of nonlinear variations of the parameters of high mobility systems. Int J Pure Appl Math 118(7):221–226 17. Schubert R, Adam C, Obst M, Mattern N, Leonhardt V, Wanielik G (2011) Empirical evaluation of vehicular models for ego motion estimation. IEEE Intell Veh Symp (IV) 534–539. https:// doi.org/10.1109/IVS.2011.5940526 18. Schubert R, Richter E, Wanielik G (2008) Comparison and evaluation of advanced motion models for vehicle tracking. In: 11th international conference on information fusion, pp 1–6. https://doi.org/10.1109/ICIF.2008.4632283 19. Shapovalov I, Maximychev E, Gafurov S, Ostankovich V, Fedorenko R (2020) Robust localization of a self-driving vehicle in a lane. In: 4th scientific school on dynamics of complex networks and their application in intellectual robotics (DCNAIR), pp 210–213 20. Caron F, Duflos E, Pomorski D, Vanheeghe P (2006) GPS/IMU data fusion using multisensor Kalman filtering: introduction of contextual aspects. Inf Fusion 7:221–230. https://doi.org/10. 1016/j.inffus.2004.07.002

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer S. Rajesh Kannan , V. G. Rajagopalan , H. Ramakrishnan , S. Sibi Selvan , and Sushanth Krishnamithran

Abstract An oil spill is the release of liquid hydrocarbon into the environment especially in oceans and is a form of man-made disaster. 35.7% of oil spilt in oceans is from oil tanker vessels. This hugely affects the marine ecosystem and also tourism, marine economy and the human health. Thus, oil spill is a biological, economical and a social disaster. This paper idea aims to detect the oil spilt from vessels accurately and also as early possible with the help of different sensors and mechanisms. This idea not only stops at that but also makes use of an oil skimming that will be pressed into action as soon as oil spill is detected. An ESP8266 microcontroller, a GPS module and also Wi-Fi signals connected to an app interface to control the entire process. This idea is a true and unique solution that employs modern technologies like IoT as well as an app interface to detect and clean up the oil spilt making it highly stable, efficient and also economical. The bot which is used to clean up the oil spilt makes use of a driver module as well as belt-based skimming mechanism to do its job efficiently. We have designed this project to be controlled manually in case of any exegesis without much strain making the response without any time lag and making recovery as soon as possible. Keywords Load cell · GPS · L298n · Belt-type oil skimmer · IoT · GUI

1 Introduction Pollution has become a major threat to the environment. There are various kinds of pollution such as water pollution, air pollution and land pollution, out of which there is another kind of major pollution that harms aquatic animals and their ecosystem called oil spill. Oil spill causes a thick dense layer of oil to be formed on the surface of ocean which prevents the sunlight from entering the surface and thereby leading S. Rajesh Kannan (B) · V. G. Rajagopalan · H. Ramakrishnan · S. Sibi Selvan · S. Krishnamithran Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering, Chennai 600119, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_26

393

394

S. Rajesh Kannan et al.

to decrease in dissolved oxygen content in water bodies. Oil spill caused by crude oil is worse, and the crude oils have very high viscosity and are transported in ships in heated temperature which causes thermal pollution and increases the water temperature in the event of oil spill. According to the study made by Center for Biological Diversity, the deep horizon oil spill which happened in New Mexico caused harm or killed more than 82,000 birds of 102 species, 6165 sea turtles and as many as 25,900 marine species. It also resulted in over 25,000 job loss, $2.3 billion loss in industrial output and an overall $1.3 billion loss in gross regional product, effects of which are even felt now. The latest oil spill in Mauritius resulted in the economic shrink of $2.1 billion of the total $15 billion tourist-dependent economy. Even the small level of oil spill in Ennore, India, caused severe damage to marine economy for weeks in surrounding regions. The oil spill incidents lead to much harmful impact due to the non-availability of portable equipment and a matured industry. We have tried to address these problems with the help of our device. Most of the oil is transported in oceans and most of which are hard crude oil. These oils are transported in the hull of the ships, which on damage lead to the spillage of oil in the oceans. We have designed a primary detection system that measures the volume of oil stored in the hull of the ship at regular time intervals and alerts in case of any spillage of oil into the oceans. By this way, we can immediately predict the disaster so that we will be getting the most required precious moments to prevent a mishap. As soon as the oil spill is detected, swarm of autonomous oil skimmers will be pressed into action. These devices can be controlled remotely with the help of IoT and cloud technologies, thus increasing the speed of action. These devices have installed GPS modules so that each and every movement of these bots can be predicted accurately. The devices can be pressed into action either by paradropping, remote release from land or sea or immediate deployment from the vessel itself. They are propelled with the help of rotors, and their stability is maintained with the help of flaps. These rotors are in turn connected to controllers that have an established connection with command centers. Once these bots reach the desired location, the flaps are deployed to maintain their stability and the skimming process will be initiated. Since these devices are small in size and have less weight, they can be easily mobilized and can be quickly deployed. We make use of belt-type oil skimmers instead of other types of oil skimmers. We have preferred belt-type oil skimmer because it has fewer mechanical parts, high efficiency and less weight than others. These skimmed off oils are then collected in a container whose size depends upon the number of belts used in the skimming process. As soon as the container is filled, the bot is propelled to a collection place where these oils can be collected. The typical capacity of our device is 5–7 L per hour whose capacity can be further increased with the help of addition of belts. Swarm of these bot deployed sooner based on the detection system alert can immensely save the oil spill from spreading and can be arrested quickly thus saving both money and ecosystem.

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

395

2 Literature Review In “Oil recovery with novel skimmer surfaces under cold climate conditions” authored by Keller and Clark [1], the skimming rate of skimmers at cold climate is tested. Various tests were conducted on the skimming capacity with emulsions and cold conditions. It was concluded that the cold climatic conditions reduce the skimming rate only to a very small extent and adjusting the inclination angle would be able to suffice those declining rate. In “Aspects of modeling an electric boat propulsion system” authored by Freire et al. [2], the design of a propulsion system with the help of electric motors for largescale boats is mentioned. The paper mentions the hardware in line methodology of getting the output from a controller and driving the motors with the help of speed regulator and providing a stable system. In “Design of a load cell with large overload capacity” authored by Aghili [3], the application of load cell in large scale of operations is explained. The paper proposes a methodology to use a special deign structure that ensures the rigidity of the structure without compromising in senor’s sensitivity. By the development of this methodology, a sensitivity of 0.2% is achieved. In “Implementation of GPS for location tracking” authored by Ariffin et al. [4], the implementation of a portable GPS module for live tracking systems is discussed. The paper discusses on how to implement the GPS live tracker for various portable applications and to get the results with at most accuracy. In “Design and efficiency comparison of various belt type oil skimmers” authored by Patel [5], the efficiency of various types of belt-type oil skimmers is discussed. It was concluded that adding more shafts and belts increases the efficiency of the skimmers greatly and even the small design changes, and perfection of design can make a significant impact in the efficiency of belt-type oil skimmers. In “Optimized Design of Electric Propulsion System for Small Crafts using the Differential Evolution Algorithm” authored by Lee et al. [6], the design of electric propulsion system is using an algorithm to determine the optimal electric motor and battery specifications for the basic requirements of a small boat implemented with an EPS. In “Development of a handy oil-skimmer” authored by Tatsuguchi et al. [7], the development of a suitable portable oil skimmers for vulnerable spots is discussed. The paper discusses the performance of portable skimmers when compared to traditional ones and also the usefulness of these skimmers at places not suitable for traditional skimmers. In “Analysis of belt type skimmer” authored by Thombare Babasaheb et al. [8], the size and type of material to be used in belt-type oil skimmer for highly efficient process are discussed. In “Review on analysis of oil skimmer” authored by Pund et al. [9], the design of efficient belt-type oil skimmer is discussed. It is concluded that the best design is capable of making the process highly efficient.

396

S. Rajesh Kannan et al.

In “A Review of Mobile Oil Skimmer” authored by Sathiyamoorthy et al. [10], the oil spill response by utilization of belt-type skimmer which is proved to be faster and highly efficient in oil spill recovery is discussed. In “Research on shipborne aided navigation system based on enhanced traffic environment” authored by Fan et al. [11], the bot on its movement through oceans is exposed to meteorological factors, climatic conditions which can be forecasted by consistent live tracking using GPS for allowing the bot to operate in safer ocean conditions which has been illustrated. In “Real-time Location Tracker for Critical Health Patient using Arduino, GPS Neo6m and GSM Sim800L in Health Care” authored by Kanani and Padole [12], the location and identification of patients through GPS location are identified, which can be modified to be used as in the same case for identification of bot GPS location. In “Microcontroller Based Controlling of Electric Vehicle” authored by Channi [13], controlling of DC motors using a microcontroller through a cloud-based platform that could be accessed using android is emphasized. The DC motors are controlled by L298N driver module, and it is concluded that it was efficient enough to operate the vehicle.

3 Smart Oil Skimmer With the increase of oil spill incidents in the world, it becomes an indomitable task to clean up the oil spill as soon as possible and prevent the further damage. For this purpose, we have used a detection system that detects the leakage of oil spill in oceans and a device that can be deployed immediately as soon as possible to clean up the oil spill effectively which can be controlled from a remote location preventing the physical contact with the heavy oil. The flow diagram of our concept is illustrated in Fig. 1.

3.1 Detection System The primary detection system makes use of a load cell connected with controller and a HX711 amplifier to amplify the values from load cell. A load cell is used to measure the weight of an object placed on its. The load cell works similar to a sensor. When load is applied to the load column, it undergoes compression and there is a change in its length. This load cell is used as a transducer as it converts one form of energy into other form, and it converts the force applied on it into change in length. This change in length cannot be known as it is. A strain gauge is attached to load cell which acts as a sensing element. The electrical resistance of the strain gauge is dependent on the length and crosssectional area. Therefore, when length increases and its area decreases, the resistance will increase. This is established by Ohm’s law for resistivity, R:

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

397

Fig. 1 Flow diagram of smart oil skimmer

R = ρl/ A

(1)

where ρ is the resistivity of load cell element (Ω m), l is the length (m), A is the crosssectional area (m2 ) and R be the electrical resistance of spring gauge [14]. A strain gauge is also a passive transducer which converts a mechanical deformation into a change of resistance. When strain gauge gets compressed, its length changes, which depends on the magnitude of force applied on top of the load cell. The resistance of the connected strain gauge undergoes a change when the length of cell varies. They are designed to make the two surfaces of the cell body bend depending on applied load. Strain placed on convex surface will stretch, while placed on concave surface will contract. The change in resistance is measured in terms of change in the output voltage, and it is very small so it needs to be amplified using a module that performs amplification. Thus, the load cell provides a voltage level equivalent to the weight placed on the load cell. The change in resistance is achieved by the Wheatstone bridge technique. Wheatstone bridge is implemented by arranging four strain gauges. The change-in-electrical resistance and impedance are measured by Wheatstone bridge with very small variations. The detected differential voltage in the center of the bridge is measured. The circuit converts variation in resistance into voltage change, and this eliminates offset voltage that appears in the voltage divider. The signal obtained from the strain gauge requires amplification. Amplification of the voltage signal from the Wheatstone bridge is necessary because the strain gauge’s Wheatstone bridge has only 2 mV/V of sensitivity. It gives a typical 3 V of cell excitation, and voltage available

398

S. Rajesh Kannan et al.

at full scale is 6 mV which is very low and hence this requires amplification. This is achieved by an instrumentation amplifier with characteristics like high differential gain, high input impedance, low offset, low input currents and low drift. Let us see the need of load cell for knowing the volume of oil leaked. We know that 1 cm2 = 1 × 10−4 m2 1 kgm2 = 0.1 Pa

(2)

Let the weight of tank filled with oil that was found using load cell be W = 5 kg (for example) and density of crude oil given as D = 820 kg/m3 that has specific gravity of 0.82. Let the assumed mass of oil tank to be nearly the same as weight, M = 5 kg. Assuming the oil-carrying container to be a perfect cylindrical carrier of radius, R = 1 m. Therefore, Area of tank be A = π × R × R = 3.14 m2

(3)

Now we have pressure exerted by liquid of mass M be P = mass of liquid/area of tank base = 1.59 kg/m2

(4)

As we already know that 1 kg/m2 = 0.1 Pa, then for pressure exerted by liquid it will be 0.159 Pa. We know that 1 Bar = 100, 000 Pa. So, in terms of bar for pressure exerted, P = 0.0159 m bar, and this is actually the hydrostatic pressure exerted by 5 kg oil. Now we have the formula to calculate the height [14] H = P/SG × D × g = 1.59/0.82 × 820 × 9.8 = 0.24 × 10−3

(5)

So, height is obtained; hence, volume can be easily calculated V = π × R × R × H = 3.14 × 1 × 0.00024 = 0.00075 m3

(6)

Now due to some unavoidable circumstances, the hull is broken and hence the carrying tank is damaged due to which oil is spilled into the ocean, so immediate action will be taken to fix the hull. By that time, say 2 kg of oil is lost (load cell will show 3 kg as it weighs only the entire container). Now we have the same parameters of SG, D and mass M , of 2 kg. Hence, the hydrostatic pressure exerted by M , be

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

P , = M , /A = 2/3.14 = 0.636 kg/m2 .

399

(7)

Height this oil can reach be H , = P , /SG × D × g = 0.636/6589.52 = 0.0000965 m.

(8)

Volume of oil after leak be V , = π × R × R × H , = 3.14 × 1 × 1 × 0.0000965 = 0.00030 m3

(9)

So, this is the volume of oil left; therefore, the oil leaked into the waters will be V − V , = 0.00075 − 0.00030 = 0.00045 m3 .

(10)

Hence, the volume of oil leaked is found to be 0.00045 m3 . This calculation can be coded into any programming language using any software available [15]. The generated code is then used to know the volume spontaneously at any instant of time. This data is useful for oil spill response and recovery crew to estimate the amount of oil lost which had to be recovered back immediately. The need to find weight involves tedious calculation and looks to contemplate us about the indirect approach in finding the volume of oil lost. Whenever the oil containers are filled with oil some space which is left unoccupied, say like 25–30% is left vacant which is indicated in Fig. 2. This is because the vessel might have numerous vigorous movements and halts throughout the course of its journey. When the ship suddenly stops, the oil inside the carrier remains to be in motion state as a result of inertia of motion. This in turn pushes or rises the oil in forward direction. Likewise, when ship sets into sudden motion after in rest, this pushes oil backward. This happens then and there in its journey, and if no space is left inside the container, the oil might leak out or overflow. Crude oil filled into containers might also exert some gases. This causes pressure inside the container; if the tank is full, the tank may not bear high pressure, and hence, oil might leak out. Partially filling the oil carriers helps to reduce this effect caused by pressure. The load cell has one end that needs to be fixed on a base, while the other end is let free. This setup is very important only then the load cell can undergo deformation. HX711 module is an amplifier breakout works to read the measured weight data from the load cell. This is a 24-bit precision A/D converter chip [3]. It is designed Fig. 2 Illustration of oil tank

Oil filled till 70%

400

S. Rajesh Kannan et al.

Fig. 3 Load cell connected with HX711 module

for high precision scale designs in electronics. The internal integration provided by this module is 128 times the integration available in gain amplifier. The input circuit in this case is pressure sensor load cell; this acts as the front end for the module. This setup along with load cell is interfaced to Arduino Uno microcontroller to know the weight. Communication is availed in HX711 by two wire interface called the clock and data which is shown in Fig. 3. This module provides higher integration and faster response when compared to any other modules. The features provided by HX711 are as follows: • • • • • • • •

The module provides two differential input channels that are selectable. It gives on-chip power supply regulator for ADC power supply and load cell. On-chip power-on-reset is available. Provides easy digital control and simple serial interface. Choosable output data rate is available, either 10SPS or 80SPS. Parallel 50 and 60 Hz supply rejection is also provided. It provides operational supply voltage at a range of 2.6–5.5 V. The range of operational temperature is − 40 to + 85 °C.

With the help of this system, we are able to measure any leakage of oil into the ocean and alert the crew so as to act accordingly.

3.2 Bot Propulsion System The movements of a system in different directions are controlled using a particular module called L298n motor driver module. This module is mainly used to communicate between the hardware motor component and the controller board. When the program is uploaded using IDE for controlling motors using a single L298n motor driver module and the GUI app. When uploading is completed, we can use the GU interface in our phones to control the motors movements and also speed of the motor. The motors are connected to the propeller for the movement in water. Our project involves the use of L298n module and GPS module, where we get the location of

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

401

the oil spill using the GPS module and then these coordinates are processed through the controller. After getting the coordinates, the bot moves to the particular location using the L298n motor driver module which communicates with motors to move in the specified direction [2]. As we can see, the driver module is working on the principle of H-Bridge control for directional movement and PWM for speed control. There are different types of L298 driver modules, but the L298n module has more stability compared to the driver modules in terms of speed and durability. The L298 driver module is a high-voltage, high-current dual full-bridge driver designed to accept standard TTL logic levels and drive inductive loads such as solenoids, DC and stepping motors. The L298N driver module communicates with the Arduino Uno microcontroller. Just by connecting the motors directly to the microcontroller will not work as it could even damage the microcontroller because Arduino pins are restricted to 50 mA current, but this is very less than 200 mA which is the requirement of a small motor. Hence, when the microcontroller is connected to the L298N driver module runs the motors as its output current is 2A. This is due to the function of driver module to convert low-current signal into high-current signal [13]. In Fig. 4, two enable inputs are provided to enable or disable the device independently of the input signals as shown. Instead of using the diesel engines for propulsion, the electric propulsion greatly reduces the pollution while maintaining the same degree of performance [16]. Electric propulsion system is needed to be designed for every vehicle for its transportation. Similarly, we have developed battery management system to power up the propulsion motors. It offers us the feature to recharge the bot whenever if required, and also it has been designed in such a way that battery lasting capacity is higher and drain is minimalized. Once charged, the propulsion system with other associated systems can withstand for 8 h. Algorithms were followed in our battery management system to attain the best power up ability to charge our bot [6]. This type of driver module is used only when: 1. High-power motor driver is required. 2. It is needed to operate different loads like motors and solenoid, etc., where an H-Bridge is required. Fig. 4 Circuit diagram of propulsion system

402

S. Rajesh Kannan et al.

The L298n module is quite different from the other modules since it can communicate faster than L293d module which has a quite a low communicating and processing speed. L293d drivers operate between 4.5 and 36 V, whereas L298n can operate at up to 46 V, L293d output current for each channel is 650 mA and it is 2 A for L298. Hence, the heat sink is provided in L298 motor drivers. The batter draining capacity of both modules is almost same, but L298n driver module has a heat sink which dissipates heat can save battery life and the voltage regulator regulates voltage in case of low supply of voltage. The double H-Bridge can handle up to: This L298N motor driver module uses two techniques for the control speed and rotation direction of the DC motors. These are pulse width modulation (PWM)—for controlling the speed. Halving the needed supply voltage can generally reduce the speed of motors due to proportionality. However, practically it is not possible to change the supply voltage as we want. The speed control can be achieved by PWM signal which is a high-frequency signal that controls the switching at a faster rate. The switching changes the average value of voltage required by the DC motors. Duty cycle of the PWM decides the value of average voltage. Duty cycle is the ratio of duration of signal HIGH to the total time period of PWM’s duty cycle, and H-Bridge is shown in Fig. 5—for controlling rotation direction. This module can control two DC motors or one stepper motor at the same time. L298n motor driver module uses the H-Bridge principle to control the direction of rotation of a DC motor as it contains two enable inputs to initiate or non-initiate further devices like motors attached to the output. In this technique, H-Bridge controls DC motor rotating direction by changing the polarity of its given input (input voltage). The H-Bridge is controlled by logic level signals. Since our designed bot needs all possible direction movements for its propulsion which can be achieved by standard TTL logic signals. Each possible logical combinations is responsible for varying operations of the DC motors used. On applying different polarities of voltages on two sides of the motors, these actions control switching to obtain HIGH or LOW logic signals to operate the motors in our desired modes. Transistors (BJT or MOSFET), with the motor at the center forming an H-like configuration. Input pins actually control the switches of the H-Bridge circuit inside L298N module. Two inputs pair up to represent the logic signals HIGH and LOW or HIGH and HIGH or LOW and LOW. Each representation is responsible for operations like brake, clockwise operation and anti-clockwise direction. We change the direction of the current flow by activating two particular switches at the same time, this way we can change the rotation direction of the motor. A small representation of working is given below. The L298N driver module can control up to four DC motors without speed control or two DC motors with speed control as illustrated in Fig. 6.

3.3 Live Tracking System The GPS module is connected with a controller that in turn is connected with the command center with the help of IoT and cloud technologies. An app is developed

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

403

Fig. 5 Implementation of propulsion system

Fig. 6 Representation of H-Bridge

to control and track the device. For prototyping purpose, we have used NEO 6 M UBLOX module and Blynk IoT platform to integrate the operations of the device, and the control is simple and effective with the use of GUI interfaces [2]. With the help of the tracking system, the live location of the device can be known and can be propelled to the desired location easily. Without the actual virtual representation of bot location, the monitoring purpose is failed and could lose track of it. Just by fetching only the geographical coordinates of the current bot location would not be an accurate information. So the GPS module is interfaced with cloud platform via Blynk that projects the bot location very accurately with all coordinates and directionrelated information [12]. Our idea is to focus on the movement of bot in the middle of the ocean or it can be unaccessible places to reach out for man power. There are several factors that influence the bot movement, and they include hydrological factors and meterological factors. With the help of the aforementioned GPS module which is interfaced with cloud technology like the Blynk platform, such influential factors are forecasted to redirect and manually operate the bot to safer conditions.

404

S. Rajesh Kannan et al.

This data obtained from the abovementioned factors can be studied so that the bot can be further improved and reengineered to operate in such difficult conditions [17]. This retrospective study is possible because of our live tracking system [11].

3.4 Skimming System After the device is driven to the spilt location on the ocean, then the skimming mechanism is initiated. Belt-type skimming mechanism is employed in our device. The rate of skimming depends on various factors such as the oil viscosity, density, temperature, belt material, the alignment and the skimming angle. We know that oil is a non-polar substance and water is a polar substance and also oil being less dense than water and that is the reason why oil floats on the surface of water. Hence, a non-polar substance type of belt is required to attract the oil. Oleophilic type of belt is used so that it offers a maximum cohesive property and a maximum skimming rate. Oil spill recovery is attained at its best efficiency when perfect type of skimmer is chosen. Belt-type skimmer is highly recommended and capable enough of skimming the oil at a much faster rate than any other types of skimmers. In our model, the belt type is equipped with pulleys that are operated by DC motors. After many revolutions, oil gets stuck to the material that is coated on the belt, and it is also easy to replace [18]. It was concluded that polyurethane is the best material where less clamor activity is required. It was concluded that the oil recovery rate is inversely proportional to the viscosity and the temperature of the surface. It was also mentioned that the pH values of the water have an effect in the recovery rate. The oil recovery rate is proportional to the speed of the belt and the surface area of the belt [19]. Thus, considering all these proofs of concepts we have designed our skimmer out of polyurethane material of area 80 cm2 and the skimming motor of RPM of about 40. The belt is inclined at a 90 ◦ angle so as to facilitate proper contact with the surface. We have designed a BMS system that is capable of supplying the skimming motor and propelling motors for up to 8 h. The BMS is made as a rechargeable type so as to reuse the batteries again and again. The life of the device can be further increased with the addition of batteries. Our design is capable of skimming about 5–7 L of oil from water depending on the above-said conditions such as temperature, density and viscosity. Much emphasis was made to increase the efficiency of the skimmers in order to deliver a perfect solution. Even though these kinds of skimmers are available in the market, they lack credible performance and are mostly manually operated. With our device, we come to a remotely operated skimmers so that humans are not required to risk their health by manually going to the hazardous locations and carrying out the task of removing the oil.

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

405

4 Results and Discussion With the inclusion of new technologies like IoT and cloud, we have come up with a unique solution that comes in handy at the times where the risk of oil spill has significantly increased and at times where people are increasingly concerned about their health. The experimental results after prototyping process have revealed that our prototype consumed very less amount of power but was able to deliver great results. With consuming just around 28.8 mAh, it was able to skim off about 5–7 L of oil from still water at normal conditions. Our battery management system was sufficient to power up the bot for up to 6 h. Our bot was able to deliver a peak RPM of about 6400, marking a max speed of about 27.358 knots and the GPS was accurate to about 2.5 m which can be further increased with the use of marine GPS. We have used polyurethane as the skimming belt material as it is proved that it is the most efficient material but we can also employ other materials depending on the conditions. Figure 9 illustrates oil spill that is being detected with the help of load cell and the abovementioned calculation which are programmed, and in case of any leakage of oil, it immediately indicates the position so that we can start to act on it. Figure 7 shows the smart oil skimmer prototype that we developed for the purpose of testing our idea, and in Table 1, the results are given. A polyurethane material is used as the skimming belt in our prototype, and Li ion batteries were used to power the entire propulsion and skimming process. It was observed that the prototype was able to recover 5–7 L of heavy oil from water every hour and capable of traveling at 27.358 knots at calm waters. The main idea of this device to carry out the task of skimming without physical involvement was a success with the prototype being able to be tracked with GPS as shown in Fig. 10 in 2.5 m of accuracy. The GPS data are obtained from the GPS module and processed and entered in a cloud-based platform so that it can be viewed from a remote place too. Figure 8 illustrates that the control of the prototype was smooth with the help of a custom-made GUI interface which enables even a semi-skilled person to control the whole process. The GUI has two different parts: one to track the location and other one to control the device. Figure 8 shows that the device can be easily controlled with the help of sliders and movable circles. Many such devices can be synchronized and can be controlled by just one person. Thus, we have made it possible to detect the oil spill as early as possible and thereby deploy the devices that can be propelled and tracked from a remote location and execute the skimming action.

5 Conclusion The oil skimmers find number of applications not only restricted to that of oil spill in oceans. The oil skimmers can be employed at various industries including thermal plants, desalination plants, offshore and onshore oil platform. The major application

406

S. Rajesh Kannan et al.

Fig. 7 Prototype developed

Table 1 Results of prototype model

Parameters

Inferred value

Battery consumed

28.8 mAh

Hours lasted

6h

Max RPM delivered by skimming motors

6400 RPM

Max operational of driver module

46 V

Max speed attained in knot

27.358

GPS accuracy

2.5 m

Belt-type skimmer material

Polyurethane material

Oil skimming capacity

5–7 L

apart from oil spill in oceans is in sewage treatment and desalination plants, instead of purifying the water at one end, simultaneous treatment of water along the flow of water bodies can be made so as to maintain the same quality of water everywhere. Based on the effluent that needs to be separated, different kinds of material belts can be used while using the same mechanism. The same mechanism can be employed to make the skimming process easy and with no physical contact, thereby preventing harm to humans involved. With the application of IoT, cloud and GPS, the traditional oil skimmers can be truly synchronized with Industry 4.0 tech.

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer Fig. 8 GUI for the control of prototype

Fig. 9 Implementation of load cell concept

407

408

S. Rajesh Kannan et al.

Fig. 10 GPS live tracking implementation

References 1. Keller AA, Clark K (2008) Oil recovery with novel skimmer surfaces under cold climate conditions. Int Oil Spill Conf Proc 2008(1):667–671 2. Freire T, Sousa DM, Branco PJC (2010) Aspects of modeling an electric boat propulsion system. In: 2010 IEEE region 8 international conference on computational technologies in electrical and electronics engineering (SIBIRCON), pp 812–817 3. Aghili F (2010) Design of a load cell with large overload capacity. Trans Can Soc Mech Eng 34:449–461 4. Ariffin AB, Aziz NHA, Othman KA (2011) Implementation of GPS for location tracking. In: 2011 IEEE control and system graduate research colloquium, pp 77–81 5. Patel M (2013) Design and efficiency comparison of various belt type oil skimmers. IJSR 4(1):998–1002 6. Lee DK, Jeong Y-K, Shin JG, Oh D-K (2014) Optimized design of electric propulsion system for small crafts using the differential evolution algorithm. Int J Precision Eng Manuf Green Technol 1(3):229–240 7. Tatsuguchi M, Mizutani M, Sano M, Fudo M, Ishida H, Fujita I (2014) Development of a handy oil-skimmer. IEEE Techno-Ocean’04 3:1464–1469 8. Dawood MA, Algawi RJ (2017) Study of operating conditions for oil skimmer apparatus from water. In: 2017 international conference on environmental impacts of the oil and gas industries: Kurdistan region of Iraq as a case study (EIOGI), pp 65–70 9. Thombare Babasaheb B, Barse Babasaheb N, Barhat Ganesh B, Kolhe Sani M, Jagtap Harshal B (2018) Analysis of belt type oil skimmer. IJARIIE 4(10):2917–2922 10. Pund R, Mhaske R, Rahane S, Rajput S (2018) Review on analysis of oil skimmer. IRJET 05(10):680–681 11. Fan Y, Huang L, Jiang D, Xu X (2018) Research on shipborne aided navigation system based on enhanced traffic environment perception. ONE 13(10):e0206402 12. Kanani P, Padole M (2020) Real-time location tracker for critical health patient using Arduino, GPS Neo6m and GSM Sim800L in health care. In: 2020 4th international conference on intelligent computing and control systems (ICICCS)

Design and Implementation of Efficient IoT-Based Smart Oil Skimmer

409

13. Channi HK (2021) Microcontroller based controlling of electric vehicle. LAP Germany Editor: LAPISBN: 978-620-3-47186-1 14. Müller Ivan A, de Brito RM, Pereira CE, Brusamarello VJ (2010) Load cells in force sensing analysis—theory and a novel application. Instrum Meas Mag IEEE 15–19 15. John W, Eren H (1999) The measurement instrumentation and sensors handbook, 1st edn. Springer 16. Son Y, Lee S, Sul S (2018) DC power system for fishing boat. In: 2018 IEEE international conference on power electronics, drives and energy systems (PEDES), pp 1–6 17. Lee S, Tewolde G, Kwon J (2014) Design and implementation of vehicle tracking system using GPS/GSM/GPRS technology and smartphone application. In: 2014 IEEE world forum on Internet of Things (WF-IoT), pp 353–358 18. Sathiyamoorthy V, Arumugam K, Arun Pragathish M, Barath BN, Baskar M, Balamurugan S (2018) A review of mobile oil skimmer. IJET 7(3):58–60 19. Khan MA, Mostafa SMG, Rahman A, Hoque MM, Islam MT, Kalam M (2021) Implementing and improving the IoT based weather monitoring and controlling double discs type oil skimmer. In: 2021 2nd international conference on robotics, electrical and signal processing techniques (ICREST), pp 499–503

Comparison of Discrete Time Sliding Manifold and Its Impact on System Dynamics Shaktikumar R. Shiledar and Gajanan M. Malwatkar

Abstract This paper presented the comparative study on development in the discrete time sliding mode control (DTSMC) especially concern with the third-order process to analysis of the system state variables. In the literature study, it was found that many researchers were fascinated by the study of DTSMC strategies implemented for control engineering applications. In this paper, different DTSMC approaches such as second-order DTSMC, DTSMC for the higher order system, and DTSMC with discrete PID sliding manifold have been included to find its effect on system dynamics such as interval state variables of the systems. The presented controller design depends on discrete controller techniques which are obtained from the laws available in literature in recent years. In this paper, the presented techniques give minimum value for the control input to reduce the effort for a controller, and it is implemented on the proposed model of the system. In the formulation of DTSMC law, the calculated values of gains from the constant matrix or gain tuning parameters are used to derive one of the laws of DTSMC. In the presented work, stable poles of the system have been given the desired value for the controller input signal. The major advantages of the presented techniques included dynamics of the state variables and effect of variables arises difficulty in implementation. The obtained simulation results show the comparative study for DTSMC the approach in the form of stabilization of the system states and its output responses. Keywords Performance · Comparison · Discrete time sling mode · Simulation · State variables · Controller action

S. R. Shiledar (B) · G. M. Malwatkar Department of Instrumentation Engineering, Government College of Engineering, Jalgaon 425002, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_27

411

412

S. R. Shiledar and G. M. Malwatkar

1 Introduction In recent days, all advance and optimized control strategies are based on digital devise. In actual, implementation of the controller in discrete time domain becomes to use the microprocessor or personal computer devices for computing the algorithms. It is well known that control system design shifted from continuous to digital becomes unstable after the sampling process. In the literature study, it was found that first time work presented in [1] studied the discrete variable structure controller theory and originated this from modification of the equivalent continues SMC reaching condition applicable to solution for reaching condition in discrete [1, 2]. However, in discrete sliding mode control (DSMC) structure case, all states do not stay on the sliding manifold instead of it make a zigzag motion around the manifold with a certain band becomes given the quasi sliding mode form. The chattering is arising mainly due to the finite sampling time, so dynamics of the system perform zigzag motion in the neighborhood of the sliding surface is unavoidable. A second-order DTSMC law or a non-switching type reaching law is used for application to communication networks [3]. Gao et al. is investigating the quasi-sliding mode design controller technique and achieving the rule of thumb for a single input single output system [4]. There were two types of state trajectories explained with sufficient conditions to fulfill some attributes to design sliding mode and reaching law [2]. The chattering effect occurs in the CSMC, when the sampling rate is high which has been minimized by selection of proper quasi-sliding mode band around the switching manifold. Misawa also inside into design the DSMC for a single input linear system [5]. In the work of [5], Lyapunov stability theory is applied to design the DSMC and Lyapunov function consider as V (k) = 21 s 2 (x). The DSMC presentation shows the trade-off between the sliding band, system modeling errors, and control of gain parameters. The performance of the Misawa [5] DSMC is compared with Furuta’s et al.’s and Chan et al.’s results [5, 6]. On the other hand, the discrete time slide mode control strategies introduced by Bartoszewicz et al. redefine the concept of slide mode show that the strategy requires less control effort to ensure better performance and system stability than regulation of Gao law [7]. In [7], designed algorithm no need to cross system states successively in each steps or iteration so that ultimately the effect of chattering was reduced as compared to results given in [2]. In light of the paper flow is in Sect. 2, included introduction of DTSMC, in Sect. 3, describe the problem formulation, Sect. 4 discussed the different DTSMC techniques, in Sect. 5, contains simulation and discussion of outcomes. Section 6 included the conclusion and comments of the newly developed prevalent controllers.

2 Preliminaries The discrete time control system analysis can be defined into two ways: first is the system derived from continues time output at the discrete instant, i.e., sampled data

Comparison of Discrete Time Sliding Manifold and Its Impact …

413

system. Second way is the system which are in the discrete form and where system states are introduced only at discrete time instant. Let us consider the representation of a CT system converted to a DT system with a defined sample time (T ) and quantization rate for the given continues signal to becomes discrete signal. x˙ = Ac x(t) ˆ + Bc u(t) ˆ + Dc u(t) y = Cc x(t)

(1)

where x(t) ˆ = [x1 , x2 , . . .] is the dynamics, matrix Ac , Bc , Cc , and Dc is the system constant matrices with proper dimensions for the system input, control input, and system output, respectively. The pair (Ac , Bc ) must be controllable in this case the solution for Eq. (1) given as follows: The solution for Eq. (1) given as follows: t x(t) = ϕ(t − tinitial )x(t ˆ initial ) +

ϕ(t − τ ) ⎡u(τ )dτ

tinitial

In discrete signal formation process consider the input signal constant during the two sampling interval to form the discrete signal. The solution of Eq. (1) exists in of tinitial ≤ τ ≤ (k + 1)T where assume the initial interval tinitial = kT the term case t tinitial ϕ(t − τ ) ⎡u(τ )dτ in Eq. (1) denoted by θ (t − kT )u(kT ) x((k + 1)T ) = ϕ(T )x(kT ) + θ (T )u(kT )

(2)

from Eq. (2) can be written for T = 1 and put θ (T ) = ⎡ x(k ˆ + 1) = ϕx(k) ˆ + ⎡u(k) y(k) = H x(k) ˆ + Du(k)

(3)

The most of the researchers are worked on continuous SMC, and the discrete counterpart of it are attempted for some of the applications. The VSC theory implementation on the computer-based system with a relatively low sampling period lead to discrete SMC. The availability of significant algorithm for implementation of DTSMC on various types of systems are well studies in literature [1, 4, 6].

3 Problem Formulation In the problem statement, consider the nth order form of the discrete time process represented in form of transfer function

414

S. R. Shiledar and G. M. Malwatkar

  (b0 ± δb0 )z P + (b1 ± δb1 )z p−1 + · · · + b p ± δb p ld z , p≤n z n + (a1 ± δa1 )z n−1 + · · · + (an ± δan )

(4)

where the term δ representing the parametric uncertainty in system. The term ld represents the delayed samples and it written as d = lTd , where T represents the sampling period. Equation (4) shows in state space discrete form as x(k + 1) = (ϕ + δϕ)x(k) ˆ + ( ⎡ + δ ⎡)u(k) + f (k) y(k) = H x(k) ˆ

(5)

where the input state variable, input control signal, and system output are represented by x (k) = Rn , u(k) = R and y(k) = R1×n , respectively. ϕ ∈ Rn×n , ⎡ ∈ Rn×1 and H ∈ R1×n are constant system matrices. The system parametric uncertainties are represented as δϕ and δ ⎡, respectively. The nonlinear external disturbances are represented by (k). In the problem formulation consider the stability and observability condition satisfied the system with (ϕ + δϕ) and (ϕ, H ). Equation (4) can be represented in form of single input and single output, linear discrete time without considering parametric uncertainties and disturbances in system as Δ

Δ

Δ

x [(k + 1)T ] = ϕx (kT ) + ⎡u(kT )

(6)

The dynamic response of the system error is calculated using the difference between reference or input state with the desired value of the signal. Here, suppose function error e(k), reference signal r(k), and system output y(k) represent as e(k) = r (k) − y(k) e(k) = x(k) ˆ − xd (k) Δ

Δ

(7)

Δ

where x (k) is the system states, and x d (k) is the system desired states

4 DTSMC Techniques The following DTSMC laws have been taken for comparison and simulation studies.

4.1 Second Order Discrete DTSMC The work given in [8] design of second-order DTSMC with observer and equivalent control. Assume the system equation fulfill the condition of matrix relation. (A, H) is controllable and observable

Comparison of Discrete Time Sliding Manifold and Its Impact …

415

x(k + 1) = ϕx(k) ˆ + ⎡u(k) + ⊓(k) y(k) = H x(k) ˆ

(8)

where ⊓(k) = δϕx(k)+δ ⎡ ˆ u(k)+d(k). On the basis of observer theory formulation of first-order discrete time sliding mode observer (DSMO) formulated as   x(k + 1) = ϕx(k) + ⎡u(k) + L y(k) − yˆ (k)   yˆ (k) + Msat K Δ

Δ

(9)

Δ

Δ

y (k) = H x (k) where M ∈ Rn , y (k) = y(k) − H x (k) is chosen as the sliding manifold, L ∈ Rn×1 is the observer gain matrix, and K denotes the constant. The formulation of the total controller is written in the following form: u(k) = u eq (k) + u sw (k)

(10)

where u eq denoted as the equivalent control law and usw as the switching control. The combined law can be represented as  −1

−βs(k) − c, ϕx(k) u eq (k) = c, ⎡ ˆ + c, (xd (k + 1))   s(k) − Msat σ where s(k) are the sliding functions and xd desired state vector trajectory. The choice for selection of M, β, σ and c, are given in [8].

4.2 DTSMC for Higher Order Systems In the contribution by Khandekar et al. [9], a robust controller design for the higher order time delay system was demonstrated. A sliding mode surface designed using the function of the state variables and generated error from the reference signal. The gain matrix or tuning parameter for the controller selected using the pole placement strategy. The stability of the proposed sliding mode was validated through the pole placement method [10]. The controller has satisfactory implemented on oscillatory behaviors type of system with minimization of the chattering of the controller. The sliding surface designed in the work includes of [9] Δ

s(k) = cx (k) − K t e(k − 1)

(11)

where c = [c1 , c2 , . . . cn ] and K t are gain matrix and it used as the tuning parameter for the designed controller. The guideline for selection of tuning parameter and

416

S. R. Shiledar and G. M. Malwatkar

constant are given precisely for system with higher order. The representation of sliding surface condition as s(k)[s(k + 1) − s(k)] < 0.

(12)

The equivalent control law presented in [9] can be a modified form of basic DTSMC. u eq (k) = (c ⎡)−1 [−cϕx(k) − cd(k) − K t e(k)]

(13)

The property of robustness to sustain to parameter variation and perturbations achieved by the introducing of a discontinues with the equivalent control law u sw (k) = αsgn(s(k))

(14)

The signum function used to switching for states on designed hyper plane. So, total control law written as u(k) = u eq (k) + u sw (k)

(15)

In discrete sliding mode controller, it is found that the proper tracking of the reference signal with improved results were obtained compared to continues sliding mode algorithm were achieved in [9, 11, 12].

4.3 DTSMC with Modified Law In work of [13], Ghabi and Dhouibi designed of the disturbance rejection estimation algorithm presented to overcome the effect of uncertainties without modification of the control law. This control strategy provided stability of close loop dynamics of the system by reducing chattering. The design of the slide mode control ensures to fast access and stability of the system (k + 1) provided as s(k + 1) = c, x(k) ˆ + c, x(k)u(k) ˆ + ζ (k) − c, xˆd (k + 1) + λs(k)

(16)

According to the reaching law in Eq. (16), the control law for the discrete uncertain system given by the expression u(k) = (c, x(k)) ˆ − 1[−c ⎡ x(k) ˆ + c, xˆd (k + 1) + s(k) − λs(k) − βsign(s(k))]

(17)

where λ is the positive constant, and ζ (k) ∈ R is equivalent disturbance included with parametric uncertainty.

Comparison of Discrete Time Sliding Manifold and Its Impact …

417

4.4 DTSMC Using Discrete PID Sliding Surface In work given by Shiledar et al. [14] designed the sliding manifold using discrete PID law with the estimation of the tuning parameter gain constants. The derived reaching law shown the guaranty of the stability also, it reduces the reaching time of states variable on sliding manifold. The sliding surface at instant s(k + 1) represented as   ˆ + 1) + k p e(k ˆ + 2) − e(k ˆ + 1) s(k + 1) = cϕx(k) ˆ + c ⎡ u(k) − K t e(k  

+ ki e(k ˆ + 2) − 2e(k ˆ + 1) + e(k) ˆ (18) ˆ + 1) + kd e(k From Eq. (18), the design of control law as   ˆ + K t (e(k ˆ + 2) − e(k ˆ + 1) u eq (k) = (K ⎡)−1 −K ϕx(k) ˆ + 1) + k p e(k  

+ ki e(k ˆ + 2) − 2e(k ˆ + 1) + e(k) ˆ (19) ˆ + 1) + kd e(k Δ

c is the sliding gain parameter, e(k) is the error function, kt is the gain constant, and k p , ki , kd , are the tuning parameter constants.

5 Results and Discussion Consider the transfer function of higher order time delay system given in [15]. G(s) =

s3

+

11s 2

1 e−0.5s + 35s + 25

The discrete form of model obtained by pole-zero method with chosen sampling interval Ts = 0.1s and representation of continues transfer function equation in z-domain as G(z) =

0.0001473z 2 + 0.0002947z + 0.0001473 −5 z z 3 − 2.1182z 2 + 1.466z − 0.3329

The state space matrices of z-transform function represented in controllable canonical form ⎡ ⎡ ⎤ ⎡ ⎤ ⎤ 0 1 0 0 0 ϕ=⎣ 0 0 1 ⎦, ⎡ = ⎣ 0 ⎦, H T = ⎣ 0.002412 ⎦ 0.3329 −1.466 2.118 0.003481 1 and D = [0] The change in the state x1 , x2 and x3 and control signal given by Mihoub et al. [8], Khandekar et al. [9], Ghabi and Dhouibi [13] and Shiledar et al. [14] are shown

418

S. R. Shiledar and G. M. Malwatkar

in Fig. 1a–c, respectively. From Fig. 1a–c, it is clear that DTSMC using discrete PID sliding surface law [14] initially provides higher rate of change of states, and after some instant it states becomes stable with low gain, while the DTSMC for higher order systems law [9] has initially low rate of change of states as compared to [14], but states are stable with high gain. The laws presented by Mihoub et al. [8] state variables are started from the initial point and all states signals stable with smooth response of gain value one as shown in Fig. 1. The DTSMC reaching algorithm suggested by Ghabi and Dhouibi [13] provides initial phase shows the similar and comparable values of the state while the after some time instant developed law of Ghabi and Dhouibi [13] state variables gain shifted from gain value less than one as shown in Fig. 1. In Fig. 1d and Fig. 2, the simulation results involved the responses of input control signals and system outputs with setting of set point value (SP) is 1 and allowable band of error consider as 2% of set point value. The sliding gain parameters for Mihoub et al. [8] are as c matrix chosen as c = [0.3005, −0.5027, 0.4979], the value of sampling step constant is chosen as Ts = 0.01, and switching gain constant M = 0.9. The constant terms for [9] included the gain matrix c = [0.3005, −0.5027, 0.4979] which is same as c matrix defined in [8] and switching constant α = 0.1 is chosen. In the algorithm of [13], the parameters are chosen as, λ = 0.025 and β = 0.014. For the work of [14], the sliding parameters include k p , ki and kd as 0.3, 0.01 and 1.65, respectively, with gain constant kt = 11.15.

6 Conclusions The development of advanced control algorithms, especially DTSMC makes attention due to its merits and feasibility of implementation for linear, nonlinear, higher order, and delay time systems. The design of the DTSMC is the basic requirement to develop a sliding surface or reaching law and formulation of the relation between sliding surface and the control input signal are presented. In recent years, DTSMC strategy is implemented to achieve the goal for high performance tracking with minimization of the high frequency oscillations. So, suppression of chattering becomes a help to reduce the wear and tear action at the final control element in the process loop. The prevalent DTSMC laws were studied, which minimize the bandwidth of the quasi-sliding mode domain and suppresses the effect of chattering. The comparative responses of the state variables, control input, and system output response were studied and simulated for different recent techniques reported in the literature. In this paper, comparative study of the second-order DTSMC, higher order systems DTSMC, DTSMC with modified law, and DTSMC using discrete PID sliding surface law presented in terms of systems dynamics state variables, control input, and system output presented with simulation results. It is remarked that there should be a low rate of change of states so that the control action shall be smooth. From the simulated result, it is observed that up to some extent all responses presented in the paper are allowable for smooth control action for higher order time delay system.

Comparison of Discrete Time Sliding Manifold and Its Impact …

Fig. 1 System states x1 , x2 , x3 and control input signal u

419

420

S. R. Shiledar and G. M. Malwatkar

Fig. 2 System output y(k)

References 1. Dote Y, Hoft RG (1980) Microprocessor based sliding mode controller for dc motor drives. In: Presented at the industrial applications society annual meeting, Cincinnati, OH 2. Gao W, Wang Y, Homaifa A (1995) Discrete-time variable structure control systems. IEEE Trans Ind Electron 42(2):117–122. https://doi.org/10.1109/41.370376 3. Veselic B, Perunicic-Drazenovic B, Milosavljevic C (2010) Improved discrete-time slidingmode position control using Euler velocity estimation. IEEE Trans Ind Electron 57(11):3840– 3847. https://doi.org/10.1109/TIE.2010.2042416 4. Gao W, Hung JC (1993) Variable structure control of nonlinear systems: a new approach. IEEE Trans Ind Electron 40(1):45–55. https://doi.org/10.1109/41.184820 5. Misawa EA (1997) Discrete-time sliding mode control: the linear case. ASME J Dyn Sys Meas Control 119(4):819–821. https://doi.org/10.1115/1.2802396 6. Furuta K (1990) Sliding mode control of a discrete system. Syst Control Lett 14(2):145–152. ISSN 0167-6911. https://doi.org/10.1016/0167-6911(90)90030-X 7. Bartoszewicz A, Zuk J (2010) Sliding mode control—basic concepts and current trends. In: 2010 IEEE international symposium on industrial electronics, pp 3772–3777. https://doi.org/ 10.1109/ISIE.2010.5637990 8. Mihoub M, Ahmed Said N, Abdennour B (2011) A second order discrete sliding mode observer for the variable structure control of a semi batch reactor. Control Eng Practice 19(10):1216– 1222 9. Khandekar AA, Malwatkar GM, Patre BM (2013) Discrete sliding mode control for robust tracking of higher order delay time systems with experimental application. ISA Trans 52(1):36– 44. ISSN 0019-0578 10. Ogata K (2003) Discrete time control systems, 2nd edn. Prentic Hall, New Jersey 11. Eker I (2010) Second-order sliding mode control with experimental application. ISA Trans 49(3):394–405. ISSN 0019-0578 12. Malwatkar GM, Khandekar AA, Asutkar VG, Waghmare LM (2008) Design of centralized PI/PID controller: interaction measure approach. In: 2008 IEEE region 10 and the third international conference on industrial and information systems

Comparison of Discrete Time Sliding Manifold and Its Impact …

421

13. Ghabi J, Dhouibi H (2018) Discrete time sliding mode controller using a disturbance compensator for nonlinear uncertain systems. Int J Control Autom Syst 16:1156–1164. https://doi.org/ 10.1007/s12555-017-0185-0 14. Shiledar SR, Malwatkar GM, Jadhav IS, Lakhekar GV (2021) Design of discrete sliding mode controller for higher order system. Reliab Theory Appl SI 1(60) 15. Wang QG, Lee TH, Fung HW, Bi Q, Zhang Y (1999) PID tuning for improved performance. IEEE transactions on control systems technology, vol 7, pp 457–465

Volkswagen Emission: An Analysis on the VW Vento Using Automotive Network Data Suprava Sarkar and Nithin Mohan

Abstract The advancement in automotive technology has surged the intricacy, cost, and size of the governing elements within its system. The demand for more and more attributes has led automation to reach its pinnacles of modernity. The automation within the network produces tons of data every second for efficient ECU (Electronic Control Unit) response. Data extraction and research play a significant role during the vehicle development processes for better evaluation of the vehicle under trial. Additionally, the central government of India has ordered that automotive manufacturers only build, market, and register BS-VI (BS6) automobiles. The rise in automobile pollution has a severe impact on global warming. Harmful chemical compounds like NOx , CO, and HC have caused severe health issues among the general public. Hence, it is essential to regulate these exhaust compounds until a permanent solution is found. This research paper seeks to analyze CAN Bus data of the Volkswagen Vento, TSI Highline Plus, petrol engine and determine its emission trail when subjected to city/highway conditions. Various ECU parameters of the vehicle have been studied, compared, and scrutinized to extract the desired data that can provide the emission details of the vehicle under trial. Keywords Volkswagen emission · Volkswagen Vento analysis · Lambda/oxygen sensor · Air/fuel ratio · ECU parameters · CAN bus network · OBD II · Telematics and fleet management

S. Sarkar Department of Instrumentation and Control, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka 576104, India e-mail: [email protected] N. Mohan (B) Influx Big Data Solutions, Influx Technology India, #2, Krishvi, Ground Floor, 2nd stage Indiranagar, BDA Colony, Domlur, Bengaluru, Karnataka 560071, India e-mail: [email protected] URL: http://www.influxbigdata.in © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_28

423

424

S. Sarkar and N. Mohan

1 Introduction The automotive industry’s major factors are the growing integration of technologically advanced solutions in the automotive sector to cater to the ever-increasing need to optimize vehicle and passenger safety, reduce emissions, and the widespread electrification of vehicles. The growing demand for autonomous vehicles has increased the need to use advanced electronic control units in modern cars, ensuring that contemporary automobiles achieve maximum efficiency of fossil fuel-powered vehicles and electric vehicles, particularly in developed regions. The installation of a small-sized onboard data logger helps to store and transmit vehicle data and vehicle routes using the Global Positioning System (GPS). The recorded data is used for diagnostics of a vehicle or the entire fleet and for understanding the driver usage patterns, which helps OEMs produce better cars. Additionally, the government’s regulations have become more stringent since the automotive sector was a significant contributor toward Global warming; the strict norms have created an immense demand for automotive testing and analysis.

2 Background Theory 2.1 CAN Bus and OBD II Overview CAN Bus is a vehicle communication network protocol that facilitates communication between all the electronic control units within the vehicle. Other communication protocols such as SPI and I2C communicate point to point, master and slave configuration. Whereas CAN Bus is a multi-master setup, any ECU connected to the network can take over the bus. The CAN bus broadcasts the message to all the ECUs present in the network, and only the interested ECU pulls in the required message. CAN allows the flexibility to install high-tech features within the car. For example, the parking assists system allows multiple sensors and ECUs, such as the parking sensors, transmission control unit, side door mirror motor, etc., to effectively communicate via the CAN Bus to park your vehicle safely. More examples include the auto brake wiping technology used in high-end cars, wiping off moisture from brake rotors during a rainy day. In this case, the rain sensor communicates via the CAN network with the breaking ECU to perform this particular function. These are the rudimentary purpose of CAN, allowing effective and efficient communication between all the ECUs within the vehicle. The On-Board Diagnostics II or OBD 2 is the self-diagnostic and reporting capability of a vehicle. The OBD port gives access to the vehicle’s subsystem data for analysis and error diagnostics. The Electronic Control Units (ECU) or nodes in a car are the electronic subsystem units responsible for specific tasks within the vehicle. For example, to detect if the car is stopped on an incline, the “hill hold” capability uses input from the vehicle’s tilt sensor (also used by the burglar alarm), and the road

Volkswagen Emission: An Analysis on the VW Vento Using …

425

speed sensors (also used by the ABS, engine control, and traction control) through the CAN bus. Similarly, the CAN bus receives data from seat belt sensors (part of the airbag controls) to verify whether seat belts are fastened, etc., to effectively communicate via the CAN Bus so you can park your vehicle safely. More examples include the anti-lock braking systems, which work with speed sensors in the vehicle to mitigate skidding on the road effectively. The OBD II port can retrieve the transmitted ECU message via a data logger. The OBD 2 is a critical aspect of the Telematics and Fleet management system that is used to measure vehicle health and condition. The fleet management system has multiple advanced, high-tech features, such as the vehicle can track down wear and tear of vehicle parts. More examples include measuring driver behavior patterns, idling time, etc.

2.2 Telematics and Data Logging A telematics system consists of a vehicle data logger that allows for the transmission and storage of telemetry data. It uses a built-in 4G modem with SIM Card to communicate over a wireless network. The data logger captures vehicle-specific data by interfacing with the onboard diagnostics (ODBII) system via CAN-BUS connector and sends the data to a centralized server along with real-time location information collected using a GPS module. The server interprets the data and makes it available to end-users via secure websites and apps designed for smartphones and tablets. Few recordable parameters include location, speed, idling time, hard acceleration or braking, fuel consumption, vehicle issues, and other telemetry data sets. When evaluated for specific events and patterns, this data can provide in-depth insights across an entire fleet (Fig. 1).

2.3 ECU Parameter IDs The On-Board Diagnostics II parameter IDs are a set of codes used to request data from the CAN Bus network of the vehicle by utilizing the Mode $01 (request live data) of the OBD II service. An automotive ECU communicates via the CAN Bus network to external test equipment to request data using the parameter IDs. An OBD II live data parameter usually consists of a unique hexadecimal identifier, the number of bits returned, start bit, size, the minimum and maximum range of physical data, and the unit (Fig. 2). Vehicle Parameters Analyzed Vehicle Speed Sensor (VSS)—The transmission output or wheel speed is measured by the Vehicle Speed Sensor (VSS). The ECM uses this data to change vehicle engine operations like ignition timing, air/fuel ratio, transmission shift points, and

426

S. Sarkar and N. Mohan

Fig. 1 The telematics and fleet management connection, configuration, and data logging steps

Fig. 2 The number of parameter IDs that were tracked during this test

start diagnostic programs. The output of the Vehicle Speed Sensor is in terms of Voltage and is communicated directly via the CAN Bus signals. Vehicle Gear Sensor—The vehicle’s transmission uses a conjunction of gears to keep the engine’s rotational speed within a safe range. The gear position sensor detects the gear position, such as first or second, and communicates the information to the ECU through the vehicle network. Engine RPM Sensor—The rotational speed or position of the crankshaft is monitored by an Engine RPM Sensor, which is an electrical device used in an internal combustion engine. The ECM uses this information to manage fuel injection, ignition timing, and other engine parameters.

Volkswagen Emission: An Analysis on the VW Vento Using …

427

Fig. 3 The relationship between a car’s air/fuel mixture and its oxygen/lambda sensor. The X-axis represents the air/fuel ratio, where 1 being a balance between the ratios. The Y-axis represents the corresponding equivalent voltage output of the oxygen sensor

Oxygen Sensor Voltage—The oxygen sensor is placed at the exhaust manifold and is used to determine the amount of unburned oxygen present in the exhaust. The fuel mixture is determined by monitoring the oxygen levels in the exhaust. The output voltage, which ranges from 0.1 to 1 V, indicates whether the fuel mixture is burning rich (with less oxygen) or lean (with more oxygen).

2.4 Relationship Between Oxygen Sensor and Air/Fuel Mixture The difference in oxygen levels across the sensing element generates a voltage across the sensor’s platinum electrodes when the air/fuel mixture is rich. There is minimal O2 in the exhaust: typically, 0.8–0.9 V. The sensor’s voltage output reduces to 0.1–0.3 V when the air/fuel mixture is lean and more oxygen is in the exhaust. The sensor’s output voltage is roughly 0.45 V when the air/fuel mixture is balanced (Fig. 3).

2.5 Relationship Between the Fuel/Air Equivalence Ratio and Relative Emissions It is seen that NOx emission gradually increases as the mixture moves toward the balanced point. The NOx production peaks when the mixture is balanced and drops down when the mixture gets leaner or richer. The NOx emission varies from approx. 1900 ppm and peaks to approx. 3100 ppm at the balanced point. Additionally, Carbon Monoxide (CO) is at its peak (approx. 40,000 ppm) when the mixture is rich and

428

S. Sarkar and N. Mohan

Fig. 4 The relationship between a car’s fuel/air mixture and its relative emissions. The X-axis represents the fuel/air equivalence ratio, where 1 is a balance point. The Y-axis represents the typical HC and NOx concentration in PPM

gradually decreases near approx. 3000 ppm as it gets closer to the balanced point (Fig. 4). Further, we see that the CO production increases as the mixture gets leaner, but the rate of change in increment is minor. Furthermore, we see that the Hydrocarbon production is at its peak (approx. 28,000 ppm), initially, when the air/fuel mixture is rich and drops down as it lets leaner. We also see that the production of HC gradually increases before the mixture balanced point. Additionally, at a 0.9 air to fuel ratio value, all the three emissions produced are relatively at their minimum. Hence at 0.9 V, the vehicle is at its efficient point, and a 0.9 V value should be maintained to emit the least emission relatively.

3 Results and Analysis 3.1 Test Route The test was conducted for 3 h 30 min with a total distance of 193 km. The vehicle was subjected to both city and highway conditions for maximum clarity in results. The data acquisition system was configured to log OBD II data and GPS and interfaced with the vehicle CAN network. The data logger acquired 140,000 data point samples from each ECU sensor under observation. Electronic Control Unit (ECU) parameters such as lambda sensor (O2 sensor) voltage, vehicle speed, RPM, vehicle power, gear, etc., were observed throughout the test (Fig. 5).

Volkswagen Emission: An Analysis on the VW Vento Using …

429

Fig. 5 The testing location and the path taken by the car

3.2 Oxygen Sensor and Vehicle Gear Relationship The X-axis represents time in seconds, and Y-axis shows the gear change. The color indicates the variation in oxygen sensor voltage with the gear change. The orange section of the graph represents low oxygen sensor voltage: < 0.45 V (lean air/fuel mixture). The blue area shows the oxygen sensor voltage greater than 0.45 V (rich air/fuel mixture), where 0.45 is the balanced point with the air/fuel mixture being 1 (Fig. 6). It is seen that the average oxygen sensor voltage is low (the color orange, < 0.45 V) throughout the journey. The low sensor voltage indicates the air/fuel mixture is lean,

Fig. 6 The change in gear with time throughout the journey

430

S. Sarkar and N. Mohan

and more oxygen is in the exhaust. During this phase, the vehicle NOx emission drops, and there is a gradual decrease in carbon monoxide and hydrocarbon emissions from their peak values. The blue phase (> 0.45 V) indicates the air/fuel mixture is rich, and less oxygen is available in the exhaust. During this phase, the vehicle NOx emission drops drastically. CO and HC emissions are high initially, as shown in Fig. 4, and with the increase in mixture richness, the CO and HC emission decreases too. The graph’s colorless sections indicate a balanced air/fuel mixture with oxygen sensor voltage being around 0.45 V. It is seen that the oxygen sensor voltage tends to increase when there is a sudden rise in gear when the vehicle starts accelerating, indicating a rich air/fuel mixture (as the color of the graph turns blue) and reduction in NOx emissions after exceeding the balanced point (according to the Fig. 4). During this phase, the HC emissions increase up to 900 ppm and CO production goes near 100 ppm. It is also seen that during vehicle idling conditions, the oxygen sensor voltage increases and produces more HC (Fig. 4). Additionally, on average, the third gear of the car, where the vehicle speed is around 40 km/hr, shows higher HC production (as we get to see more color blue throughout the graph). In contrast, the vehicle keeps a lower air/fuel mixture throughout the journey, excluding gear three, as the color orange dominates the entire graph. Majority of the time, the vehicle produces more elevated CO and HC emissions, and low NOx emissions. The car rarely achieves an oxygen sensor voltage around 0.45 V, where the air/fuel mixture is balanced and produces high NOx emissions (Fig. 7). It is seen that the oxygen sensor voltage has majorly concentrated in two places— between 0.1 to 0.2 V and 0.7 to 0.8 V. The 0.1–0.2 V range indicates the vehicle had a lean air/fuel mixture, and the 0.7–0.8 V range indicates rich air to fuel mixture. The

Fig. 7 The causal relationship between the oxygen sensor voltage and the change in gear. The X-axis represents the change in oxygen sensor voltage, which varies from 0 to 1 V. The Y-axis shows the gear change from 0th gear to 7th gear

Volkswagen Emission: An Analysis on the VW Vento Using …

431

car achieved a balanced point of 0.45 V rarely. The data shows that the vehicle had low NOx emissions, but the CO and HC emissions were high throughout the journey. Furthermore, the data shows higher data point concentration around the 0.1–0.2 V range, indicating low NOx emissions (1000 ppm) and moderate HC emissions (700 ppm). Gear 0–1 shows lower data point concentrations of a rich mixture, indicating low emissions of CO.

3.3 Vehicle Speed and Oxygen Sensor Relationship The X-axis shows the change in time (s), and the Y-axis shows the change in vehicle speed (km/h). The color of the graph indicates the change in oxygen sensor voltage with vehicle speed. Orange sections of the graph show low oxygen sensor voltage (< 0.45 V), where 0.45 V being the median, and the blue section shows high oxygen sensor voltage (> 0.45 V) (Fig. 8). It is seen that the average oxygen sensor voltage is low (the color orange, < 0.45 V) throughout the journey. The low sensor voltage indicates the air/fuel mixture is lean, and more oxygen is in the exhaust. During this phase, the vehicle NOx emission drops, and there is a gradual decrease in carbon monoxide and hydrocarbon emissions from their peak values. The blue phase (> 0.45 V) indicates the air/fuel mixture is rich, and less oxygen is available in the exhaust. During this phase, the vehicle NOx emission drops drastically. CO and HC emissions are high initially, as shown in Fig. 4, and with the increase in mixture richness, the CO and HC emission decreases too. The graph’s colorless sections indicate a balanced air/fuel mixture with oxygen sensor voltage being around 0.45 V (Fig. 9).

Fig. 8 The relationship between the change in vehicle speed and oxygen sensor voltage with respect to time

432

S. Sarkar and N. Mohan

Fig. 9 The causal relationship between the change in vehicle speed and oxygen sensor voltage with respect to time. The X-axis represents the change is vehicle speed and the Y-axis shows the change in oxygen sensor voltage from 0 to 1 V

Additionally, it is seen that, with a sudden rise in acceleration/vehicle speed, the air/fuel mixture gets rich (as the color of the graph shifts to blue), indicating higher HC production (Fig. 4). According to the data, we see that above 100 km/h, the probability of HC emission is high, and at a range of 20 km/h–60 km/h, the vehicle purely emits CO. The vehicle rarely reaches its balance point, where NOx production peaks (Fig. 4). The oxygen sensor voltage appears to have concentrated in two areas: between 0.1 and 0.2 V and 0.7 and 0.8 V. The range of 0.1–0.2 V suggests a lean air/fuel combination, whereas the range of 0.7–0.8 V indicates a rich air/fuel mixture. Rarely did the vehicle reach a balanced point of 0.45 V. The car had low NOx emissions, but significant CO and HC emissions throughout the drive, according to the statistics. Additionally, the data indicates greater data point concentrations in the 0.1–0.2 V range, indicating low NOx (1000 ppm) and moderate HC emissions (700 ppm). From 0 to 30 km/h, vehicle speed results in minimal CO emissions but a greater concentration of HC.

3.4 Engine RPM and Oxygen Sensor Relationship The X-axis represents time in seconds, and the Y-axis shows the engine RPM variation. The color indicates the variation in oxygen sensor voltage with the change in engine RPM. The orange section of the graph represents low oxygen sensor voltage: < 0.45 V (lean air/fuel mixture). The blue area shows the oxygen sensor voltage

Volkswagen Emission: An Analysis on the VW Vento Using …

433

Fig. 10 The change in RPM with time throughout the journey

greater than 0.45 V (rich air/fuel mixture), where 0.45 is the balanced point with the air/fuel mixture being 1 (Fig. 10). It is seen that the average oxygen sensor voltage is low (the color orange, < 0.45 V) throughout the journey. The low sensor voltage indicates the air/fuel mixture is lean, and more oxygen is in the exhaust (as the color of the graph is orange). During this phase, the vehicle NOx emission drops (Fig. 4), and there is a gradual decrease in carbon monoxide and hydrocarbon emissions from their peak values. The blue phase (> 0.45 V) indicates the air/fuel mixture is rich, and less oxygen is available in the exhaust. During this phase, the vehicle NOx emission drops drastically. CO and HC emissions are high initially, as shown in Fig. 4, and with the increase in mixture richness, the CO and HC emission decreases too. The colorless sections of the graph indicate a balanced air/fuel mixture with oxygen sensor voltage being around 0.45 V. Additionally, it is seen that above 2500 RPM (vehicle speed 40 km/h), the vehicle tends to produce more HC from the exhaust, as the graph gets bluer above this data point. The sudden increase in acceleration tends to make higher HC at the exhaust as well. On average, RPM below 2500 tends to produce more CO and HC and reduces NOx emissions. The oxygen sensor voltage being 0.45 V, where the air/fuel mixture is balanced, is rarely seen throughout the journey.

4 Conclusion 4.1 Vehicle Idling The data analysis shows us that during the vehicle’s idling, the system provides a high oxygen sensor voltage, indicating a rich air/fuel mixture at the exhaust. During

434

S. Sarkar and N. Mohan

this phase, the car produces Hydrocarbons, whereas the NOx and Carbon monoxide emissions reduce. The hydrocarbon production increases to approx. 900 ppm in this phase, as shown in Fig. 4. In contrast, the carbon monoxide emission tends to drop toward 100 ppm and NOx toward 1000 ppm. To reduce any emissions during the idling period, the driver should switch off his engine for efficiency. Switching off the engine during an idling state not only reduces emissions but also saves fuel and money.

4.2 Acceleration The vehicle’s oxygen sensor voltage shoots above 0.45 V during a sudden rise in acceleration, indicating a rich air/fuel mixture at the exhaust. During this phase, the car produces Hydrocarbons, whereas the NOx and Carbon monoxide emissions reduce. The hydrocarbon production increases to 900 ppm in this phase, as shown in Fig. 4. In contrast, the carbon monoxide emission tends to drop toward 100 ppm and NOx toward 1000 ppm. On average, the driving pattern has not surged the acceleration multiple times and has maintained an oxygen sensor voltage below 0.45 V, indicating a lean air/fuel mixture at the exhaust. The Carbon monoxide and HC production peaks during this phase, reaching up to 40,000 ppm and 1400 ppm, respectively. The NOx production is low, around 1900 ppm. The vehicle rarely achieves a stable air/fuel ratio throughout the journey. Hence, the driver should avoid sudden acceleration of the vehicle, as suddenly accelerating the vehicle causes higher fuel consumption and increases HC emissions at the exhaust.

4.3 Vehicle Speed It is seen that vehicle speed above 42 km/h (2500RPM) produces a rich air/fuel mixture at the exhaust. During this phase, the car produces Hydrocarbons, whereas the NOx and Carbon monoxide emissions reduce. The hydrocarbon production increases to 900 ppm in this phase, as shown in Fig. 4. In contrast, the carbon monoxide emission tends to drop toward 100 ppm and NOx toward 1000 ppm. We see that RPM speed above 2500 directly relates to the air/fuel mixture as per the data shown above. At speeds less than 40 km/h, the vehicle produces low oxygen sensor voltage, indicating a low air/fuel mixture at the exhaust. The Carbon monoxide and HC production peaks during this phase, reaching up to 40,000 ppm and 1400 ppm, respectively. The NOx production is low, around 1900 ppm. The vehicle rarely achieves a stable air/fuel ratio throughout the journey. Hence, the driver should maintain speeds below 42 km/h at gear values between 0–2 for better fuel efficiency, as speeds above 42 km tend to cause a rich air to fuel mixture.

Volkswagen Emission: An Analysis on the VW Vento Using …

435

4.4 Emissions The vehicle fluctuates from a lean air–fuel mixture and a rich air–fuel mixture throughout the journey, indicating high Carbon Monoxide and Hydrocarbon emissions. The vehicle rarely maintains a balanced point of a lambda value around 1, indicating low NOx emissions (fluctuating around 3000 ppm) frequency at the exhaust. The use of transition metal oxides as described in [6] or customized sensors such as [7] alongwith accelerometers as described in [8] could facilitate continuous monitoring of oxides of nitrogen and sulfur and correlate with acceleration enabling better understanding and possibility of regulation. Finally, it can be concluded that Volkswagen Vento, TSI Highline Plus, petrol engine had low NOx emissions throughout the journey. The vehicle emits higher Hydrocarbon emissions than Carbon Monoxide. Furthermore, the efficiency of the vehicle depends on the catalytic converter that determines the concentration of the emission particles released into the atmosphere.

References 1. “ECU” is a Three letter answer for all the innovative features in your car: know how the story unfolded. Embitel (2017) 2. Etschberger K, Controller area network, basics protocols, chips and applications. ISBN 3-00007376-0 (www.ixxat.com) 3. Lawrenz W, CAN systems engineering, from theory to practical applications. ISBN 0-38794939-9 4. SAE International (2018) J1939: serial control and communications heavy duty vehicle network 5. Understanding CAN DBC, telematics & fleet management. Influx Big Data Solutions 6. Nag P, Sadani K (2015) Factors governing gas sensing characteristics of some transition metal oxides. Int J Control Theory Appl 8(3):1225–1233 7. Sadani K, Nag P (2016) Encapsulated piezoresistor cantilevers as affinity sensors: a review. Int J Control Theory Appl 9(39):205–215 8. Sadani K, Prabhakar DA, Nag P (2018) A cardio pulmonary resuscitation device for stretchers. Int J Eng Technol 7(2.21):62–5

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective Application as Smart Wheelchairs Leon Muli Suryavanshi, Ananth Jnana Chandraraj, Kshetrimayum Lochan, and Pooja Nag

Abstract The wheelchair is used to assist in movement and heavily relied on by people who have significant difficulty in walking. Manual wheelchairs can be used by people with good upper limb strength and dexterity, however people suffering from any kind of locomotive disabilities in the arms would require assistance. This article reports the proof of concept of a smart electric wheelchair that can be controlled by utilizing hand gestures. A prototype with wheeled mobile robot was developed and involved the use of data fusion from sensors mounted on human hand to create a scheme for motor control of the wheels. Hand gesture dependent data from accelerometers and gyroscope were inputs to control the wheelchair in the required direction. The prototype of the wheelchair, PID based controller and a hand mounted sensor unit (HMSU) was able to function reliably and efficiently on urban terrain. Keywords Electric wheelchair (ECW) · Gesture-controlled/control · IMU (inertial measurement unit) sensor · PID · Filter · Hand mounted sensor unit (HMSU)

1 Introduction A large number of people around the world suffer from locomotive disabilities [1]. This may be permanent in some cases or a temporary injury. Hence, wheelchairs have become extremely important for such individuals to move around. Patients who cannot operate manual wheelchairs effectively would always require additional help. This limits the freedom of the patients, as they have to constantly ask for assistance to move around. Conventional electric wheelchairs (ECWs), are operated with a joystick input. This is suitable for the majority of people that need ECWs, however Leon Muli Suryavanshi, Ananth Jnana Chandraraj—Authors have contributed equally to the work L. M. Suryavanshi · A. J. Chandraraj · K. Lochan · P. Nag (B) Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal, Karnataka 576104, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_29

437

438

L. M. Suryavanshi et al.

a small percentage may have additional disabilities that make it difficult to operate a joystick. One such disability is Symbrachydactyly [2], in which an individual has missing/underdeveloped/webbed fingers and/or toes. The purpose of this work was therefore aimed to assist such individuals that have finger-related disabilities in addition to locomotive disabilities, by providing an alternative control scheme to a joystick, at a minimum cost. This could be used to easily and intuitively maneuver the ECW. Multiple iterations of alternative wheelchair control system have been conceptualized and realized with a portion of them relying on IMU sensors in different capacities. This project demonstrates the prospects of utilizing IMU sensors as the main element to not only derive control signals, but also provide stabilizing data based on the wheelchair’s angular orientation, to be used in a PID control algorithm. Furthermore, the use of two IMU sensors enables proper functioning of the control system on sloped surfaces.

1.1 Related Work Over the years, several types of smart electric wheelchairs have been reported, which utilize different strategies to sense the direction command from the patient, and accordingly maneuver the wheelchair. This includes, mounting of sensors on the patient hand or head to sense certain gestures or movement [3, 4], muscle activity monitoring through electromyogram (EMG) [5], tracking of eye movement [6], facial expressions [7], a combination of few of the above [8] or voice recognition [9]. Khan [3], proposed a flex sensor-based control of wheelchair. The flex sensors were attached to a glove and finger movement controlled the motors. Four fingers were used to derive five different outputs corresponding to forward, left, right, reverse and stop. However, flex sensors do not produce consistent change in resistance over long periods of usage and easily get damaged. Machangpa and Chingthamb [4], designed a wheelchair that could be controlled by head movement. Gyroscope, accelerometer and ultrasonic sensors were mounted on a hat and used to sense the head movement. The data from the sensors were passed through a novel algorithm for processing. This enabled each user to have a custom threshold for their head movements. The ultrasonic sensor was used to prevent accidental collisions. The roll and pitch angles were taken into consideration in processing when determining the direction to move. Rakasena and Herdiman [5], designed a wheelchair controlled by muscle activity in the arm and utilized surface electromyography sensors (SEMGs) to provide signals based on the user’s muscle activity. This design was mainly aimed at people that have some sort of muscle dystrophy disabilities. An electrode was placed on the wrist flexor muscle and wrist extensor muscle. Direction of movement was controlled by performing different kinds of grip movement i.e. spherical, cylindrical, lateral, hook and tip grasp. These gave different output potential readings which could be mapped to different motions of the wheelchair. However, a limitation of such a method is that EMG sensors require the user’s skin to be cleaned of any oils produced by the body in order to get reliable readings from the sensor. Rajesh and Mantur [6],

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective …

439

designed an eye tracking based wheelchair control system with an attached camera mounted on the user’s head. The camera kept track of the position of the pupil and the wheelchair moved in the direction the user looked. In such a system, the field of vision is hindered and the user had to wait two seconds for any change in direction or stop. Rabhi et al. [7], proposed a wheel chair controlled using facial expressions. Yi et al. [8], proposed a combination or a hybrid of EMG signals and head gestures which worked as the input to the wheelchair. Rabhi et al. [10], designed an Electric Wheel Chair controlled by a touchscreen device such as a tablet. Although non-contact approaches such as these [6–8, 10–12] are frequently used for sensory feedback, they run a risk of faulty alarms or actuation as well. The advantage of the system was that it required less effort when compared to a traditional joystick powered wheel chair and also gave a much quicker speed response. Choudhari et al. [13], proposed a unique method to control wheelchairs—an electrooculography-based system to control the motion of the wheel chair (forward, left, right turn and stop) through number of eye blinks. While some methods are promising, most of them suffer from issues of user discomfort, reliability, incapability of traversing over slopes or a complicated approach which thereby increases cost. This work reports of a simple approach of using accelerometer and gyroscope sensor data fusion and an efficient control scheme to develop a robust prototype, which can be easily scaled up to build an efficient and affordable smart electric wheelchair.

2 Methodology 2.1 The Hand Mounted Sensor Unit (HMSU) The proposed system comprised of two 6-axis inertial measurement unit (IMU) MPU-6050 sensors [2], consisting of an accelerometer and gyroscope. These sensors were connected to Arduino Nanos [14] for processing. The first IMU was mounted on the user’s hand and detected the orientation of their hand (Fig. 1b), based on which the motor was actuated via a PWM signal. This acted as the hand mounted sensor unit (HMSU). A minimum threshold was set for the angles required to cause the motor to be actuated (Table 1). The HMSU communicated to the actuating Arduino wirelessly via nrfl01 + modules (Fig. 1a).

2.2 Motor Actuator A second Arduino was mounted on the wheelchair and connected to two motors via a motor driver (Fig. 2). Optical speed encoders that were attached to the motors and the second IMU sensor were both used to send feedback signals. This feedback was passed through a PID system which was used to maintain a constant speed of

440

L. M. Suryavanshi et al.

Fig. 1 a Hand mounted sensor unit circuit connection; b type of gestures used to control wheelchair (with the tip of fingers being the nose of the roll axis and base of the wrist being tail end of the roll axis)

Table 1 Mapping for motor actuation and control

Direction of motor actuation

Tilt angle of user’s hand Roll

Forward Left Right Back

Pitch

≤ 30°

≥ 30°

≥ −30°

≥ 30°

≤ −30°

≤ 30°

≤ −30°

≥ -30°

≥ 30°

≤ 30°

≥ 30°

≥ −30°

≤ 30°

≤ −30°

≥ −30°

≤ −30°

the wheelchair in order to enable it to traverse surfaces with varying gradients and maintain a straight path when required, since a differential drive system was used. The IMU sensor connected to the actuating Arduino had two purposes. One was to detect changes in yaw angle. A rotation toward the right was taken as positive change, while a rotation toward the left was taken as negative change. If the wheelchair was headed on a straight path, then the yaw angle would be approximately zero. This information was used to determine the orientation of the wheelchair and fed into the control loop to ensure stable orientation when in motion. The second purpose was to determine the roll and pitch angle of the wheelchair. This is because the orientation of the HMSU’s IMU sensor is affected by two entities,

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective …

441

Fig. 2 Actuating Arduino circuit connection

one being the user’s hand and the other being the wheelchair itself. If we take the example of a scenario where the wheelchair is traversing an inclined surface, say a 15° slope, then the hand mounted IMU sensor would not be able to distinguish between the cause of the change in its orientation and its reading would therefore be erroneous (it would have an offset of 15°). Hence, in order to correct any offset created by the incline of the terrain being traversed, the orientation of the wheelchair in terms of pitch and roll was always calculated by the actuating Arduino and sent periodically to the hand-mounted Arduino wirelessly. The change in yaw angle as well as the speed detected by the encoders is fed to the PID algorithm in a feedback loop. The feedback loop consisted of two PID algorithms, with one responding to the speed and the other responding to the change in yaw angle.

2.3 PID Control Loop Assuming a situation where the wheelchair is meant to move forward or back, in order to ensure that it moves in a straight path, the wheels must rotate at the same RPM. However, observing the speed is not enough, since each wheel might experience different levels of sudden disturbances, caused by either varying surface friction or obstacles. Hence the requirement for detecting the yaw angle. The PID control loop (Fig. 3) was programmed, and the PID gains were found mainly by trial and error. The method was implemented with the assumption that the motor driver would output consistent current that was proportional to the PWM signal being written to it.

442

L. M. Suryavanshi et al.

Fig. 3 PID control loop for actuating motors. Speed PID Sample Time (Ts1) = 0.25 s, Yaw PID Sample Time (Ts2) = Program execution time of Arduino microcontroller

The control scheme was applied in the prototype (Figs. 4 and 5) and the system response was tested.

Fig. 4 Prototype setup top view

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective …

443

Fig. 5 Prototype setup bottom view

2.4 Calculating Angles Form IMU Sensor The MPU-6050 contains a gyroscope and accelerometer which individually produce their own readings, a sensor fusion algorithm was used to get the highest accuracy and speed response possible with the hardware. The first step involved calculation of the roll and pitch angles from each sensor in the module. The sensitivity and subsequently the range of the accelerometer and gyroscope sensors are selected by writing to the control register. The highest sensitivity was used in this case, with the accelerometer having a max sensitivity of 16,384 LSB/g (least significant bits per g) and range of ± 2 g. While the gyroscope had a max sensitivity of 131 LSB/°/s (least significant bits per degree per second) and a range of ± 250°/s.

2.4.1

Euler Angle Calculations from Accelerometer Data

The accelerometer readings were used to calculate Euler angles [15], using the concept of rotation matrices. However, since the accelerometer cannot detect yaw angles, any orders of the rotation matrix which contain the yaw angle (RXZY , RYZX ,

444

L. M. Suryavanshi et al.

RZXY , RZYX ) have to be left out. This leaves only two orders out of the original six orders. These two orders are RXYZ and RYXZ . The accelerometer outputs the gravity vector in terms of g-force, therefore by working backward, it is possible to find the rotation done about a given base vector. This base vector was taken to be the gravity vector when the sensor was at rest and upright. Therefore, the output given can be assumed to be: ⎛

⎞ ⎛ ⎞ GV X 0 G V = ⎝ G V Y ⎠ = R⎝ 0 ⎠ GV Z 1

(1)

where GV GVX GVY GVZ R

The gravitational vector The X axis g-force The Y axis g-force The Z axis g-force Rotation matrix (RXYZ and RYXZ ).

The pitch ‘θ ’ can hence be deduced: ⎛

⎞ ⎛ ⎞ GV X − sin(θ ) G V = ⎝ G V Y ⎠ = ⎝ sin(ϕ) cos(θ ) ⎠ GV Z cos(ϕ) cos(θ ) ⎛ ⎞ −G VX ⎠ (θ ) X Y Z = tan−1 ⎝ / 2 G V Y + G 2V Z

(2)

(3)

Similarly, the roll ‘ϕ’ can be found: ⎛

⎞ ⎛ ⎞ GV X − sin(θ ) cos(ϕ) ⎠ GV = ⎝ GVY ⎠ = ⎝ sin(ϕ) GV Z cos(ϕ) cos(θ ) ⎞ ⎛ −G V Y ⎠ (ϕ)Y X Z = tan−1 ⎝ / 2 2 GV X + GV Z

2.4.2

(4)

(5)

Gyroscope Data and Sensor Fusion

The gyroscope produces the angular velocity experienced via the Coriolis effect about the three axes. Therefore, the roll, pitch and yaw can be found by discrete

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective …

445

Fig. 6 Complementary filter flow chart. The complementary filter (Fig. 6) is capable of producing roll and pitch values with a suitable amount of accuracy and speed

integration of these values: n Σ

(ωi × Δt)

(6)

i=1

The gyroscope is less prone to noise but suffers from gyroscope drift, which causes the angles to drift overtime from the true value, hence the use of a sensor fusion algorithm. The fusion algorithm involves passing the calculated gyroscope angles through a discrete high pass filter, while the Euler angles deduced form accelerometer are passed through a discrete low pass filter. This is because the gyroscope is not affected by any linear movement/jerks and only responds rotational motion and hence has barely any high frequency noise. The accelerometer on the other hand has a lot of high frequency noise since it responds to any kind of motion. The filtered values are then added to give the final roll and pitch. The resulting equation is a complementary filter as shown below in the case of pitch ‘θ ’: θ out[n] = β(θ out[n − 1] + ω[n] × Δt) + (1 − β)(θ accel[n])

(7)

3 Results and Discussion Figure 7a shows the change in pitch angle with time as the IMU sensor was rotated arbitrarily. The blue line depicts the output given by the complementary filter. As can be seen, the blue line hugs red line but does ignore most of the noise. However, it should be noted that that the complementary filter can only produce reliable readings in the range of ± 90°. This is because of gimbal lock which occurs when calculating Euler angles with a limited number of rotation orders. This means that the

446

L. M. Suryavanshi et al.

accelerometer sensor can comprehend angles on the range of ± 90°, beyond which the values wrap around back to 0°. Figure 7b shows the PWM signal being written by the PID controller based on the yaw angle given as input to the PID, with a setpoint of zero degrees. In the given instance shown here, the model was manually rotated toward the right by 40°. This caused the PID controller to slow down the left wheel by decreasing the PWM value (shown by Green line) and keep the right wheel moving faster (Red line) so as to turn the model back toward the left till the setpoint of zero degrees was reached. As the setpoint was being approached (shown by Blue line), the left wheel started to increase speed. The system exhibited slight overshoot in PWM signal being written to the motors, but the was able to maintain the system close to the setpoint. Once very near/at the setpoint, the left motor resumes to maintain the same speed as right motor. The data shown above (Fig. 7b) was taken from the prototype wheelchair in motion.

Fig. 7 a Filtered pitch angle vs unfiltered pitch reading against time (red graph: raw accelerometer reading, blue graph: filtered sensor fusion reading); b variation in PWM output and wheelchair yaw angle with time (red: PWM signal written to right motor, green: PWM written to left motor, blue: yaw angle of wheelchair)

Hand Gesture-Controlled Wheeled Mobile Robot for Prospective …

447

4 Conclusion This paper puts forward the concept of using IMU sensors to create a control scheme that uses hand gestures. The MPU-6050 is to be mounted on the user’s hand and produces input signals, an Arduino Nano generates control signals based on the input signal using sensor fusion and Euler angles, which are sent wirelessly to the wheelchair’s actuation system. A PID control loop actuates the wheelchair motors, considering the speed of the wheelchair (using encoders), the orientation of the wheelchair (yaw angle). The PID ensures that the speed of the wheelchair remains relatively constant and that over all operational stability is maintained. The MPU6050 attached to the wheelchair corrects offsets that arise in the HMSU due to inclined surfaces. The prototype wheelchair operates as expected and responds to the hand mounted sensor unit with minimum delay and minor errors. The current system was developed at minimum cost which would make it more easily available to the general public and sets a good foundation for the potential improvement, via implementation of additional sensors, better hardware and further consideration of ergonomics.

References 1. WHO Webpage. https://www.who.int/publications/i/item/9789241547482 2. Woodside JC, Light TR (2016) Symbrachydactyly—diagnosis, function, and treatment. J Hand Surg 41(1):135–143. https://doi.org/10.1016/j.jhsa.2015.06.114 3. Khan HA (2018) The economical design of a hand-gesture and Bluetooth controlled wheelchair by integrating indigenous components: mobility aid for the disabled. IOP conference series: materials science and engineering, vol 473, Foundation University Islamabad, South Korea 4. Machangpa JW, Chingthamb TS (2018) Head gesture controlled wheelchair for quadriplegic patients. In: Procedia computer science, international conference on computational intelligence and data science, vol 132, pp 342–351 5. Rakasena EPG, Herdiman L (2019) Electric wheelchair with forward-reverse control using electromyography (EMG) control of arm muscle. In: Journal of physics: conference series, international conference on applied science and technology (iCAST on Engineering Science) 24–25 Oct 2019, vol 1450, Bali, Indonesia 6. Rajesh A, Mantur M (2017) Eyeball gesture controlled automatic wheelchair using deep learning. In: 2017 IEEE region 10 humanitarian technology conference (R10-HTC), Dhaka, pp 387–391. https://doi.org/10.1109/R10-HTC.2017.8288981 7. Rabhi Y, Mrabet M, Fnaiech F (2018) A facial expression controlled wheelchair for people with disabilities. Comput Methods Programs Biomed 165:89–105 8. Yi Z, Xiaolin F, Yuan L (2015) Intelligent wheelchair system based on sEMG and head gesture. J China Univ Posts Telecommun 22(2):74–80, 95 9. Joseph C, Aswin S, Sanjeev Prasad J (2019) Voice and gesture controlled wheelchair. In: Proceedings of the third international conference on computing methodologies and communication (ICCMC 2019) 10. Rabhi Y, Mrabet M, Fnaiech F, Gorce P, Miri I, Dziri C (2018) Intelligent touchscreen joystick for controlling electric wheelchair. IRBM 39(3):180–193 11. Nag P, Mathew A, Sadani K (2017) A non-contact flow monitoring system for rotameters. ICAR

448

L. M. Suryavanshi et al.

12. Nag P, Shinde S, Sadani K (2018) Developing a smart navigator for surveillance in unmanned zones. Ind Interact Innov Sci Eng Technol, 255–263 13. Choudhari AM, Porwal P, Jonnalagedda V, Mériaudeau F (2019) An electrooculography based human machine interface for wheelchair control. Bio Cybern Biomed Eng 39(3):673–685 14. Arduino webpage, Arduino Nano Every. https://store-usa.arduino.cc/products/arduino-nanoevery-with-headers?queryID=undefined 15. Ramessur SA, Oree V (2019) Hand gesture controller for robotic-wheelchair using microelectromechanical sensor ADXL 345. Smart Sustain Eng Next Gener Appl 561:3–10

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test K. Aneesha Acharya and Amartya Choudhary

Abstract Manual dexterity is the ability to use the hands in a coordinated way to grasp and manipulate objects and to demonstrate slight precise movement. This research aims to develop an outcome measurement unit (pegboard) that tells about timing information of the hand activity and the accountability of the pegs movement. This paper seeks to create an automated pegboard with the help of photo-interrupter sensors, Infrared sensors, microcontroller, and adequate circuitry to reduce human error in time calculation. This pegboard test consists of two methods to assess hand– eye coordination: (1) Insertion of the peg and (2) Removal of the pegs. ATmega 2560 microcontroller helps in interfacing the hardware model with the software. Photointerrupter helps in sensing removal and insertion of the pegs inside the peg holes. An inbuilt timer calculates the activity period in milliseconds. Test sessions showed that the average time taken to complete the test rises as the age increases, indicating the loss of hand–eye coordination and increment in underlying neural-related conditions as the age increases. Keywords Dexterity · Pegboard · Photo-interrupter sensor · Infrared sensor · Microcontroller

1 Introduction Manual dexterity is the ability to make coordinated hand and finger movements to grasp and manipulate objects [1]. It includes a combination of muscular-skeletal and neurological functions to execute small precise actions. A well-developed manual dexterity over time means that a person can adequately assess, plan, and execute a task within a normal time range. The test used in this paper is the modified Nine-Hole Peg Test (NHPT), which is also considered a gold standard in manual dexterity tests [2]. NHPT is used widely in stroke, multiple sclerosis research. A manual pegboard K. Aneesha Acharya (B) · A. Choudhary Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal 576104, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_30

449

450

K. Aneesha Acharya and A. Choudhary

Fig. 1 Traditional Nine-Hole Peg Test (NHPT) set-up

test kit consists of a pegboard with a defined number of holes, pegs, and a stopwatch, as shown in Fig. 1. Rivera et al. [3] describe how the automatic pegboard replaces the traditional pegboard device using sensors and embedded electronics. Using a customized electronic pegboard, the authors studied the different intraindividual variability in fine motor skills of varying age groups. Children of various age groups performed the test, and the result emphasizes that the children with lower dysfunction show less intraindividual variability. Jobbágy et al. [4] introduced a smart 9-hole peg tester model, which comprises a standard 9-hole pegboard but with Light Emitted Diodes (LEDs) next to each hole. In this model, one can freely choose the order of the peg placement and removal. This smart 9-hole peg tester (s-9-HPT) analyzes a healthy person’s finger dexterity and a patient suffering from a stroke. Wang et al. [5] focus on manual dexterity as a critical aspect of motor function across the different age span. The test was conducted on 4319 subjects using a manual 9-hole pegboard. The dominant hand completed the test more quickly as compared to the non-dominant hand. The presence of impairments in dexterity across different age span is possible to study using standards of the 9-HPT. Research related to fine motor assessment in multiple sclerosis (MS) and Parkinson’s disease population also use the pegboard test [6, 7]. With the upcoming technologies and advancements in the automation, the shortcomings of the manual pegboard test can be rectified [8]. The automation of the pegboard test certainly will be cost-friendly and lead to a more accurate result [9]. This work aims to develop an outcome measurement unit (pegboard) that tells about timing information of the hand activity and the accountability of the pegs movement. This paper is structured as follows: Sect. 2 contains methodology, which explains experimental procedure and algorithm developed in pegboard test set-up development. Section 3, results are obtained with our method, followed by a discussion in Sect. 4, and the concluding remarks in Sect. 5.

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test

451

2 Methodology The conventional pegboard test had a few shortcomings, such as the clinician need to operate a stopwatch during the trial, keeping an eye on peg movement, and continuous observation required. Stopwatch handling is a tedious and time-consuming job. The main idea behind automated pegboard development is to eliminate all human errors which occur during the test process and make a more efficient assessment tool. The equipment used here are the photo-interrupters (MOC 7811), LED lights, microcontroller (Arduino Mega 2560), resistors (10 kΩ and 100 Ω), breadboards, single strand wires, jumper wire, and a laptop. Figure 2 shows the block diagram representation of the pegboard test set-up. A photo-interrupter detects the insertion and removal of pegs. Photo-interrupter makes the system more efficient and less error-prone. It detects the presence of any object once it has been inserted right between its opening; hence will help to detect the pegs inserted. The next part consists of selecting the LEDs and resistor as there is always a chance to blow the LEDs up if the resistor is not connected. LEDs guide the patient about the peg placement. It is a 9-hole peg test, and hence we require nine sensors and nine LEDs attached to each sensor. The next component in this project is the infrared (IR) sensor which counts the number of hand movements. Hand movement count ensures that one peg is lifted in each movement and the participant obeys all the rules related to peg displacement. LCDs or the screen itself used to display the result message, and lastly, to transfer the data to the portable devices to track the progress, an HC-05 Bluetooth module is required.

Fig. 2 Block diagram representation of the pegboard test set-up

452

K. Aneesha Acharya and A. Choudhary

2.1 Algorithm of Automated Pegboard Test The flowchart shown in Fig. 3 gives a basic idea of how the logic works in an automated pegboard test set-up. IR sensor connected to digital pin 6 of the Arduino Mega 2560 board. Nine photo-interrupter sensors have been initialized with a specific Personal Identification Number (PIN). Only one peg is supposed to be picked up during the peg test, and the IR sensor is positioned properly near the peg holder. When the participant’s hand approaches the peg holder to collect one peg, the IR sensor detects this movement. IR sensor triggers the timer, and once the peg has been inserted in the pegboard, it increments the counter. IR sensor keeps track of hand movement during the test. After hand holding the peg from the peg holder, the next step is to detect correct peg placement. The output of the photo-interrupter circuit is observed for the proper placement of peg placement. If the output given by this photo-interrupter sensor is high, the second LED lights up, indicating the hole in which the peg is supposed to be kept. In between this process, if the subject cannot keep a single peg into the indicated sensor within 250 s, a “time out” message will be displayed on the screen. This is particularly required when the clinician examines the patient population and monitors the progress in multiple sessions. In healthy adults, pegboard tests can complete within 50 s and for the patient population, time given is five times more (250 s). The “time out” feature helps in identifying the intentional delay from the patient population. In this automated pegboard test, guided LEDs have been used to indicate the position where the peg is to be inserted. In case the subject keeps the peg in a sensor that he was not supposed to keep, the subsequent LED will not light up. This indicates that previous peg movement was wrong and participant can correct it immediately. Hence, LED arrangements guide the participant during the peg insertion and removal process. Once all the nine pegs have been inserted, and the hand movement sensed by the IR sensor count is 9, a message to remove the peg is displayed on the screen. Once the message to remove has appeared, the subject is supposed to remove the pegs out from the sensors in reverse order. The code first checks whether the peg has been taken out of the sensor or not by checking the output of the photo-interrupter sensor. During removal of the peg, a low signal has been detected. The next thing that’s detected is the hand movement of the subject. Once both the sensor’s output and hand movement have been detected, the counter corresponding to hand movement is incremented. If the time taken by the subject to remove the peg from the sensor is more than 250 s, the “time out” message will be displayed, and the test will be terminated. Once all the pegs from the nine sensors have been removed, the hand movement count and the total time taken to perform the test will be displayed on the screen.

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test Fig. 3 A flowchart explaining the programming of the set-up

453

454

K. Aneesha Acharya and A. Choudhary

3 Results The initial part of the research work was to measure the peg insertion time. In this part, time was measured only till the last peg was inserted without using the third person’s interference with the help of LED lights. These LED arrangements guide the subject to place the pegs in an orderly fashion. Figure 4 shows the subject performing the test by inserting the pegs into the sensor. Once the insertion part is done, it is time to test the removal part and cross verify the results with the manual stopwatch. In the removal part, another important thing to be added was the glowing of LED’s backward, with the difference being the fact that as soon as the peg was removed, subsequent LED will light up contrary to the last part in which the LED used to light up once the peg was inserted. A specific pattern is followed, and this is ensured using the infrared (IR) sensor. The circuit has been coded so that unless the IR sensor has not detected one hand movement, the sensor will not detect the peg, making sure that the subject follows the rule. Once insertion and removal parts were done, the total time and hand movement were displayed on the screen, as shown in Fig. 5. HC-05 Bluetooth module is used, which transmits the data on the screen to mobile phone using an application called Bluetooth terminal (Fig. 6). Overall, five subjects of different age groups were tested, and the results are displayed as shown in Table 1. It can also be inferred from Table 1 that the time taken to complete the test increases as the age increases. In all the cases, the right hand is the dominant hand, and hence the test was done using the same as well. Thus, Table 1 shows the results obtained when each of the five subjects was tested for their reflexes associated with their right hand. The subjects were of varying with same hand dominance (right-hand user). This observation leads to believe that the average reaction time tends to gradually increase in the age range of 20–30 years, whereas the rise is exponential when seen in the range of 40–80 years. Figure 7 was the graphical analysis of the subjects when they were tested for the first time. The hand–eye coordination of all the subjects is improved by practicing pegboard tests at regular intervals. Normative data from the study [10] also confirm the same.

Fig. 4 Subject inserting the pegs

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test

455

Fig. 5 Output of the pegboard test experiment in the Arduino IDE environment

A separate pegboard test session was conducted after the two weeks of training, and the results are shown in Table 2. As shown in Table 2 and Fig. 8, the reaction time improved after repeatedly practicing the same test over a significant time. Hence it can be inferred that the same test was used to detect and improvise the patient’s condition.

4 Discussions The advantages provided by the automated pegboard to the traditional pegboard are as follows: • Presents the result with greater accuracy by using internal timer and appropriate sensors. • It allows registering the time for each one of the pegs. • It allows knowing if two or more pegs move at the same time. • It allows comparing the time taken to complete the test using the right hand and the left hand with greater accuracy. • It allows determining the number of pegs moved successfully during the test.

456

K. Aneesha Acharya and A. Choudhary

Fig. 6 Transmission of data using HC-05 Bluetooth module

Table 1 Age of the subject with the time taken for the test S. No.

Age

Time taken (s)

Person 1

09

25

Person 2

22

29

Person 3

45

35

Person 4

52

36

Person 5

80

60

• It gives a visual indication if the peg has been inserted properly or not. When the peg is inserted correctly, the LED goes off, and when the peg is removed, the LED is on. The pegboard was constructed with the help of breadboards, single strand wire, and jumper wire. Also, the insertion and removal sequence of the pegs were successfully achieved. The timer was incorporated into the system to calculate the whole

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test

Fig. 7 Graph representing the time taken for the pegboard test with age Table 2 Improved performance during pegboard test S. No.

Age

Time taken (s)

Person 1

09

22

Person 2

22

26

Person 3

45

34

Person 4

52

34

Person 5

80

56

Fig. 8 Rehabilitation graph showing improvement in the test performance

457

458

K. Aneesha Acharya and A. Choudhary

process’s completion time using the millis () function. All the functions of the pegboard were demonstrated using the flowcharts, block diagrams, and serial monitor and serial plotter outputs. In future work, portability of the device will be increased by developing pegboard in a 3D printer with proper Printer Circuit Board (PCB).

5 Conclusion Conducting the automated pegboard test gives vital information related to the individual participant’s hand–eye coordination and enhances the fine motor skills of the subject. One advantage of this test over the traditional tests is that this test can be conducted without any external supervision and can give accurate results compared to the conventional tests. The test process can be incorporated by including the diversified demographical nature of the participants in the future study. Due to the present pandemic situation, we have considered convenience sampling rather than random selection from the population. These samples individually might not be representative of the general population. The prototype will be transferred on the PCB board considering the design specifications and portable in future work.

References 1. Kellor M, Frost J, Silberberg N, Iversen I, Cummings R (1971) Hand strength and dexterity. Am J Occup Therapy Off Publ Am Occup Therapy Assoc 25(2):77–83 PMID: 5551515 2. Johansson GM, Häger CK (2019) A modified standardized Nine-Hole Peg test for valid and reliable kinematic assessment of dexterity post-stroke. J NeuroEng Rehabil 16(8). https://doi. org/10.1186/s12984-019-0479-y 3. Rivera D, García A, Ortega JE, Alarcos B, van der Meulen K, Velasco JR, Del Barrio C (2019) Intraindividual variability measurement of fine manual motor skills in children using an electronic pegboard: cohort study. JMIR Mhealth Uhealth 7(8). https://doi.org/10.2196/12434 4. Jobbágy Á, Marik AR, Fazekas G (2018) Quantification of the upper extremity motor functions of stroke patients using a smart Nine-Hole Peg tester. J Healthc Eng. https://doi.org/10.1155/ 2018/7425858 5. Wang YC, Bohannon RW, Kapellusch J, Garg A, Gershon RC (2015) Dexterity as measured with the 9-Hole Peg Test (9-HPT) across the age span. J Hand Ther 28(1):53–59. https://doi. org/10.1016/j.jht.2014.09.002 6. Feys P, Lamers I, Francis G, Benedict R, Phillips G, LaRocca N, Hudson LD, Rudick R (2017) Multiple sclerosis outcome assessments consortium. The Nine-Hole Peg Test as a manual dexterity performance measure for multiple sclerosis. Multiple Scler 23(5):711–720. https:// doi.org/10.1177/1352458517690824 7. Earhart GM, Cavanaugh JT, Ellis T, Ford MP, Foreman KB, Dibble L (2011) The 9-hole PEG test of upper extremity function: average values, test-retest reliability, and factors contributing to performance in people with Parkinson disease. J Neurol Phys Ther 35(4):157–163. https:// doi.org/10.1097/NPT.0b013e318235da08 8. Bowler M, Amirabdollahian F, Dautenhahn K (2011) Using an embedded reality approach to improve test reliability for NHPT tasks. IEEE Int Conf Rehab Robot. https://doi.org/10.1109/ ICORR.2011.5975343

Manual Dexterity Assessment Using a Nine-Hole Pegboard Test

459

9. Lemmens RJ, Janssen-Potten YJ, Timmermans AA, Smeets RJ, Seelen HA (2015) Recognizing complex upper extremity activities using body worn sensors. PLoS One 10(3). https://doi.org/ 10.1371/journal.pone.0118642 10. Mathiowetz V, Weber K, Kashman N, Volland G (1985) Adult norms for the Nine-Hole Peg test of finger dexterity. Occup Therapy J Res 5(1):24–38. https://doi.org/10.1177/153944928 500500102

Implementation of Indoor Navigation Control for Two-Wheeled Self-balancing Robot B. Vignesh, Deepa Jose, and P. Nirmal Kumar

Abstract Two-wheeled self-balancing robots (TWSBR) overcomes the limitations of conventional three or four-wheeled robots by offering a zero-turning radius with two motors. The usage of conventional dc motors with encoders in these two-wheeled self-balancing robots requires a complex design and tuning of cascaded controllers. In this paper, we implement a model that reduces the complexity of the design, and we have implemented a simple low-cost indoor navigation technique along with it, to know about the robustness of the design. The proposed system uses stepper motors in place of dc motors such that a single controller can be used that reduces the complexity of the design. We have implemented laser-based navigational control along with obstacle avoidance control on this robot for the indoor environments to show the stability of the robot during the motion for reaching the destination goal along with the detection and avoidance of the obstacles in the path of the robot. Keywords ROS · Lidar · Two-wheeled self-balancing robot · Obstacle avoidance · Navigation

1 Introduction In today’s scenario robotics are developed in different structures that have their own advantages when compared with other robotic structures. The conventional four-wheeled robots which are used for various applications in indoor and outdoor environments have their own limitations such as requirement of high number of wheels and motors which increases the energy consumption of the robot and also B. Vignesh · P. Nirmal Kumar Electronics and Communication Engineering, College of Engineering, Anna University, Guindy, Chennai 600025, India e-mail: [email protected] D. Jose (B) Department of Electronics and Communication, KCG College of Technology, Chennai, Tamil Nadu 600097, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_31

461

462

B. Vignesh et al.

these robots require additional spaces on the path for making some rotations or when changing its directions which makes it difficult to control over narrow paths. Such limitations were overcome by the two-wheeled self-balancing robots (TWSBR) which requires only two motors for locomotion and hence the power requirement is less and also it requires zero turning radius which is the most advantage of the robot making it suitable for narrow spaces [1]. Many papers were proposed with different types of cascaded controllers for better positional and directional stability. Cascaded controller like PD-PI was proposed [2] where the PD controller is maintaining the balance and the PI controller is tuned for the position control of the TWSBR and compared with the adaptive fuzzy controller. A two-level adaptive controller along with PD controller [3] was designed to tackle the instability of the TWSBR which provided a better stability, but the implementation is complex due to multiple controllers being cascaded. As given in [4] the sliding mode-based controllers offer a robust control of the TWSBR over the disturbance and noise. Cascaded PID controllers [5] were designed for TWSBR with conventional DC-geared motors. In the above-discussed papers have commonly used conventional DC-geared motors with encoders to identify the position of the shaft and to rotate it precisely. And also, the cascaded form of controllers discussed above are quite difficult to tune for a good stability. Some of the indoor navigational methodologies were studied to find a way for low-cost robust navigational technique. Ultrasonic based obstacle avoidance along with Fuzzy controller [6] were adopted for collision-free movement. A 3D SLAM was developed by using a rotating 2D laser scanner [7] that was aimed to provide a 3D map with 2D lidar sensor with a complex algorithm. Autonomous mobile robots were designed in the ROS environment [8] for establishing a mapping and path planning for indoor navigation. ROS is nowadays widely used in various robotic applications because of its dedicated platform for robotics. In this paper, we proposed a TWSBR model with single PID controller by using stepper motors in place of DC geared motors along with the integration of the 2D Lidar and ultrasonic sensors for a low-cost collision free navigation of the robot in the indoor environments.

2 Proposed System Design The proposed system has three sections where each section does different operations. The overall block diagram of the proposed system is given in Fig. 1. Those three sections are (i) Balancing controller design (ii) obstacle avoidance controller (iii) navigational controller. A. Balancing Controller Design It is the important design of the robot which ensures the stability of the robot at all circumstances that contains PID controller, IMU sensor, motor control, As said

Implementation of Indoor Navigation Control for Two-Wheeled …

463

Fig. 1 Overall block diagram of the TWSBR model

earlier, since we use Stepper motors in TWSBR a single PID controller is sufficient for maintaining the stability of the robot. The robot’s upright position is maintained by computing the error from the actual setpoint to the current angle of the robot. The current angle of the robot is measured by the MPU 6050 Inertial Measurement Unit (IMU) sensor. The values out from the sensor are not accurate; some offsets are there due to environmental noises. These will affect the stability of the system, Sun et al. [8] has used a kalman filter that gives more accurate sensor values at a cost higher computation. Since we are implementing in a low-cost hardware like an arduino, implementation of kalman filter increases the time lag which causes instability of the system. Hence, we have designed a complementary filter for the filtering purposes. The block diagram of the complementary filter is shown in Fig. 2. The complementary filter is simple to design and requires lower computations time. The equation corresponding to the complementary filter is shown in Eq. (1), where α is the constant calculated from the sampling frequency of the code. Tilt_angle = α ∗ (acc_Angle) + (1 − α) ∗ (gyro_angle)

Fig. 2 Complementary filter

(1)

464

B. Vignesh et al.

The error calculated is given as input to the PID controller. The PID controllers are linear aThe PID controller requires the value of three gains proportional gain (P), integral gain (I), derivative gain (D). Each gain has certain significance in some parameters like the peak overshoot, settling time, rise time, etc. By tuning the PID gain values by trial-and-error method offers a better controller for the robot. The equation for the PID controller is given in the Eq. (2).  pid_out = K p ∗ err + K i

err ∗ dt +

derr ∗ Kd dt

(2)

where K p —proportional gain, K i —integral gain, K d —differential gain. The output of the PID controllers is sent as pulse width modulation (PWM) pulses to the stepper motor drivers. The advantage of the stepper motors is the rotation angle of the shaft can be controlled by the number of pulses given to the motor driver and the speed of the motor is controlled by adjusting the frequency of the pulses. We use NEMA 17 stepper motors which is a low-cost stepper motor that offers a torque of 4.44 kg cm. Since it is a two-wheeled robot, I have implemented a differential drive system for the robot which makes turning or rotation easier. Simple driver circuitry and elimination of the encoder makes the design simpler and more effective while compared to the cascaded controllers. Compared to other controllers like fuzzy or LQR which might be more robust compared to PID are difficult in realtime implementation. The PID controllers are simple and its robustness is sufficient for the robot to maintain its stability during directional movements [9, 10]. To execute the directional commands unlike the normal robots the direction of the TWSBR is altered by changing the setpoint of the robot. Making the setpoint to positive makes the robot to move forward and setpoint to negative makes the robot to move backward. The directional commands are given from the other two sections based on the obstacles around the robot and optimal path chosen by the algorithm. The flowchart for the balancing algorithm is shown in Fig. 3. B. Navigational Controller For adding an indoor navigation technique, the methods discussed earlier uses sensors like lidar, RGBD camera, odometry sensors, etc., by fusing different sensor values together a sophisticated navigation can be implemented but this comes with higher hardware and computational costs. In this paper, I have used 2D lidar sensors for creating a map of the indoor environment and path planner creates a path based on the goal, and the directional commands are sent to the microcontroller of the balancing controller unit. And also, the model receives the directional commands from the obstacle avoidance section where both of these commands are fused in such a way that the model is able to locate itself in an environment and navigate to the respective goal without colliding with the dynamic obstacles present in the particular chosen path. This way of integrating the data from the lidar and ultrasonic sensors provides a better navigation with a low-cost investment. The slamtec 2D lidar is used which detects the distance of the obstacles by the laser triangulation method. The lidar sensor covers a distance of 8 m in radius with

Implementation of Indoor Navigation Control for Two-Wheeled …

465

Fig. 3 Flowchart of the balancing algorithm

a 360 angular rate. The sampling frequency of the lidar sensor is 5.5 Hz that gives approximately 4000 readings in a minute. The accuracy of the distance is high when compared with ultrasonic sensors with resolution of 0.5°. This navigation controller is developed in the Robotic Operating System (ROS) which is a dedicated middleware for the robotic applications. In the ROS we develop

466

B. Vignesh et al.

the code in blocks where each block of the code is called a node. One node for getting the data from the lidar, another node for creating the map from obtained data from the lidar and the other node for identifying the respective optimal path for the given destination and sends the commands to the balancing controller unit. Raspberry pi 4 is used for implementing this technique in ROS. At first every node was developed in the ROS environment of pi and simulated in the available sample home environments of the ROS successfully. But on real-time execution with realtime data from the lidar sensor since the GPU of the Raspberry pi is not sufficient for running all nodes along with the application softwares like Rviz and Gmapping simultaneously the outcomes on movement of the robot were not accurate. So, I have used the ROS toolbox of matlab software in remote pc for creating the map and running the path planner in it for finding the optimal path and directional commands are sent from the pc to the balancing controller unit. C. A* Algorithm The robot at first is made to move around the given indoor environment such that a map is generated from the lidar data. Then the local plannar uses A* algorithm to find the optimal route for the selected destination goal in the created map. The A* algorithm assign cost values to different motions in the adjacent cells and choses the cells with lower cost and repeat the algorithm. After generating the path, the directional commands are sent to the Balancing Controller Unit (BCU) via bluetooth for producing the respective movements. D. Obstacle Avoidance Unit The lidar sensor is capable of detecting the obstacles only to its range of vicinity. The obstacles that are located on the floor which are very much smaller are not detected by the lidar which might result in collision of the robot with the obstacle. To overcome this issue, we added a section called obstacle avoidance unit. Here, the ultrasonic sensors are used for obstacle detection which offers better obstacle detection at lower price. The developed model senses the nearby obstacles which are in the range of 50 cm that could collapse the stability of the robot. When the obstacle is detected the location of the obstacles is first determined by measuring the distance of the object from various sensors placed at different angles and then the direction of the movement is chosen. Whenever an object is detected the robot holds the directional commands sent by the navigational unit and waits for some seconds to find whether the object is moved from the path, if not moved then it avoids that particular obstacle by moving around it where now the directional commands are sent by obstacle avoidance unit and after overcoming the obstacle once again it resumes the navigational unit and receives the directional commands from the navigational unit for reaching the destination. We have used six ultrasonic sensors which are placed at different angles and different heights such that it covers 150° of area in front of the robot. Three sensors are placed at the lower part of the robot in a tilted manner so that they are capable of detecting obstacles on the floor which have a height above 1 cm. the sensors are tilted by an angle of 50°. And the other three sensors are placed in the top below the

Implementation of Indoor Navigation Control for Two-Wheeled …

467

lidar sensor where one in the middle is facing straight forward and the other two are facing towards the diagonal direction of the robot. This kind of placement of sensors offers an optimum obstacle avoidance for the TWSBR model.

3 Structure of the TWSBR Model The structure of the robot looks similar to the arrangement of the stacks, where each stack contains certain components of the robot. The structure of the robot is built over acrylic material which provides low weight structure. The stepper motors are attached under the base layer of acrylic sheet and above the base layer contains the obstacle avoidance unit placed on it that is interfaced with the ultrasonic sensors. The complete structure of the developed TWSBR model is shown in Fig. 4. The first layer consists of the balancing controller unit that comprises the MPU 6050 sensor and motor driver and microcontroller for maintaining the stability of the TWSBR model. The second layer sheet consists of the battery pack that powers the entire setup. A Lithium-ion battery of 12v 1800 maH battery is used which offers higher current density at lower weight. The third layer houses the raspberry pi 4 board which is going to fetch the data from the lidar sensor and send it to the computer via Fig. 4 Complete structure of the TWSBR model

468

B. Vignesh et al.

the ROS master URI protocol for further mapping and path planning. The Rplidar is mounted to the fourth layer which is the top most layer of the robot structure.

4 Results and Discussion The balancing algorithm was first experimented. Instead of modelling the controller in MATLAB and tuning the gain values, I have directly implemented on the hardware. First, the proportional gain is set such that the robot is able to balance with oscillatory movements; then secondly, the differential gain is adjusted to damp out the oscillatory movements, and lastly, the integral gain is set to nullify the offset in the position of the robot. The responses of the pid controller for disturbances and directional commands are plotted as a graph is shown below in Fig. 5a, b. From the above responses, it may be seen that output of the pid controller is near to zero when the robot is at still position. And pid is positive for forward direction and pid is negative for backward. Fig. 5 a Pid_output response on disturbances. b Pid_output response on direction movements

Implementation of Indoor Navigation Control for Two-Wheeled …

469

After computing the balancing algorithm, the obstacle controller was tested the ultrasonic sensors were able to detect the obstacles and sent the right directional commands to the balancing controller unit when an obstacle is detected which is shown in the Fig. 6a–c are given. At last, the navigation controller was tested for mapping and path planning. As said earlier the developed code was tested first with the simulational environments available in ROS environment is shown in Figs. 7 and 8 Fig. 9 shows the map created for simulation environment and Fig. 10 shows the execution of the path planning and navigation in the simulation window on the created map. After successful simulation, the controller was implemented in raspberry pi and MATLAB. The robot was able to create the map successfully and planned an optimal path for the particular destination which is shown in Figs. 11 and 12. After checking the functionalities of each controller unit everything was integrated together for a complete operation. As discussed in the abstract we are able to implement a TWSBR model with a single PID controller and also, we have successfully implemented a low-cost navigation system for the robot to navigate in the given indoor environment along with the obstacle avoidance.

5 Conclusion The proposed paper has discussed the implementation of the lidar and ultrasonic sensor-based navigation for a two wheeled self-balancing robot. Instead of using conventional DC motor with encoders we have used stepper motors which also produced a better stability like the dc motors which reduced the cost of the hardware and also the need of tuning multiple controllers are eliminated where a single PID controller is sufficient for maintaining the stability of the robot. The navigation using ROS environment by the data from the lidar for map creation and fusing the values of the ultrasonic sensors for a collision free navigation in the determined path defined by the A* star algorithm were successfully integrated together providing a better navigation technique for the robot with low-cost hardware.

6 Future Work The proposed TWSBR model is capable of maintaining the stability in the event of navigation and collision avoidance. The balancing algorithm has to modified such that the robot moves with same stability on the slopes. At present, the robot is able to stand still on a slope surface of about 20° but finds it difficult to move from the flat surface to the sloped surface which must be eliminated by the modification in the algorithm. Similarly, I have an idea on implementing swarm robotic concept to this structure such that multiple TWSBR robots can be coordinated for carrying

470 Fig. 6 a–c Directional commands sent to the BCU based on the position of the obstacle detected

B. Vignesh et al.

Implementation of Indoor Navigation Control for Two-Wheeled …

471

Fig. 6 (continued)

Fig. 7 Home environment for simulation

the load of different weights by varying the orientation and number of the robots involved in the load carrying task which ensures the efficient utilization of the robots on dynamically changing load weights.

472

Fig. 8 Initial mapping

Fig. 9 Complete map of the environment

B. Vignesh et al.

Implementation of Indoor Navigation Control for Two-Wheeled …

Fig. 10 Simulation of A* path planning in map

Fig. 11 Real-time binary occupancy grid map

473

474

B. Vignesh et al.

Fig. 12 Real-time path planning and navigation

References 1. Zimit A, Yap HJ, Hamza M, Siradjuddin I, Hendrik B, Herawan T (2018) Modelling and experimental analysis two-wheeled self balance robot using PID controller. https://doi.org/10. 1007/978-3-319-95165-2_48 2. Iwendi C, Alqarni MA, Anajemba JH, Alfakeeh AS, Zhang Z, Bashir AK (2019) Robust navigational control of a two-wheeled self-balancing robot in a sensed environment. IEEE Access 7:82337–82348. https://doi.org/10.1109/ACCESS.2019.2923916 3. Wasif A, Raza D, Rasheed W, Farooq Z, Ali SQ (2013) Design and implementation of a two wheel self balancing robot with a two level adaptive control. In: Eighth international conference on digital information management (ICDIM 2013), Islamabad, Pakistan, pp 187–193. https:// doi.org/10.1109/ICDIM.2013.6694021 4. Abeygunawardhana PKW, Defoort M, Murakami T (2010) Self-sustaining control of two-wheel mobile manipulator using sliding mode control. In: The 11th IEEE international workshop on advanced motion control, Nagaoka, Japan 5. Pratama D, Ardilla F, Binugroho EH, Pramadihanto D (2015) Tilt set-point correction system for balancing robot using PID controller. In: 2015 international conference on control, electronics, renewable energy and communications (ICCEREC), Bandung, Indonesia, pp 129–135. https://doi.org/10.1109/ICCEREC.2015.7337031 6. Ruan X, Li W (2014) Ultrasonic sensor based two-wheeled self-balancing robot obstacle avoidance control system. In: 2014 IEEE international conference on mechatronics and automation, Tianjin, China, pp 896–900. https://doi.org/10.1109/ICMA.2014.6885816 7. Fang Z, Zhao S, Wen S (2017) A real-time and low-cost 3D SLAM system based on a continuously rotating 2D laser scanner. In: 2017 IEEE 7th annual international conference on CYBER technology in automation, control, and intelligent systems (CYBER), Honolulu, HI, USA, pp 454–459. https://doi.org/10.1109/CYBER.2017.8446162 8. Sun F, Yu Z, Yang H (2014) A design for two-wheeled self-balancing robot based on Kalman filter and LQR. In: 2014 international conference on mechatronics and control (ICMC), Jinzhou, China, pp 612–616. https://doi.org/10.1109/ICMC.2014.7231628 9. Natarajan E, Hong LW, Ramasamy M, Hou CC, Sengottuvelu R (2018) Design and development of a robot gripper for food industries using Coanda effect. In: 2018 IEEE 4th international symposium in robotics and manufacturing automation, ROMA, https://doi.org/10.1109/ROM A46407.2018.8986699 10. Someshwaran M, Jose D, Paul JP (2020) Autonomous unmanned ground vehicle for enhancement of defence strategies. Lect Notes Netw Syst 89:873–880

Application of NIR Spectroscopy with Chemometrics for Discrimination of Indian Black Pepper Berries Arnab Giri, Dilip Sing, Sudarshana Ghosh Dastidar, Pallab Kanti Halder, Nanaocha Sharma, Pulok K. Mukherjee, and Rajib Bandyopadhyay

Abstract The objective of this study was to discriminate the samples with different piperine content using near-infrared reflectance (NIR) spectroscopy. Twenty black pepper samples were collected from different parts of West Bengal, India. Clustering analysis was carried out using principal component analysis (PCA) and linear discriminant analysis (LDA) on the collected samples. The cluster separability criterion was used for quantifying the separability. The separation index was obtained as 3368.4, which establishes the utility of the NIR instrument for discrimination of the black pepper samples based on their piperine content. Keywords NIR spectroscopy · Medicinal plant · Piperine · Black pepper · Principal component analysis (PCA) · Linear discriminant analysis (LDA) · Separability index

1 Introduction Piper nigrum L. (family—Piperaceae), commonly known as black pepper, is found in India, China, Brazil, and Southeast Asia. Being considered as the “King of Spices,” it is one of the earliest spices used for seasoning to food and to enhance the flavor of other spices [1]. Black pepper is a natural antioxidant and is extensively used in anti-inflammatory, anti-cancer, anti-tumor, anti-periodic, anti-bacterial, anti-fungal analgesic, and anti-pyretic treatment in traditional medicine like, Ayurveda, Unani, A. Giri · D. Sing (B) · S. G. Dastidar · R. Bandyopadhyay Department of Instrumentation and Electronics Engineering, Jadavpur University, Salt Lake Campus, Kolkata, India e-mail: [email protected] P. K. Halder · P. K. Mukherjee School of Natural Product Studies, Jadavpur University, Kolkata, India N. Sharma · P. K. Mukherjee Department of Biotechnology, Institute of Bioresources and Sustainable Development, Government of India, Imphal, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_32

475

476

A. Giri et al.

and traditional Chinese medicine. It also contains vitamin A, C, E, K, niacin, βcarotene, and cholesterol-reducing properties [2]. Piperine is the major alkaloid present in black pepper, responsible for its pungent smell, taste, and medicinal properties. The concentration of piperine in black pepper varies from 2 to 9% depending on climate, place of origin, and growing conditions, which regulates the degree of therapeutic effectiveness [3]. Due to this varying quality and its vast commercial importance, quality control and market grading of black pepper are needed by estimating piperine, the marker molecule concentration. Traditional quality assessment of black pepper involves visual and olfactory segregation, which does not consider the piperine concentration. Modern physiochemical studies involve estimation of piperine molecule concentration using analytical techniques like chromatography, mass spectrometry. However, these analytical processes are limited in applications as they are time-consuming, expensive, complicated, labor-extensive, and involve samples [4, 5]. NIR spectroscopy has proved to be an effective, low-cost, non-invasive, and accurate analytical method for estimating chemical constituents. The basic molecular bonds of X–H (X:C, O, N, and S), which are the primary basic components of any organic molecule, mainly reflect different features of the NIR spectrum. Nearinfrared reflectance spectroscopy (NIR) is an accurate and non-invasive analytical method, which is used in diverse industrial applications to study the characteristics of different chemical constituents. This is particularly useful in study of medicinal plants [6–8]. The technique offers several advantages over other techniques like localized surface plasmon resonance (LSPR) [9, 10], evanescent wave-based measurements [11], and surface-enhanced Raman measurements involving the synthesis and use of customized nanoparticle systems [12, 13]. NIR techniques offer higher sensitivity and require minimum sample preparation procedure. In this paper, the feasibility of NIR spectroscopy was studied for the quantification of black pepper. NIR spectroscopy has been used for cluster analysis to find the discrimination of black pepper samples with different piperine content in them. The NIR spectral data of black pepper samples were subjected to principal component analysis (PCA) and linear discriminant analysis (LDA). In addition, the cluster separability criterion has been used for quantitative measurement, and the separation index was calculated.

2 Materials and Methods 2.1 Sample Collection Specimens of P. nigrum L. were collected from the local markets and plantations of Kolkata and nearby districts of West Bengal, India. Twenty samples were considered for experimentation.

Application of NIR Spectroscopy with Chemometrics …

477

2.2 Chemical Analysis The RP-HPLC analysis of the black pepper berries was carried out using piperine as the marker compound. The RP-HPLC method was validated based on calibration curve with high correlation coefficient and low LOD, LOQ values. The solvent system was optimized (methanol: water 1% glacial acetic acid—60:40) with retention time for piperine at 5.5 min. The percentage content of piperine in the different samples varied from 0.53% w/w to 6.5% w/w. The details of experimental procedure and result can be found in [14].

2.3 Spectra Acquisition NIR spectra of the black pepper samples were collected using the DWARF-Star NIR spectrometer from StellarNet Inc. USA. It contains an InGaAs detector array of 256 diodes. The wavelength range is 900–1700 nm. The spectrometer is accommodated with a bi-directional fiber-optic bundle installed at the high-intensity contact probe, tungsten halogen lamp with 12 V/20 W bulb as the light source and a sample holder. The block diagram and the experimental setup are shown in Figs. 1 and 2, respectively. For each sample, 5 replicates of spectra were collected and stored for analysis. Fig. 1 Block diagram

Tungsten Lamp Source

Sample in sample holder

Bandpass Filter

Focusing Lens

Detector

478

A. Giri et al.

Fig. 2 Experimental setup

2.4 Software SpectraWiz, Spectrometer OS v5.3, 2013 software was used for collecting the NIR spectra. PCA and LDA techniques were implemented in Matlab V10.0 (Mathworks Co., USA).

3 Data Analysis 3.1 Principal Component Analysis (PCA) Principal component analysis (PCA) is a multivariate method to analyze a data table whose datasets are related by quantitative dependent variables [15]. The primary objective of the technique in this experiment was to obtain the important statistical measures to represent as principle components and display the similarities as spot maps on orthogonal axes.

3.2 Linear Discriminant Analysis (LDA) Linear discriminant analysis (LDA) is a very common technique for dimensionality reduction problems as a preprocessing step for machine learning and pattern

Application of NIR Spectroscopy with Chemometrics …

479

classification applications [11]. For reduction features, LDA technique uses a lesser dimension hyper-plane and in a direction that maximizes the separation between the samples.

3.3 Separability Index (SI) Separability is the quantitative measure to segregate the fairly different set of points in a dataset. The separability measure is obtained by the ratio of the trace of the “within class scatter matrix” (SW) to that of the “between class scatter matrix” (SB) [14].

4 Result and Discussion 4.1 Wavelength Versus Absorbance Plot The wavelength versus absorbance plot of the spectral data of black pepper samples is shown in Fig. 1. The X axis of each of the spectral profile represents the wavelength (in nm) in the range of 900–1700 nm, and the Y axis represents the absorbance.

4.2 Principal Component Analysis PCA analysis was applied on the black pepper spectral data. The PCA plot for the black pepper samples by considering the first two principal components is shown in Fig. 3. For each sample, five replicates were available, and thus, there are 60 points on the plot. The information content in the first two axes of PCA is PC1-89.31% and PC2-6.46%. The PCA plot revealed that the samples of same variety have been observed to appear as clusters, and no overlaps were observed among them (Fig. 4).

4.3 Discrimination Using Linear Discriminant Analysis The spectral data were subjected to LDA analysis. A cluster separation criterion has been used in our study for quantitative measurement, and the separation index was obtained as 3368.4. Thus, the discrimination obtained is better than PCA. The LDA plot is shown in Fig. 3 (Fig. 5).

480

A. Giri et al.

Fig. 3 Wavelength versus absorbance plot PCA SCORE PLOT OF BLACK PEPPER SAMPLES

0.8

S1 S2 S3 S4

0.6

S5

PC2 Explained: 22.622426 perc

S6 S7

0.4

S8 S9 S10

0.2

S11 S12 S13

0

S14 S15 S16 S17

-0.2

S18 S19 S20

-0.4

-0.6

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

PC1 Explained: 74.144083 perc

Fig. 4 PCA plot of black pepper samples

0.8

1

1.2

1.4

Application of NIR Spectroscopy with Chemometrics …

481

LDA plot of Black pepper samples 0.1

S1 S2 S3

0.08

S4 S5 S6 S7

0.06

S8 S9 S10

0.04

S11

LDA2

S12 S13 S14

0.02

S15 S16 S17

0

S18 S19 S20

-0.02

-0.04

-0.06 -0.12

-0.1

-0.08

-0.06

-0.04

-0.02

0

0.02

0.04

0.06

LDA1

Fig. 5 LDA plot of black pepper samples

5 Conclusion The results highlight that different samples of black pepper with variation in piperine content in them could be distinguished visually by PCA and LDA analysis. The separability index is obtained as 3368.4 which shows a very good discrimination using this technique. The results demonstrate that NIR spectroscopy is capable of discrimination between different black pepper samples without any prior chemical analysis. Acknowledgements This project was funded by the National Medicinal Plant Board, Ministry of AYUSH, Government of India [grant numbers: Z.18017/187/CSS/R&D/WB-02/2017/-18-NMPBIV A, 2017]. The authors express their gratitude toward the Institute of Bioresources and Sustainable Development, an autonomous institute under Departmentof Biotechnology (DBT), Government of India, Imphal, India for necessary help and support through IBSD-JU joint collaboration.

References 1. Black Pepper Plant, Editors of Encyclopaedia Britannica 2. Goswami T, Meghwal M (2012) Chemical composition, nutritional, medicinal and functional properties of black pepper: a review. J Nutr Food Sci 01(S1). https://doi.org/10.4172/scientifi creports.172

482

A. Giri et al.

3. Stojanovi´c-Radi´c Z et al (2019) Piperine-a major principle of black pepper: a review of its bioactivity and studies. Appl Sci 9(20):1–29. https://doi.org/10.3390/app9204270 4. Mukherjee PK, Bahadur S, Chaudhary SK, Kar A, Mukherjee K (2015) Quality related safety issue-evidence-based validation of herbal medicine farm to pharma. In: Evidence-based validation of herbal medicine, Elsevier Inc., pp 1–28 5. Mukherjee PK (2019) Chapter 17—Plant metabolomics and quality evaluation of herbal drugs. In: Mukherjee PK (ed) Quality control and evaluation of herbal drugs. Elsevier, pp 629–653 6. Sing D et al (2020) “Prediction of andrographolide content in Andrographis paniculata Using NIR Spectroscopy. IEEE Appl Signal Process Conf (ASPCON) 2020:335–338. https://doi.org/ 10.1109/ASPCON49795.2020.9276668 7. Sing D et al (2021) Estimation of andrographolides and gradation of Andrographis paniculata leaves using near infrared spectroscopy together with support vector machine. Front Pharmacol 12(May):1–8. https://doi.org/10.3389/fphar.2021.629833 8. Qu JH et al (2015) Applications of near-infrared spectroscopy in food safety evaluation and control: a review of recent research advances. Crit Rev Food Sci Nutr 55(13):1939–1954. https://doi.org/10.1080/10408398.2013.871693 9. Sadani K, Nag P, Mukherji S (2019) LSPR based optical fiber sensor with chitosan capped gold nanoparticles on BSA for trace detection of Hg (II) in water, soil and food samples. Biosens Bioelectron 134:90–96 10. Nag P, Sadani K, Mukherji S, Mukherji S (2020) Beta-lactam antibiotics induced bacteriolysis on LSPR sensors for assessment of antimicrobial resistance and quantification of antibiotics. Sensors Actuators B Chem 311:127945 11. Nag P, Sadani K, Mohapatra S, Mukherji S, Mukherji S (2021) Evanescent wave optical fiber sensors using enzymatic hydrolysis on nanostructured polyaniline for detection of β-lactam antibiotics in food and environment. Anal Chem 93(4):2299–2308 12. Nag P et al (2021) Polyphenol stabilized copper nanoparticle formulations for rapid disinfection of bacteria and virus on diverse surfaces. Nanotechnology 13. Sadani K et al (2020) A point of use sensor assay for detecting purely viral versus viral-bacterial samples. Sensors Actuators B Chem 322:128562 14. Sing D, et al (2021) Rapid estimation of piperine in black pepper: exploration of Raman spectroscopy. Phytochem Anal 15. López del Val JA, Alonso Pérez de Agreda JP (1993) Principal components analysis. Aten Primaria 12(6):333–338. https://doi.org/10.5455/ijlr.20170415115235 16. Tharwat A, Gaber T, Ibrahim A, Hassanien AE (2017) Linear discriminant analysis: a detailed tutorial. AI Commun 30(2):169–190. https://doi.org/10.3233/AIC-170729

A Comparative Study Between Partial Least Squares and Principal Component Regression for Nondestructive Quantification of Piperine Contents in Black Pepper by Raman Spectroscopy Dilip Sing, Sudarshana Ghosh Dastidar, Wasim Akram, Sourav Guchhait, Shibu Narayan Jana, Subhadip Banerjee, Pulok Kumar Mukherjee, and Rajib Bandyopadhyay Abstract The aim of this work was to compare principal component regression (PCR) and partial least squares (PLS) regression methods while estimating the piperine content in black pepper using Raman spectroscopy. The calibration and prediction models of the regression analysis on Raman spectra were developed using PCR and PLS algorithm. The efficiency of the developed models was evaluated by means of root mean square error of calibration (RMSEC), root mean square error of prediction (RMSEP), and correlation coefficient (R2 ). For PCR algorithm, these parameters were obtained as 0.1, 0.1, and 0.97, respectively; and for PLS regression, the parameters were found as 0.05, 0.08, and 0.99, respectively. The results revealed that Raman spectroscopy with PCR and PLS algorithm could be used for determining the concentration of piperine in black pepper with an accuracy of 92.35% and 94.74% respectively. Keywords Raman spectroscopy · Partial least squares (PLS) · Principal component regression (PCR) · Medicinal plant · Black pepper · Piperine

D. Sing (B) · S. G. Dastidar · W. Akram · S. Guchhait · R. Bandyopadhyay Department of Instrumentation and Electronics Engineering, Jadavpur University, Salt Lake Campus, Kolkata 700106, India e-mail: [email protected] S. N. Jana · S. Banerjee · P. K. Mukherjee School of Natural Product Studies, Jadavpur University, Kolkata 700032, India P. K. Mukherjee Institute of Bioresources and Sustainable Development, An Autonomous Institute under Department of Biotechnology, Government of India, Imphal 795004, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_33

483

484

D. Sing et al.

1 Introduction The multivariate calibration methods are utilized abundantly for extracting relevant information from spectral data for predicting concentrations of analytes in complex samples [1–5]. There are several regression models that have been used for the development of calibration models. In this regard, PCR and PLS are two of the commonly used regression analysis techniques. PCR takes care of the collinearity problem with a few factors. However, the collinearity problem could also be handled by PLS algorithm utilizing fewer latent variables than with PCR. Several reports reveal that PLS attains the minimum mean square error (MSE) with a lesser latent variable than PCR. Therefore, partial lest square regression provides better optimization of latent variables with less time as compared to PCR [6]. Here in this experiment, we have used Raman spectroscopy with black pepper samples. In black pepper, piperine is the most important marker molecule. It has lots of different pharmacological features like anti-inflammatory [7], anticancer [7], antibacterial, antioxidant [8], antidepressant [9], analgesic, and anti-pyretic [10]. For quality assessment of black pepper, the prediction of piperine content in black pepper is extremely important. The conventional methods for quantifying piperine involve high-end chromatographic and mass spectroscopic technologies. These techniques in combination with chemometrics help in assessment of the secondary metabolites present in plants [11, 12]. These methods are laboratory-based, which need tedious processing of samples causing sample destruction. Other detection techniques have been evolving that includes more advantages, such as in situ estimation without prerequisite of chemicals, portability, and low cost such as [13–15]. An alternative low-cost and effective analytical technique based on Raman spectroscopy shows enormous application in numerous fields and products like in essential oil [16], material science [17], geology [18], and medicine [19]. In our previous study, we explored the efficiency of Raman spectroscopy for estimation of piperine in black pepper [20]. In this paper, we present the comparison of the performances of PCR and PLS algorithms for the data of black pepper with Raman shift in the range of 1097–1627 cm−1 .

2 Materials and Methods 2.1 Black Pepper Samples and Chemical Analysis Twelve Piper nigrum samples were collected from different geographical locations of India. For chemical analysis, methanol extracts were prepared and then the analyses were carried out using the RP-HPLC system (Waters, USA). It is equipped with a LC-30AD pump, ultraviolet/visible detector (three Lines degasser) with volume of four hundred micro-litter and a rheodyne 7725i injector with twenty micro-litter loop. The details of experimental procedure and result can be found in [20].

A Comparative Study Between Partial Least Squares and Principal …

485

2.2 Spectra Acquisition The spectrum of Raman shift was collected using an IDRaman micro-Microscope from Ocean Optics, Inc., Florida, USA. It was equipped with the laser source (785 nm) and has an output power of 100 mW (40 mW at sample). A thermo-electric cooler was used for minimizing the effect due to variation in temperature. The groove frequency of the grating was 1200 lines/mm, and the spectral response range was 200–3200 cm−1 . Resolution was maintained at 8 cm−1 . The average of sixteen was taken for each sample, and the exposure time was set at 1 s. Before acquiring each spectral response, a clean and dry sample container is used for taking the background spectrum.

2.3 Principal Component Regression Principal component regression (PCR) is a data analysis method based on linear regression. In PCR, the spectral points primarily break up into suitable components (latent variables). The important relevant information of the spectral data is contained in the initial few latent variables [21].

2.4 Partial Least Squares Regression Partial least squares (PLS) regression is a mathematical model like principal components regression [20, 22]. In PLS, training data with maximum variance is projected to a new space to create the calibration model.

2.5 Software ToupView Program and the OceanView Program software (version 7, Ocean Optics, Inc., Florida, USA) were used for controlling the instrument. The implementation of the PCR and PLS regression technique was done in the R Studio.

3 Results and Discussions The raw data collected from the Raman spectrometer was used to PLS and PCR regression technique for building the calibration model. The detailed experimental workflow can be found in Fig. 2 in our paper [1] for training and testing data set

486

D. Sing et al.

with chemical analysis. The whole data was divided in two parts. Here, the number of data for calibration and prediction is 9 and 3, respectively. In this process, the number of components was decided by computing the lowest root mean square error cross-validation (RMSECV). The optimal model was chosen with the overall lowest RMSECV and highest prediction accuracy. The number of components versus RMSECV plot for raw spectral data using PCR and PLS algorithm is shown in Figs. 1 and 2, respectively. The calibration model for black pepper samples was built using 80% of the data and 20% of the data was used for testing. PCR and PLS regression algorithms were used and the content of piperine in black pepper was predicted. The results are shown in Table 1. FourPCR and PLS components were chosen as optimal. The average accuracy of prediction for PCR and PLS regression was found to be 92.35% and 94.73%, respectively. Also, the root mean square error of prediction (RMSEP) of three samples for PCR and PLS regression was found to be 0.1 and 0.08, respectively. The correlation coefficient (R2 ) for test samples for PCR and PLS regression was found to be 0.97 and 0.99, respectively. This revealed good correlation between reference concentration and predicted concentration of piperine in black pepper. The maximum and minimum standard deviations of the predicted piperine content were ± 0.012 and ± 0.167, respectively, for PCR and ± 0.006 and ± 0.138, respectively, for PLS regression. Also, the average RPD (residual prediction deviation) value of the test data set for PCR and PLS regression was obtained as 7.01 and 8.92, respectively.

Fig. 1 Number of components versus RMSECV using PCR algorithm

A Comparative Study Between Partial Least Squares and Principal …

487

Fig. 2 Number of components versus RMSECV using PLS algorithm

Table 1 Result of PCR and PLSR models of piperine in black pepper Model

LV

Calibration

Prediction

RMSEC

RC

2

RPDc

RMSEP

Rp 2

RPDc

PCR

4

0.1

0.96

5.12

0.1

0.97

7.01

PLS

4

0.05

0.99

10.43

0.08

0.99

8.92

4 Conclusion A comparative study has been conducted using principal component regression and partial least squares regression algorithms on the spectral data from a Raman spectrometer for predicting estimating piperine content in black pepper seed. The efficacy of Raman spectrometer for quality estimation of black pepper has been illustrated. The accuracy of prediction of piperine in black pepper samples using PCR and PLS was obtained as 92.35% and 94.73%, respectively. The correlation coefficient (R2 ) for test samples for PCR and PLS regression was found to be 0.97 and 0.99, respectively. Also, the root mean square error of prediction (RMSEP) for PCR and PLS regression was found to be 0.1 and 0.08, respectively. The average RPD value of the test data for PCR and PLS algorithm was obtained as 7.01 and 8.92, respectively. The results reveal the efficacy of both the regression analysis methods for the estimation of piperine in the black pepper using Raman spectral data.

488

D. Sing et al.

Acknowledgements The present study was sponsored by the National Medicinal Plant Board, Ministry of AYUSH, Government of India. The authors are grateful to the Institute of Bioresources and Sustainable Development, an autonomous institute under Department of Biotechnology (DBT), Government of India, Imphal, India, for necessary help and support through IBSD-JU joint collaboration.

References 1. Sing D et al (2021) Estimation of andrographolides and gradation of Andrographis paniculata leaves using near infrared spectroscopy together with support vector machine. Front Pharmacol 12(May):1–8. https://doi.org/10.3389/fphar.2021.629833 2. Sing D et al (2020) Prediction of andrographolide content in Andrographis paniculata using NIR spectroscopy. IEEE Appl Signal Process Conf (ASPCON) 2020:335–338. https://doi.org/ 10.1109/ASPCON49795.2020.9276668 3. Balabin RM, Safieva RZ, Lomakina EI (2007) Comparison of linear and nonlinear calibration models based on near infrared (NIR) spectroscopy data for gasoline properties prediction. Chemom Intell Lab Syst 88(2):183–188. https://doi.org/10.1016/j.chemolab.2007.04.006 4. Hemmateenejad B, Akhond M, Samari F (2007) A comparative study between PCR and PLS in simultaneous spectrophotometric determination of diphenylamine, aniline, and phenol: effect of wavelength selection. Spectrochim Acta Part A Mol Biomol Spectrosc 67(3–4):958–965. https://doi.org/10.1016/j.saa.2006.09.014 5. Ghasemi J, Niazi A (2001) Simultaneous determination of cobalt and nickel. Comparison of prediction ability of PCR and PLS using original, first and second derivative spectra. Microchem J 68(1):1–11. https://doi.org/10.1016/S0026-265X(00)00159-4 6. Yeniay O, Göktas A (2002) A comparison of partial least squares regression with other prediction methods. Hacettepe J Math Stat 31:99–111 7. Tasleem F, Azhar I, Ali SN, Perveen S, Mahmood ZA (2014) Analgesic and anti-inflammatory activities of Piper nigrum L. Asian Pac J Trop Med 7(S1):S461–S468. https://doi.org/10.1016/ S1995-7645(14)60275-3 8. Zarai Z, Boujelbene E, Ben Salem N, Gargouri Y, Sayari A (2013) Antioxidant and antimicrobial activities of various solvent extracts, piperine and piperic acid from Piper nigrum. LWT Food Sci Technol 50(2):634–641. https://doi.org/10.1016/j.lwt.2012.07.036 9. Li S, Wang C, Li W, Koike K, Nikaido T, Wang MW (2007) Antidepressant-like effects of piperine and its derivative, antiepilepsirine. J Asian Nat Prod Res 9(5):421–430. https://doi. org/10.1080/10286020500384302 10. Parmar VS et al (1997) Phytochemistry of the genus Piper. Phytochemistry 46(4):597–673. https://doi.org/10.1016/S0031-9422(97)00328-2 11. Mukherjee P et al (2016) Metabolomics of medicinal plants—a versatile tool for standardization of herbal products and quality evaluation of ayurvedic formulations. Curr Sci 111:1624. https:// doi.org/10.18520/cs/v111/i10/1624-1630 12. Mukherjee PK (2019) Chapter 17—Plant metabolomics and quality evaluation of herbal drugs. In: Mukherjee PK (ed) Quality control and evaluation of herbal drugs. Elsevier, pp 629–653 13. Nag P, Sadani K, Mohapatra S, Mukherji S, Mukherji S (2021) Evanescent wave optical fiber sensors using enzymatic hydrolysis on nanostructured polyaniline for detection of β-lactam antibiotics in food and environment. Anal Chem 93(4):2299–2308 14. Nag P, Sadani K, Mukherji S, Mukherji S (2020) Beta-lactam antibiotics induced bacteriolysis on LSPR sensors for assessment of antimicrobial resistance and quantification of antibiotics. Sensors Actuators B Chem. 311:127945 15. Sadani K et al (2020) A point of use sensor assay for detecting purely viral versus viral-bacterial samples. Sensors Actuators B Chem 322:128562

A Comparative Study Between Partial Least Squares and Principal …

489

16. Hanif MA et al (2017) Evaluation of the effects of Zinc on the chemical composition and biological activity of basil essential oil by using Raman spectroscopy. Ind Crops Prod 96:91– 101. https://doi.org/10.1016/j.indcrop.2016.10.058 17. Zhang X, Tan QH, Bin Wu J, Shi W, Tan PH (2016) Review on the Raman spectroscopy of different types of layered materials. Nanoscale 8(12), pp 6435–6450. https://doi.org/10.1039/ c5nr07205k 18. Jehliˇcka J, Edwards HGM (2008) Raman spectroscopy as a tool for the non-destructive identification of organic minerals in the geological record. Organic Geochem 39(4):371–386. https:// doi.org/10.1016/j.orggeochem.2008.01.005 19. Nabiev I, Chourpa I, Manfait M (1994) Applications of Raman and surface-enhanced Raman scattering spectroscopy in medicine. J Raman Spectrosc 25(1):13–23. https://doi.org/10.1002/ jrs.1250250104. 20. Sing D et al (2021) Rapid estimation of piperine in black pepper: exploration of Raman spectroscopy. Phytochem Anal 21. Lee KM, Herrman TJ (2016) Determination and prediction of fumonisin contamination in maize by surface-enhanced Raman spectroscopy (SERS). Food Bioprocess Technol 9(4):588–603. https://doi.org/10.1007/s11947-015-1654-1 22. SNJ, Dilip Sing PKM, Mallik R, Dastidar SG, Bandyopadhyay R, Banerjee S (2020) Prediction of andrographolide content in Andrographis paniculata using NIR spectroscopy. IEEE Appl Signal Process Conf 335–338

Power Quality Data Mining Using Hybrid Feature Extraction Technique Vidhya Sivaramakrishnan, Balaji Mahadevan, and Kamaraj Vijayarajan

Abstract This paper proposes an approach for identifying the nature of power quality disturbances using hybrid feature extraction technique combining S transform and Hilbert transform. Using the combined features obtained, the classification is performed using Extreme Learning Machine (ELM). The effectiveness of the proposed approach is tested using wide spectrum of power quality disturbances. The comparison with existing methods indicates that the proposed hybrid signal processing approach for feature extraction results in improved classification accuracy. Sensitivity of the proposed approach is examined for signals with noise and the results are presented. Keywords Power quality disturbance · S transform · Hilbert transform · Extreme learning machine

1 Introduction The distortion in the current and voltage signals lead to failure and malfunctioning of several sensitive loads. These distortions are mainly attributed to the switching load capacitors, power electronic equipments and natural phenomena like lightning [1, 2]. The recent advancements in classification algorithms along with efficient feature extraction techniques have encouraged the researchers to take up studies in power quality data analysis and classification. Numerous literatures have been reported on improving the efficiency of classification using new variants in soft computing V. Sivaramakrishnan (B) New Prince Bhavani College of Engineering and Technology, Chennai, India e-mail: [email protected] B. Mahadevan · K. Vijayarajan Sri Sivasubramaniya Nadar College of Engineering, Chennai, India e-mail: [email protected] K. Vijayarajan e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_34

491

492

V. Sivaramakrishnan et al.

and efficient signal processing techniques. The major challenge lies in selecting appropriate features that form the basis of classifying power quality disturbance. The time frequency algorithms have been extensively applied to extract the features and thereby recognize the type of disturbance. In [3], the authors have applied wavelet technique for extracting the features for classifying power quality disturbance. The performance of the classifier with different wavelet family has been tested under normal and noisy conditions. An S transform based feature extraction technique has been proposed in [4]. The statistical features extracted by applying S transform form the input to fuzzy expert system based classifier. The main advantage of S transform is that it has superior characteristics when the signals are prone to noise. Owing to this merit, it has widely applied for detection and classification of various time series events. The classification of events using S transform and TT transform has been discussed in [4]. The patterns generated as the result of applying transformation forms the basis for classification. Feature extraction using Hilbert transform has been discussed in [5]. The results of this method are compared with other feature extraction methods to highlight its merits. The authors have employed Teager energy operator and Hilbert transform for detecting voltage flicker events [6]. For efficient classification of power quality signals, the authors have integrated wavelet and Hilbert transform [7]. In this method, the features are extracted by applying Hilbert transform on all the decomposed coefficients obtained through wavelet transform. The results indicate that the proposed method has enhanced accuracy of classification. From the literature, it is evident that integrating two transformation techniques results for feature extraction results in improved classification accuracy. However, the classification with random noise signals remains a challenge and there is a need to explore efficient feature extraction and classification techniques. In this trend, the concept of integrating two transformation techniques for feature extraction generates interest. This paper focuses on a hybrid feature extraction technique using S transform and Hilbert transform. To highlight the merits, the performance of the hybrid feature extraction technique is compared with other feature extraction methods. The feature classification is done using extreme learning machine.

2 S Transform The S transform [8] is a feature extraction technique that provides the required data for analyzing the disturbances [8, 9]. This method incorporates the features of shortterm Fourier transform and the wavelet transform. The S transform is derived from wavelet transform by changing the phase of the window function or mother wavelet. The S transform  of the signal, y(t), is the product of the signal and a phase correction function e− j2 f t [8]. The S transform of y(t) is defined as

Power Quality Data Mining Using Hybrid Feature Extraction Technique

∞ S(τ, f ) =

493

y(t)ga f (τ − t) e−i2f t dt

(1)

−∞

where ga f (τ ) is the Gaussian modulation function defined as 

|f| − ga f = √ e 2π

τ2 2σ 2



(2)

and 1 |f|

σ =

(3)

The expression becomes ∞ S(τ, f ) = −∞



|f| y(t) √ e 2π

−

(τ − t)2 f 2/



2

e−i2



ft

dt

(4)

The discrete Fourier transform of the time series y(t) is expressed as   N −1 n  ) nk − (2i N = H y(kT )e NT K =0

(5)

/ By substituting f → n N T and τ → j T . the discrete S transform is obtained as below

N −1 m+n n  = G(m, n)ei2mπ j/N H S j T, NT N T m=0 where 2 2 2 G(m, n) = e−2π m /n and j, m, n = 0, 1,…, N − 1. The discrete inverse of the S transform can be obtained as ⎡ ⎤ N −1  N −1  n ⎣ ⎦ei2nk y(kT ) = S( j T, N T n=0 j=0

(6)

(7)

The rows and columns of the matrix obtained as output of S transform correspond to frequency and time, respectively. The features extracted from time, frequency, amplitude and from high frequency and low frequency areas characterize the power quality events.

494

V. Sivaramakrishnan et al.

Figure 1 illustrates the signal without disturbance together with the characteristics obtained from S transform. The characteristics in Fig. 2 depict these plots for sag disturbance.

Fig. 1 Signal without disturbance and its S transform characteristics

Fig. 2 Sag disturbance and its S transform characteristics

Power Quality Data Mining Using Hybrid Feature Extraction Technique

495

Even though the study of literature show that the S transform exhibits superior characteristics, the performance under noisy conditions remains a challenge. Hence, this work explores the combination of S transform with Hilbert transform for improved classification accuracy.

3 Hilbert Transform The Hilbert transform results in a signal whose amplitude is the same as that of the original signal and the phase component lags the original by 90°. The result of Hilbert transform is the convolution of original signal y(t) with function πt1 [5, 10]. Hilbert transform can be expressed as H (y(t)) = zˆ (t) = 

 H z (t) = y(t) = 



1 πt 1 πt

∗ y(t)

(8)

∗ zˆ (t)

(9)

As the integrals defining the convolution do not converge. HT can be expressed as below: 1 H (y(t)) = P V π

∞ −∞

1 y(τ ) dτ = P V (t − τ ) π

∞ −∞

y(t − τ ) dτ τ

(10)

where PV represents the Cauchy’s principal value of the singular integral. The discrete Hilbert transform is given by 1 1 = y H (t) = y(t) ∗ πt π

∞ −∞

y(λ) dλ t −λ

(11)

The discrete Hilbert transform output is 90° phase shifted and the complete signal is expressed as yc [k] = y[k] + j y H [k]

(12)

The envelope of the original signal is then defined as |y A [k]| =

/

y 2 [k] + y H2 [k]

(13)

The envelope of Hilbert transforms aids in discriminating different PQ disturbances.

496

V. Sivaramakrishnan et al.

Figure 3 shows the Hilbert transform envelope for normal signal, sag signal and sag signal with noise. In this work, the features extracted from S transform [11] and Hilbert transform are used for identifying power quality disturbances namely PQD1-sag, PQD2Harmonic, PQD3-Interruption, PQD4-Swell, PQD5-Swell with Harmonic, PQD6Sag with Harmonic, PQD7-Flicker, PQD8-Oscillatory Transient. The 9 features extracted from Hilbert envelope given in Table 1 together with 20 features extracted

Fig. 3 a Signal without disturbance. b Output of Hilbert transform for signal without disturbance. c Sag disturbance. d Output of Hilbert transform for sag signal. e Output of Hilbert transform for normal signal and sag signal with noise

Power Quality Data Mining Using Hybrid Feature Extraction Technique Table 1 Features extracted from Hilbert transform

497

F1—mean of the Hilbert array F2—standard deviation of the Hilbert array F3—maximum value of the Hilbert array F4—energy of the Hilbert array F5—kurtosis of the Hilbert array F6—skewness of the Hilbert array F7—median of the Hilbert array F8—range of the Hilbert array F9—minimum value of the Hilbert array

from S transform [12] are used for identifying the nature of disturbance. These features have been obtained by programming in MATLAB/Simulink environment.

4 Extreme Learning Machine Classifier Extreme Learning Machine (ELM) classifier consists of input layer, a single hidden layer and an output layer. In ELM, input weights are selected randomly and the output weights are calculated analytically [13]. The various activation functions for the hidden layer used in ELM include Sigmoid, Sine, Gaussian and hard limiting functions. Linear Activation function is applied for the output neurons. The merits of ELM include faster learning speed and enhanced generalization performance [11, 12, 14, 15]. The output Ok of the ELM network composed of H hidden neurons is described as Ok =

H 

wk j G j (V , b, X i ), k = 1, 2, . . . , C

(14)

j=1

where V represents (H × n) input weight, b represents (H × 1) bias values for each hidden neuron, W represents (C × H) output weight, C is multi-class classification representing distinct classes with N observations (X i , T i , i = 1, 2, …, n), Gj (.) is the output of the jth hidden neuron and Gj (.) is the activation function. The output of the jth hidden neuron Gj (.) with sigmoidal activation function is described as   N  k G j (V , b, X i ) = tanh b j + v jk xi (15) k=1

Equation (14) can be expressed in matrix form as

498

V. Sivaramakrishnan et al.

O = W Oh

(16)

where Oh is an H × n matrix which is given by ⎞ G 1 (V , b, X 1 ) G 1 (V , b, X 2 ) . . . G 1 (V, b, X N ) ⎟ ⎜ .. .. .. Oh = ⎝ ⎠ . . . ⎛

G H (V , b, X 1 ) G H (V, b, X 2 ) . . . G H (V , b, X N ) The target tik is given as tik = 1 if ci = k = −1 otherwise k = 1, 2, . . . C

(17)

where ci is the class label for X i In ELM, the output weights are calculated by presuming the network output Y equal to the coded class label t, W = O Oh+

(18)

where Oh+ is the Moore–Penrose pseudo inverse of matrix Oh The estimated class label is computed as 

C i = arg max Oik

(19)

k=1,2...c

5 Extreme Learning Machine Classifier The evaluation of performance of the proposed technique is done by taking into consideration different types of power quality events. For each disturbance, around 125 samples are generated using Matlab. In addition 30, 40 and 50 dB noise is added to the original signal to analyze the performance for signals with noise. For these signals, the features are obtained and the training and testing data set is generated. The classification accuracy obtained based on these data sets are illustrated in Tables 2 and 3. From Tables 2 and 3, it is evident that the combined feature extraction method has better classification accuracy in comparison with S transform based feature extraction method. For the signals with noise, the combined feature extraction method performs better. The S transform technique performs better for sag and swells signals with 30 dB noise. The analysis of the proposed method with Support Vector Machine (SVM) classifier is presented in Tables 4 and 5, respectively. The result indicates

Power Quality Data Mining Using Hybrid Feature Extraction Technique

499

Table 2 Training accuracy for the combined feature extraction method Signal Without noise

50 dB noise

40 dB noise

30 dB noise

S transform Hybrid S transform Hybrid S transform Hybrid S transform Hybrid features features features features PQD1 99

100

99

100

100

100

96

100

PQD2 100

100

99

100

100

100

99

100

PQD3 100

100

100

100

100

100

100

100

PQD4 100

100

100

100

100

100

90

100

PQD5 100

100

99

100

99

100

99

100

PQD6 98

100

99

100

100

100

95

100

PQD7 100

100

99

100

100

100

81

100

PQD8 100

100

100

100

99

99.15

100

100

Table 3 Testing accuracy for the combined feature extraction method Signal Without noise

50 dB noise

40 dB noise

30 dB noise

S transform Hybrid S transform Hybrid S transform Hybrid S transform Hybrid features features features features features features features features PQD1 95

100

93

100

80

100

93

90

PQD2 100

100

100

100

100

100

77

100

PQD3 100

100

100

100

100

100

98

100

PQD4 100

100

100

100

100

100

97

94.11

PQD5 100

100

100

100

100

98.14

100

100

PQD6 100

100

100

98.14

100

100

100

100

PQD7 100

100

100

100

95

100

95

100

PQD8 100

100

98

100

90

100

98

100

that the classification accuracy of both ELM and SVM classifiers has improved with hybrid feature selection. Table 4 Training accuracy with different classifiers Feature selection

ELM (%)

SVM (%)

S transform

98

97.5

Hybrid features

99.83

99.42

Table 5 Testing Accuracy with different classifiers Feature selection

ELM (%)

SVM (%)

S transform

94

94

Hybrid features

99.05

98.5

500

V. Sivaramakrishnan et al.

6 Performance of Classifiers with Real Time and Dataset The performance of the proposed hybrid feature extraction technique is evaluated with the power quality event data set available in national database repository of power system events (https://pqmon.epri.com/). From the database data set with respect to events sag, oscillatory transient and interruption are taken and given as input to the proposed algorithm employing feature extraction and ELM classifier. Figure 4 illustrates the waveform of the events obtained from the data set. The accuracy of classification results are illustrated in Table 6 and the results signify the accuracy of hybrid feature extraction in classification of power quality events. Site0006

Site0010

Event 3557 - 2006-02-12 20:29:36.9340

Event 2787 - 2005-07-27 21:55:35.6620 12000

10000

10000

8000

6000

6000

4000

4000 INST AN VOLTS

INST AN VOLTS

8000

2000 0

-2000

2000 0

-2000

-4000

-4000

-6000

-6000

-8000

-8000

-10000

-10000

-12000 0

200

400

600

800

1000

Time (s)

Electrotek Concepts®

1200

1400

0

PQDiffractor®

Electrotek Concepts®

100

200

300

400

500 Time (s)

(a)

600

700

800

900 PQDiffractor®

(b) Site0010 Event 2857 - 2006-05-19 19:44:57.5420

3000 2500 2000 1500 INST AN VOLTS

1000 500 0 -500

-1000 -1500 -2000 -2500 -3000

0

Electrotek Concepts®

200

400

600 Time (s)

(c)

800

1000

1200 PQDiffractor®

Fig. 4 Waveforms obtained from the national database repository of power system events. a Sag waveform. b Oscillatory transient. c Interruption

Power Quality Data Mining Using Hybrid Feature Extraction Technique

501

Table 6 Classification with respect to real-time data Events

Total

Number of correct classification

Sag

9

8

Oscillatory transient

2

1

Interruption

1

0

7 Conclusion This work investigates the performance of hybrid feature extraction method combining S transform and Hilbert transform for classifying single and combined power quality disturbances. ELM is used as the classifier for classifying the disturbances. The analysis illustrates that the hybrid feature extraction technique improves the classification accuracy. The proposed approach performs better for noisy signals which have been a major concern in the area of power quality disturbance classification. The hybrid feature extraction technique improves the classification accuracy of both ELM and support vector machine classifier to indicate that this technique can satisfactorily classify power quality disturbances. The performance of the proposed feature extraction techniques is tested on real-time power quality events obtained from national database repository of power system events and the classification accuracy from these data sets highlight the efficacy of the proposed feature extraction technique. Future work will focus on optimization of features using evolutionary computation algorithm.

References 1. Ahsan MK, Pan T, Li Z (2018) A three decades of marvellous significant review of power quality events regarding detection & classification. J Power Energy Eng 6:1–37 2. Moravej Z, Banihashemi SA, Velayati MH (2009) Power quality events classification and recognition using a novel support vector algorithm. J Energy Conv Manag 50:3071–3077 3. Uyar M, Yildirim S, Gencoglu MT (2008) An effective wavelet-based feature extraction method for classification of power quality disturbance signals. J Electr Power Syst Res 78:1747–1755 4. Behera HS, Dash PK, Biswal B, Power quality time series data mining using S-transform and fuzzy expert system. J Appl Soft Comput 10:945–955 5. Suja S, Jerome J (2010) Pattern recognition of power signal disturbances using S Transform and TT transform. J Electr Power Energy Syst 32:37–53 6. Jayasree T, Devaraj D, Sukanesh R (2010) Power quality disturbance classification using Hilbert transform and RBF networks. J Neurocomput 73:1451–1456 7. Abdel-Galil TK, El-Saadany EF, Salama MMA (2004) Online tracking of voltage flicker utilizing energy operator and Hilbert transform. IEEE Trans Power Delivery 19(2):861–867 8. Sahani M, Mishra S, Ipsita A, Upadhyay B (2016) Detection and classification of power quality event using hybrid wavelet-Hilbert transform and extreme learning machine. In: 2016 international conference on circuit, power and computing technologies, pp 1–6 9. Stockwell RG, Mansinha L, Lowe RP (1996) Localization of the complex spectrum: the Stransform. IEEE Trans Signal Process 44:998–1001

502

V. Sivaramakrishnan et al.

10. Rodríguez JA, Aguado F, Martín JJ, López F, Mu˜noz JE, Ruiz (2012) Rule-based classification of power quality disturbances using S-transform. J Electr Power Syst Res 86:113–121 11. Vidhya S, Kamaraj V (2017) Particle swarm optimized extreme learning machine for feature classification in power quality data mining. Automatika 58(4):487–494 12. Huang G-B, Zhu Q-Y, Siew C-K (2006) Extreme learning machine: theory and applications. J Eurocomput 70(1–3):489–501 13. Abd-el-Maleka MB, Abdelsalam AK, Hassana OE (2018) Novel approach using Hilbert transform for multiple broken rotor bars fault location detection for three phase induction motor. ISA Trans 80:439–457 14. Ahila R, Sadasivam V, Manimala K (2015) An integrated PSO for parameter determination and feature selection of ELM and its application in classification of power system disturbances. J Appl Soft Comput 32:23–37 15. Mishra M (2019) Power quality disturbance detection and classification using signal processing and soft computing techniques: a comprehensive review. Int Trans Electr Energy Syst 1–42

Low-Cost, IOT-Based Child Safety Monitoring Robot with User-Friendly Mobile App Kalyan Kasturi, Rajani Dharanikota, Khaleelu Rehman, Senthilkumar Meyyappan, and Akhil Kommineni

Abstract Current population trend in India is moving from single-income to dualincome families, and there is an increasing need for parents to remotely monitor the safety of their children, while they are busy at work. In this research work, we propose a low-cost child safety monitoring robot based on IOT technology that can perform real-time live video streaming of the children and send notifications about child safety to the parents. Using this low-cost, IOT-based child safety monitoring robot, parents can monitor their child’s activities in real time through live video feed by controlling the robot’s movement in real time from anywhere using a user-friendly mobile app. In addition to this, the parents will get alerted automatically through notifications in case of any gas leakage or fire accident. The main advantage of the proposed system is instead of using mobile network (GSM) we are using Webhooks for sending alerts in case of gas leakage or fire accident which reduces the cost of hardware and works perfectly when connected to the home network. The entire IOT-based child safety monitoring system developed in this research paper makes of low-cost, common electronic components and hence provides a cost-effective IOT system for child safety monitoring. Keywords NodeMCU · Arduino · Real-time database · Webhooks · Wireless camera · Child safety and child monitoring

1 Introduction The Internet of Things (IOT) technology consists of a network of connected objects and modernizes the human society in several aspects [1]. The Internet of Things has become a popular technology bandwagon with multitudes of applications in wide-ranging scientific fields such as smart cities, healthcare systems, automobile systems and financial managements systems. The IOT technology is also used in the development of smart houses [2]. Most of the connected objects in the IOT network K. Kasturi (B) · R. Dharanikota · K. Rehman · S. Meyyappan · A. Kommineni ECE Department, Nalla Malla Reddy Engineering College, Hyderabad 500088, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_35

503

504

K. Kasturi et al.

are based on some kind of microcontroller device and used to collect information and perform monitoring in houses or offices [3]. There has been research work conducted in the area of security and surveillance based on IOT technology. The general idea of research work in this field is to employ a digital camera which is connected to the Internet. The digital camera connected to the Internet or called as a web camera serves as a connected object which can be used to perform some form of surveillance in some particular location of interest on a particular subject or person of interest. The web camera connects to the Internet which serves as the backbone network of the IOT system. Zafar et al. [4] investigated about building a real-time monitoring system with the help of IOT devices and cloud technology in their research work using Arduino and cloud service. The research work performed by Deshmukh et al. [5] discussed the use of Internet-based camera or webcam technology for the purpose of intelligent surveillance. Ramakrishna and Swathi [6] discussed using IOT technology for developing a smart security surveillance system. In addition to the surveillance research, another related area of interest in home security or safety is the detection of gas leakage system. In the country of India, in which millions of households totally and completely depend on the LPG gas for cooking their food materials on a daily basis, the gas leakage detection is a topic of significant interest. Gupta [7] discussed developing a gas leakage detection and alerting system. Sharma [8] developed a microcontroller-based LPG gas leakage detector which sends alert with the help of GSM module in case of any gas leakage. In this research paper, we develop a low-cost IOT-based child safety monitoring robot that performs video surveillance of children present in the house and generates live video feed which can be monitored by their parents who are present at office or some other location outside their home. We also developed a complete IOT-based child monitoring as well as safety system, by performing gas leakage detection and flame detection. The low-cost IOT-based child safety monitoring robot can be very conveniently monitored by the parents using a simple mobile app, which was developed as part of the proposed research work.

2 Model of IOT-Based Child Safety Monitoring Robot This research work is aimed at developing a low-cost, IOT-based robot which can be controlled remotely from anywhere by using an android application to help parents to monitor their child’s activities. Nowadays, mostly CC cameras are being used for this monitoring purpose. But children will not stay at a single location or room always in the house. Fixing cameras at each and every location in the house is practically not possible, and also, it is a costly process, and maintaining all those cameras is a tough task for the parents. To overcome all these problems, in this research work, we propose low-cost, IOT-based robot which is capable of monitoring the child’s activities and automatically alerting the parents in case of any danger in the house, especially near the surroundings of the child. This low-cost, IOT-based robot can be used by the

Low-Cost, IOT-Based Child Safety Monitoring Robot …

505

parent to monitor their child’s activities whenever they want from anywhere like office, workplace, etc., using an android application which will be user-friendly and can be used by anyone with basic knowledge of using smartphones. Currently, for alerting the user in case of fire detection and gas leakage detection, GSM-based systems are being widely used. GSM-based alerting systems consist of SIM card which is used for sending alerts in case of any accident. The main purpose of low-cost, IOT-based robot developed in this research work is for indoor use. GSMbased alerting systems may not work properly due to various reasons like low signal strength. So, in case of any danger in the house if there is no proper signal strength then there may be a delay in sending the alert to the parent. This delay may result in a great danger as we know fire and LPG leak can cause huge danger in confined places like houses if no proper action is taken as soon as possible. To overcome all these problems, we are using Webhooks in place of GSM-based alerting system. Webhooks are automated messages sent from apps when something happens. They have a message—or payload—and are sent to a unique URL—essentially the app’s phone number or address. Webhooks are almost always faster than polling and require less work on your end. They are much like SMS notifications. So, when something like gas leakage occurs, the sensor will send the alert to the NodeMCU. As we know NodeMCU is an IOT development board which is capable of connecting to the Internet and capable of sending and receiving the data using HTTP requests. In this research paper, we are making use of cloud computing to connect both the android application and robot virtually through Internet. So, this process makes the possibility of controlling the robot from anywhere with the help of Internet. A block diagram of the low-cost, IOT-based child safety monitoring system is as shown in Fig. 1. Initially, both the robot and the user’s smartphone in which the controller application is installed should be connected to the Internet. Once the user initiates a command, then that particular instruction will be first sent to the real-time database. Real-time database works in such a way that if any changes occur in the existing data present in the database, then the new data will be updated in the configured application or device. Here, the real-time database is connected to the NodeMCU. So, as soon as the command is initiated by the user, that particular command will be sent in real time to the NodeMCU which is one of the main components of the child safety monitoring robot. From the NodeMCU, the data will be transferred to Arduino UNO via serial communication. Then, based on the instruction received, Arduino UNO sends the control signals to the L298N motor driver module which controls the movement of DC motors. The above-mentioned process completes within seconds which makes it possible to control the robot’s movement in real time. Wireless camera will be mounted on the dashboard of the robot. The camera will be configured with home WIFI network. So, the user can watch the live video feed in order to monitor the child’s activities as well as observing the surroundings in order to control the robot’s movement.

506

K. Kasturi et al.

Fig. 1 Block diagram representation of the proposed low-cost, IOT-based child safety monitoring robotic system

For the detection of fire and gas leakage, two different sensors are being used: MQ5 sensor for gas leakage detection and IR flame sensor for fire detection. Both the IR flame sensor and MQ5 sensor will be directly connected to the NodeMCU. In case of gas leakage or fire detection, sensors will send the data to NodeMCU. NodeMCU is preconfigured in such a way that in case of fire or gas leakage detection a particular set of commands must be executed. The preconfigured logic consists of a code to send a GET request to the preconfigured Webhook URL. GET is one of the HTTP methods used to communicate with the servers. GET is used to retrieve and request data from a specified resource in a server. GET is one of the most popular HTTP request techniques. In simple words, the GET method is used to retrieve whatever information is identified by the request-URL. Once the logic is triggered, a GET request will be initiated to the Webhook URL. A Webhook URL works in such a way that if it is triggered then an alert will be automatically sent to preconfigured email address or phone number with the preconfigured alert message like “FIRE DETECTED” or “GAS DETECTED”. Using Webhooks instead of GSM-based alerting system reduces the cost as well as size and weight of the hardware.

Low-Cost, IOT-Based Child Safety Monitoring Robot …

507

3 Components and Technologies Employed in Child Safety Monitoring Robotic System In this research work, we are making use of NodeMCU [9] for connecting Arduino UNO with the real-time database because Arduino UNO [10] itself does not have any inbuilt hardware to connect to the Internet. NodeMCU is a low-cost open-source IOT platform. It initially included firmware which runs on the ESP8266 Wi-Fi SOC from Espressif Systems and hardware which was based on the ESP-12 module. We are making use of Arduino UNO for sending the control signals to motor driver module in order to control the movement of DC motors. L298N driver module [11] is employed to control the movement of robot in different directions based on the instructions received from the Arduino UNO microcontroller. MQ5 gas sensor and IR flame sensor are employed for detecting gas leakage and fire accident. An android application which is used to control the movement of the robot is developed using MIT app inventor. MIT app inventor is an intuitive, visual programming environment that allows everyone to build fully functional apps for smartphones. We used Google Firebase, a real-time database, to transfer the instructions from android application to the robot in real time.

4 Demonstration Kit The entire hardware components for the child safety monitoring robotic system were connected together to develop the demonstration kit. Figure 2 illustrates the side view of the demonstration kit showing the important hardware parts. We can observe in Fig. 2 that the Arduino UNO board which is the heart of the child safety monitoring robotic system can be located at the center of the kit. Wireless camera which presents the live video feed for child monitoring is fixed at the front side of the robot’s chassis for the purpose of monitoring as well as guiding the user in order to move the robot and can be seen on the far right-hand side of the picture. A top view of the demonstration kit is shown in Fig. 3. The motor drivers which control the DC motors used to move the child safety monitoring robot in various directions are connected to the kit, as shown in the picture. Remaining hardware components such as MQ5 gas sensor and IR flame sensor are placed on the chassis as shown in Fig. 3. Power supply for all the hardware components is given through onboard rechargeable battery. The entire cost of all components used in the child monitoring system is less than Rs. 5000, and hence, this child monitoring system can be built with very less cost.

508

K. Kasturi et al.

Fig. 2 Side view of the demonstration kit of the low-cost, IOT-based child safety monitoring robot

5 Experimental Results If a parent who is at some location other than home such as workplace or office wants to check the status of their child who was alone in the house, then the parent can start the child safety monitoring by launching the android application and corresponding application for wireless camera which were already installed in his/her smartphone. The android mobile application is developed in a user-friendly approach so that by clicking the left/right/front/back arrows on the mobile app the parent can move the IOT-based child safety monitoring robot as per their requirements and the parent can watch the child’s activities in real time. The working of the low-cost, IOT-based child safety monitoring robot was successfully verified by moving the demonstration kit presented in Figs. 2 and 3 in various directions and collecting the live video feed of the child present in the home in real time. In Fig. 4, the live video feed of child collected in real time by the low-cost, IOT-based child safety monitoring robot is presented. In Fig. 4, we can observe the child who is seated, and in addition, we can note the time and date stamp at the top, right-hand corner of the picture. In case of fire accident or gas leakage detection near the child, an alert notification will be automatically sent to the parent through SMS or email based on their preference. Figure 5 shown below illustrates the notification which will be sent to the parent in case of fire accident. We can clearly observe that the fire detection message was sent by seeing “What: fire_detected” header displayed in the picture. From Fig. 5, we can also observe that the fire detection message was sent using Webhooks. In

Low-Cost, IOT-Based Child Safety Monitoring Robot … Fig. 3 Top view of the demonstration kit of the low-cost, IOT-based child safety monitoring robot

Fig. 4 Live video feed of child with time and date stamp of the low-cost, IOT-based child safety monitoring robot

509

510

K. Kasturi et al.

Fig. 5 Fire detection message using Webhooks delivered by the demonstration kit of the low-cost, IOT-based child safety monitoring robot

addition, we can observe the time and date stamp of fire detection notification in Fig. 5.

6 Conclusion Currently, IP cameras are available for child monitoring purpose. It is not possible to fix cameras at each and every location in the house. Installing cameras at each and every location in the house is costly process, and also, it is not possible to monitor all the cameras. So, by using this low-cost IOT-based child safety monitoring robot, parents can monitor their children’s activities from their workplace, etc., and will get alerted automatically in case of gas leakage or fire accident. Once the robot is connected to the Internet, it can be controlled from any location such as office, workplace and almost anywhere in the world. Parent can control the robot’s movements by operating the left/right/front/back arrows on the mobile app and can move the robot in all directions. The parent can watch the live video feed through the onboard IP camera. Also, the child monitoring system can be built with very less cost. This low-cost child monitoring system will be very useful for the parents whose children will reach home from schools before their parents arrive from their workplace.

References 1. Atzori L, Lera A, Morabito G (2010) The Internet of Things: a survey. Comput Netw 54(15):2787–2805 2. Kasturi KS, Reddy PV, Rao AN, Vinod S (2016) A review of architecture and applications for Internet of Things. Adv Nat Appl Sci 10(9):261–266 3. Kasturi K, Sri Mourya K, Mahendra I, Maheshwari V (2017) Gesture recognition based device control using MEMS accelerometer. Int J Eng Sci Adv Comput Bio-Technol 8(4):256–264 4. Zafar S, Miraj G, Baloch R, Murtaza D, Arshad K (2018) An IoT based real-time environmental monitoring system using Arduino and cloud service. Eng Technol Appl Sci Res 8(4):3238–3242 5. Deshmukh A, Wadaskar H, Zade L, Dhakate N, Karmore P (2013) Webcam based intelligent surveillance system. Int J Eng Stud 2(8):2278–4721 6. Ramakrishna U, Swathi N (2016) Design and implementation of an IoT based smart security surveillance system. Int J Sci Eng Technol Res 5(4):697–702

Low-Cost, IOT-Based Child Safety Monitoring Robot …

511

7. Gupta A (2017) Economical and optimal gas leakage detection and alert system. Int J Sci Res Publ 7(11):260–263 8. Microcontroller Based LPG Gas Leakage Detector using GSM Module: Engineers Garage. https://www.engineersgarage.com/contributions/microcontroller-based-lpg-gas-leakage-det ector-using-gsm-module/. Last accessed 03 Feb 2021 9. NODEMCU—A perfect board for IOT. https://www.circuito.io/blog/nodemcu-esp8266/. Last accessed 16 May 2021 10. Arduino UNO. https://www.arduino.cc/en/Main/arduinoBoardUno>. Last accessed 16 May 2021 11. L298N Driver Module. https://components101.com/modules/l293n-motor-driver-module/. Last accessed 17 May 2021

Secure Image Classification Using Deep Learning K. Gururaj, Alaka Ananth, and Sachin S. Bhat

Abstract Machine learning and security are the buzzwords these days. Just like other fields, privacy concern is a major issue in machine learning systems as well. Current privacy techniques focus on allowing multiple input parties to collaboratively train machine learning models without releasing their private data in its original form. One of the most sensitive data in this regard is medical images. Usage of such data for collectively training models might be against the policies of hospitals, which assure patients that their information would be kept confidential. In such a scenario, privacy preserving machine learning poses several advantages over the conventional methods. In this paper, we have implemented a secure machine learning model based on the multi-party protocol described in SecureML (Mohassel and Zhang in 2017 IEEE symposium on security and privacy. IEEE, pp 19–38, 2017, [1]), on the medical dataset of X-ray images for pneumonia. The performance of these privacy preserving techniques against conventional machine learning algorithms is evaluated. Keywords Image classification · Neural network · Linear regression · X-ray images

K. Gururaj · A. Ananth NMAM Institute of Technology, Nitte, India S. S. Bhat (B) Shri Madhwa Vadiraja Institute of Technology and Management, Bantakal, Udupi, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement and Instrumentation, Lecture Notes in Electrical Engineering 957, https://doi.org/10.1007/978-981-19-6913-3_36

513

514

K. Gururaj et al.

1 Introduction Machine learning (ML) algorithms are widely popular in today’s applications. They are commonly used to produce predicting models for use in pharmacy, accounting, advisory systems, and vulnerability detection. Machine learning aims to transform industries through its novel and impactful applications. Machine learning involves manipulating huge quantities of training data. Consequently, large amount of data collection leads to the problem of its privacy. With rising concerns of digital breaches and security concerns, ensuring privacy of data in machine learning is an important concern. For instance, machine learning models that predict human diseases using images require sensitive medical data. While these algorithms indeed benefit humans, they require access to data that are usually private to hospitals. In such scenarios, hospitals, legal rules, and restrictions may prohibit them from outsourcing their medical data. With the standardization of newer and better machine learning techniques, there is a need to integrate data privacy with machine learning arises, which led to the concept called privacy preserving machine learning. Privacy preserving ML through secured multi-party computation (MPC) is a propitious solution that enables distinct entities to train different models on their combined data without revealing any information beyond the outcome. The leakage of data used for training could be potentially leaked by the models, and this has been studied extensively [2– 4]. However, the assertions given by privacy preserving machine learning provide a first line of defense that can be supported by integrating with other privacy-related methods such as differential privacy [5, 6]. Recently, there have been data collection schemes that need privacy as a necessary feature owing to the sensitivity of the data. The most common setting for privacy preserving machine learning is one where clients and servers interact in the following way: Multiple clients encrypt their data and send it to the server, which performs training on the combined data to classify new data samples. This being said, conventional training algorithms do not incorporate privacy inherently. Therefore, privacy preserving machine learning solutions aim at optimizing this gap [7–10]. In the last few years, privacy (especially differential privacy) has acquired popularity as a formal and quantifiable measure of privacy risk in data publishing [11, 12]. The most common scenario considered in these privacy preserving techniques is the server-aided model. Here, owner (otherwise known as clients) sends encrypted data to multiple servers which perform the training procedure on combining similarly sent data. There are also techniques where the server has a pre-trained model and uses it to classify new data samples. In this paper, we provide an implementation of a secure deep learning model using PySyft [13] that uses convolutional neural networks to recognize pneumonia in chest X-ray images [14]. We provide evaluation of both these models, and present graphs of variation of the model accuracy against different parameters. Remaining part of the paper is organized as follows. Section 2 provides an overview of the concepts as well as the preliminaries need to understand the rest of the sections. We give a theoretical background to secure machine learning in

Secure Image Classification Using Deep Learning

515

Sect. 3. Section 4 describes our work and implementation of two secure models. Analyzed results are presented in Sect. 5. We conclude in Sect. 6.

2 Background The problem of secure computation [16] has been well studied in cryptography. In the setting of multi-party computation, two or more parties engage in a protocol. The parties each have their inputs, and agree on a common functionality that accepts these inputs and produces an output. For example, in the widely popular problem of multi-party private set intersection, the parties’ inputs are sets of elements, and the functionality computes the intersection of these sets. The goal of any Secure MPC protocol is to achieve two main concerns: the protocol should output the correct value according to the defined functionality, and each party should learn no additional information than what can be inferred from its own input and the obtained output. The main focus in this paper is on the semi-honest model, also otherwise known as the honest-but-curious model. Here, the parties are assumed to follow the specified Secure MPC protocol, but they may try to learn additional information from the protocol transcript. Semi-honest security in the context of machine learning captures a realistic scenario, since for currently used popular machine learning libraries, maliciously modifying the code is prevented through software attestations during deployment. Therefore, all that the parties can hope to obtain is the information received during the protocol, which mainly constitutes the transcript. Oblivious transfer (OT) is a commonly used cryptographic functionality which is widely used in MPC protocols. Here, a sender S has two strings s0 and s1 , and the receiver has a choice bit b. The OT protocol allows the receiver to obtain sb without gaining any information about s1−b . Moreover, the sender does not gain any information about b. Multiple oblivious transfers can be executed efficiently with a reduced cost using the OT extension technique [17]. OT protocols can be applied to realize any secure functionality using Yao’s garbled circuit approach [18]. This approach provides a generic mechanism for constructing a semi-honest secure twoparty protocol to compute a functionality f , using its Boolean circuit representation. In a garbled circuit protocol, one party (a circuit-generator) generates an encrypted version of the Boolean circuit for f . The other party (the circuit-evaluator) obliviously computes the output of the circuit without learning any intermediate values.

3 Privacy Preserving Machine Learning with SecureML In 2017, Mohassel and Zhang proposed SecureML [1], a privacy preserving machine learning technique that designs efficient protocols for various types of regressions and artificial neural networks. They consider the two-server model, where parties

516

K. Gururaj et al.

distribute their sensitive data among two servers S 1 and S 2 . It is assumed that the two servers do not collude, and thus, train models jointly using two-party MPC protocols.

3.1 Server Architecture There are n parties P1 , P2 … Pn , who wish to collectively train their data. This data may be sensitive, and the parties (also referred to as clients in context of a client– server architecture) do not wish to share their private data information to other parties. The work in [1] does not make any assumption about the distribution of client data and can also be output shares of a previous MPC protocol’s output. For reasons of efficiency, the servers outsource their data to remote servers, who are more computationally capable than clients. This scenario has been widely used in machine learning-based applications today, where clients take the role of lightweight mobile devices who outsource computation to servers. An advantage of a server-aided approach is that clients can send their data (as secret shares) to servers in a setup phase that is independent of the training phase. In our implementation described in Sect. 4, we focus on the two-server model, where the servers are assumed to not collude (i.e., there is no common semi-honest adversary A that simultaneously corrupts both the servers).

3.2 Secret Sharing In the SecureML protocol, the two servers exchange information between them as secret shares. Because of the assumption that the two servers never collude, it is impossible for one server to obtain information about the data with the other server and any of the clients. Three different secret sharing techniques are used: Boolean, Additive, and Yao shares. A finite field is used to generate additive shares. To additively secret share a field element a ∈ Z 2 l, a server chooses a random a0 ∈ Z 2 l, and sends a1 = a a0 mod 2l . The shares can be added and multiplied using corresponding operations over the finite field, as described in [1]. Similar to arithmetic sharing, Boolean shares can be obtained through similar operations in Z 2 . In this case, addition of shares corresponds to bit-wise XOR, and multiplication of shares corresponds to bit-wise AND. Finally, Yao shares are generated based on the encrypted values of wires present in the garbled circuit [18]. The advantage of having three different shares is that certain type of shares is used to carry out specific operations very efficiently. For instance, if two arithmetic values need to be multiplied, doing so with Boolean shares is very inefficient. With arithmetic sharing, a multiplication of the shares efficiently results in shares of the multiplication of the arithmetic values in the field.

Secure Image Classification Using Deep Learning

517

3.3 Definition of Security As mentioned in Sect. 2, we assume that the adversary A is semi-honest, meaning that the parties corrupted by A follow the specified protocol without deviating from it. However, any information received during the protocol is analyzed by the corrupted parties to obtain additional information. Although the two servers do not collude, an adversary can corrupt a server and a subset of the clients. The security requirement states that the adversary should learn only the inputs of the parties it has corrupted and should obtain no information about the other parties’ data than what is implied by the output. This definition is attributed to the real-world and ideal-world models defined in multi-party computation [16]. In an ideal world, a trusted third party exists that cannot be corrupted and does not reveal information to any of the parties. In such a case, the parties send their inputs to this trusted third party over a private channel. The third party locally computes the desired functionality, and outputs them to each party. In this world, any additional information obtained by the clients is a function of its own input and the local output. In the real world, where parties can be corrupted, a secure MPC protocol should be designed so that the corrupted parties learn no more than what can be obtained from their collective inputs and the obtained protocol output. This protocol, therefore, would “simulate” the ideal world, and the MPC protocol is said to securely evaluate the desired functionality. In the context of privacy preserving machine learning, the functionality to be evaluated should train the required model in our implementation. Therefore, the execution of the secure MPC protocol should not leak any information to the adversary, regarding the model parameters and the input dataset of honest clients used for training.

4 Implementation of the Two-Server Model Over the Chest X-ray Dataset We provide implementation of secure machine learning models in the semi-honest two-party setting. This model is a protocol for deep convolutional neural networks implemented using the PySyft library [13] on Pneumonia X-ray images [14]. We highlight the differences in performance of privacy preserving algorithms against ideal (non-secure) machine learning algorithms in Sect. 5. In the remainder of this section, we discuss on the implementation architecture and the code optimizations used to make the protocols more efficient. We implemented privacy preserving machine learning model on chest X-ray images [14] to recognize pneumonia. This dataset, being a collection of medical images, has more features and pixel-information in each image. Using a simple technique such as linear regression would perform poorly on this data. We, therefore, decided to use a more complicated machine learning technique using neural networks.

518

K. Gururaj et al.

More specifically, we use convolutional neural networks (CNNs) to train over this data. The advantage is that these type of neural networks are great at capturing localized information in images, which greatly improves accuracy. However, the implementation for a neural network is highly complicated, with multiple back-propagated derivatives at each node. Hence, a native C++ implementation posed many difficulties. We decided to use a pre-built Python library to achieve this task. We use PyTorch (a popular machine learning library for Python) as well as the PySyft [13] to implement a convolutional neural network. PySyft is a library for secure, private machine learning. It provides efficient implementations of many privacy preserving machine learning models. The setting used in our implementation is again the two-server model, where the servers do not collude. The general flow of the PyTorch script is as follows. Two servers are present that hold information corresponding to a model that has been trained on a specific dataset. To begin with, the server characterizes and trains a model with this private training data (securely, using a privacy preserving machine learning technique). At this point, the server connects with a client C, who holds its own portion of the information related to the dataset, and wishes to obtain the server’s model and make a few predictions. The server encrypts its trained model (which is a neural network in this case). The client encodes its information which it requires to query on the server’s model. At this point, both the server and client use the encrypted information present with them (the trained model and the encoded query, respectively). Finally, the server performs the query computation securely and this result is sent back to the client in an encrypted manner. This query is sent back in such a way that the server infers nothing about the client’s query data—neither the sources of information nor the final prediction results. The computation is secure in the honest-but-curious adversary model which is standard in many MPC frameworks as well. The structure of our implementation is as follows. We first split the dataset into three sets, the training, validation, and test sets. We also initialize two workers (corresponding to two servers training the data). This is followed by a data share function that sends data to these workers thereby splitting the dataset into nearly equal sets. A traditional convolutional neural network is used to train the medical image data. Two 2D convolutional layers (of sizes 3 × 20 × 5 and 20 × 50 × 5) are implemented in the neural network, followed by two fully connected (or linear) layers. We also used two rectified linear units (ReLUs) as the activation functions along with two layers for max-pooling. In PyTorch, the most optimal way to implement a network is to write a separate class. The class contains an initialization function as well as forward function with relevant dimensions. The hyperparameters used to train the model are given in Table 1. The training of the dataset is then launched. Now, the server securely sends the model to the workers holding the data. Because the model contains sensitive information, the server does not wish to disclose its weights. So, we use secret sharing (described in Sect. 3.2) to share the model among the workers (who are assumed to not collude) without revealing any information. This is followed by a testing function that performs an encrypted evaluation. The model weights, data inputs, prediction,

Secure Image Classification Using Deep Learning Table 1 Hyperparameters fine-tuned

Hyperparameters

519 Selection

Input

224 × 224

Activation function

ReLu

Optimizer

SGD

Error rate

Categorical cross entropy

Learning rate

0.0005–0.01

Kernel size

3×3

Momentum

0.9

Epochs

100

Batch size

32–512

Pooling

Max-pooling

Padding

Zero padding

Searching algorithm

Randomized search

and the target used for scoring are all encrypted, thereby revealing no additional information to the adversary.

5 Results For the secure convolutional neural network, we used the PySyft [13] library for the popular machine learning framework PyTorch. We applied different techniques such as max-pooling, SGD optimizers, and ReLU layers to find the optimal hyperparameters for the network. This secure model was trained on the medical dataset containing X-ray images to detect Pneumonia. Since we use a convolutional neural network, we decided to move from a simpler dataset such as MNIST, to a more complicated medical image-based dataset [14]. This allows our model to capture localized information more effectively and identify complex patterns that simple linear regression could not. The experiments were carried out on an Intel×86 processor running on Linux environment. We implemented a convolutional neural network to securely train a model for the X-ray medical datasetto detect pneumonia [15]. Figure 1 shows a few images from this dataset. As described in Sect. 4, the network architecture for the CNN consists of two 2D convolutional layers (of sizes 3 × 20 × 5 and 20 × 50 × 5), two linear layers, and two max-pooling layers. The ReLU activation function was used in PyTorch, along with the SGD optimizer.

520

K. Gururaj et al.

Fig. 1 X-ray images [14] to detect pneumonia

We obtained several graphs in 3 to compare the behavior of accuracy of the trained model against different parameters. Figure 2a shows the variation of accuracy against the number of epochs. The accuracy slightly increases in the beginning as the number of epochs’ increases, but it later converges to a constant value. Figure 2b shows the variation of accuracy against the batch size used for training. The optimal batch size to obtain highest accuracy was found to be 128. Figure 2c shows the variation of accuracy with learning rate. As the learning rate increases, the accuracy of our secure neural network decreases. This can be attributed to an increase in aggressiveness of the optimization equation. With a higher learning rate, the model overshoots the global minimum and tends to oscillate about that point. This leads to a decrease in overall accuracy. We found that a learning rate of 0.0005 was optimal to obtain maximum accuracy.

6 Conclusion In this work, we implemented practical algorithms for privacy preserving machine learning techniques. Specifically, we used the convolutional neural network to securely classify X-ray images to detect pneumonia. Using the PySyft library [13], we analyzed the practicality of using complex machine learning algorithms in the context of secure multi-party protocols. The performance of this privacy preserving technique was evaluated against conventional machine learning algorithms.

Secure Image Classification Using Deep Learning

521

Fig. 2 a Accuracy versus epochs, b accuracy versus batch size, c accuracy versus learning rate, d total execution time versus iterations, e setup time versus iterations, f online time versus iterations

522

K. Gururaj et al.

Fig. 2 (continued)

References 1. Mohassel P, Zhang Y (2017) Secureml: a system for scalable privacy-preserving machine learning. In: 2017 IEEE symposium on security and privacy. IEEE, pp 19–38 2. Tramèr F, Juels A, Reiter MK, Ristenpart T (2016) Stealing machine learning models via prediction Apis. In: Proceedings of the 25th USENIX conference on security symposium, SEC’16, (USA). USENIX Association, pp 601–618 3. Song T, Ristenpart T, Shmatikov U (2017) Machine learning models that remember too much. In: Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, CCS’17, (New York, NY, USA). Association for Computing Machinery, pp 587–601 4. Shokri R, Stronati M, Song C, Shmatikov V (2017) Membership inference attacks against machine learning models. In: 2017 IEEE symposium on security and privacy (SP), pp 3–18 5. Abadi M, Chu A, Goodfellow I, McMahan HB, Mironov I, Talwar K, Zhang L (2016) Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, CCS ’16, Association for Computing Machinery, New York, NY, USA, pp 308–318 6. McMahan HB, Ramage D, Talwar K, Zhang L (2017) Learning differentially private language models without losing accuracy. ArXiv, vol. abs/1710.06963

Secure Image Classification Using Deep Learning

523

7. Mohassel P, Rosulek M, Zhang Y (2015) Fast and secure three-party computation: the garbled circuit approach. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, CCS ’15, Association for Computing Machinery, New York, NY, USA, pp 591–602 8. Furukawa J, Lindell Y, Nof A, Weinstein O (2017) High-throughput secure three-party computation for malicious adversaries and an honest majority. In: Coron J-S, Nielsen JB (eds) Advances in cryptology—EUROCRYPT 2017. Springer International Publishing, Cham, pp 225–255 9. Araki T, Furukawa J, Lindell Y, Nof A, Ohara K (2016) High-throughput semi-honest secure three-party computation with an honest majority. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, CCS ’16, Association for Computing Machinery, New York, NY, USA, pp 805–817 10. Mohassel P, Rindal P (2018) Aby3: a mixed protocol framework for machine learning. In: Proceedings of the 2018 ACM SIGSAC conference on computer and communications security, CCS ’18, Association for Computing Machinery, New York, NY, USA, pp 35–52 11. Dwork C, Nissim K (2004) Privacy-preserving datamining on vertically partitioned databases. In: Franklin M (ed) Advances in cryptology—CRYPTO 2004. Springer, Berlin, Heidelberg, pp 528–544 12. Dwork C (2006) Differential privacy in automata, languages and programming. Springer, Berlin, Heidelberg, pp 1–12 13. OpenMinded, “PySyft.” www.openmined.org 14. Mooney P, Chest X-Ray images. www.kaggle.com/paultimothymooney/chest-xray-pneumonia 15. Bishop CM (2006) Pattern recognition and machine learning, information science and statistics. Springer, Berlin, Heidelberg 16. Oded G (2009) Foundations of cryptography: volume 2, basic applications, 1st edn. Cambridge University Press, USA 17. Ishai Y, Kilian J, Nissim K, Petrank E (2003) Extending oblivious transfers efficiently. In: Boneh D (ed) Advances in cryptology—CRYPTO 2003. Springer, Berlin, Heidelberg, pp 145–161 18. Yao C (1986) How to generate and exchange secrets. In: Proceedings of the 27th annual symposium on foundations of computer science, SFCS ’86, IEEE Computer Society, USA, pp 162–167